Jeremy N BailensonDepartment of CommunicationStanford UniversityStanford CA 94305 USABailensonstanfordedu

Kim SwinthDepartment of PsychologyUniversity of California SantaBarbaraSanta Barbara CA 93106 USA

Crystal HoytSchool of LeadershipUniversity of RichmondRichmond VA 23173 USA

Susan PerskyAlex DimovJim BlascovichDepartment of PsychologyUniversity of California SantaBarbaraSanta Barbara CA 93106 USA

Presence Vol 14 No 4 August 2005 379ndash393

copy 2005 by the Massachusetts Institute of Technology

The Independent and InteractiveEffects of Embodied-AgentAppearance and Behavior on Self-Report Cognitive andBehavioral Markers ofCopresence in Immersive VirtualEnvironments

Abstract

The current study examined how assessments of copresence in an immersive vir-tual environment are influenced by variations in how much an embodied agent re-sembles a human being in appearance and behavior We measured the extent towhich virtual representations were both perceived and treated as if they were hu-man via self-report behavioral and cognitive dependent measures Distinctive pat-terns of findings emerged with respect to the behavior and appearance of embod-ied agents depending on the definition and operationalization of copresenceIndependent and interactive effects for appearance and behavior were found sug-gesting that assessing the impact of behavioral realism on copresence without takinginto account the appearance of the embodied agent (and vice versa) can lead tomisleading conclusions Consistent with the results of previous research copresencewas lowest when there was a large mismatch between the appearance and behav-ioral realism of an embodied agent

1 Introduction

The use of avatars (ie digital representations of human users) to repre-sent individuals within multiuser virtual environments in real time has in-creased considerably in the past decade underscoring the need to better un-derstand how users experience mediated interactions with others In additionthe use of embodied agents (ie digital representations of computer programsthat have been designed to interact with or on behalf of a human) as interac-tive guides or as interactants within virtual environments has prompted inquiryinto how people perceive and respond to nonhuman social entities

In this paper we examine how definitions and operationalizations of copres-ence (also commonly referred to as social presence) are influenced by variationsin the human resemblance of embodied agents in terms of appearance and be-havior within an immersive virtual environment In doing so we explore howand when definitional and measurement differences regarding copresence leadto convergent and divergent findings with respect to the independent and in-

Bailenson et al 379

teractive effects of the appearance and behavior of em-bodied agents Our multimethod approach includedself-report behavioral and cognitive dependent mea-sures

11 Defining and OperationalizingCopresence

One of the difficulties in understanding and syn-thesizing the accumulating literature on copresence isthe plethora of definitions and operalizations used byinvestigators as can be seen in the thorough discussionby Schroeder (2002) Unfortunately this difficulty hasled to inconsistent and often contradictory findings re-ported in the copresence literature Heeter (1992) forexample defined copresence (or social presence) as theextent to which other beings both living and syntheticexist in a virtual world and appear to react to humaninteractants Nowak (2001) on the other hand definedcopresence as the sense of connection with anothermind Russo (2001) argued that copresence (or whatshe called mediated presence) is the perception by a com-municator that another person in a mediated or onlineenvironment is real immediate or present ConverselySlater and his colleagues have defined copresence as thesense of being and acting with others in a virtual place(Slater Sadagic Usoh amp Schroeder 2000) FinallyBlascovich and his colleagues have defined copresence asthe extent to which individuals treat embodied agents asif they were other real human beings (Blascovich et al2002)

Although there is no consensual definition of cop-resence or agreement about how best to assess itexisting research on copresence can be roughly di-vided into two camps studies that focus on peoplersquosperceptions of embodied agents versus studies thatfocus on peoplersquos social responses to embodied agentsResearchers from the first camp (eg Heeter 1992Biocca 1997 Lombard amp Ditton 1997 GarauSlater Bee amp Sasse 2001 Lessiter Freeman Keoghamp Davidoff 2001 define copresence in terms of a us-errsquos perception or feeling that others are cosituatedwithin an interpersonal environment (Swinth amp Blas-covich 2002) Researchers in this

tradition take the view that copresence occurs when auser not only perceives or feels that he or she isldquopresentrdquo within a virtual environment but that heor she perceives or feels that others are present thereas well Because this view of copresence focuses onpeoplersquos phenomenological experiences researchersadopting this orientation frequently operationalizecopresence in terms of a userrsquos self-reported percep-tion andor feeling that they are not alone in a vir-tual environment

Another group of researchers however views copres-ence as a social task-related or physiological response toembodied agents (eg Bailenson et al 2004 Blasco-vich et al 2002 Reeves amp Nass 1996 BailensonBeall amp Blascovich 2002 Beall Bailenson LoomisBlascovich amp Rex 2003 Blascovich et al 2002 Lee ampNass 2004 Bente Ruggenberg Tietz amp Wortberg2004 Beall Bailenson Loomis Blascovich amp Rex2003 Meehan Insko Whitton amp Brooks 2004Slater 2004 Bente Ruggenberg Tietz amp Wortberg2004) Researchers adopting this orientation argue thatcopresence occurs when people treat embodied agentsas if they were other real people and they operationalizecopresence using a variety of methods including theanalysis of physiological data examination of a variety ofnonverbal and verbal behaviors or the assessment oftask performance cooperation interaction satisfactionliking or other social responses

In the current study participants were asked to exam-ine embodied agents that varied in behavior and appear-ance We then measured the extent to which the em-bodied agents were perceived and treated as if they werehuman via self-report behavioral responses and cogni-tive dependent measures Adopting this approach en-abled us to compare the two general ways of definingcopresence described above by examining both partici-pantsrsquo self-reported perceptions of embodied agents aswell as their self-reported cognitive and behavioral re-sponses to them Moreover by employing three distinctmeasurement modalities (self-report questionnaires be-havioral responses and cognitive responses) we wereable to distinguish converging and diverging patternspertinent to each of the two general ways of defining

380 PRESENCE VOLUME 14 NUMBER 4

copresence We discuss each of these response modali-ties in turn

12 Self-Report Markers of Copresence

In order to tap the phenomenological aspects ofcopresence we administered a self-report questionnairethat asked participants to report on the extent to whichthey perceived or felt like there was another social entitywith them in the virtual environment In addition weadministered two self-report measures designed to assessdifferent kinds of social responses people might have toembodied agents liking of an embodied agent and will-ingness to perform embarrassing acts in front of an em-bodied agent We hypothesized that all three indicatorsof copresence would be correlated positively for copres-ence and liking and negatively between copresenceliking and willingness to perform embarrassing acts

13 Behavioral Markers of Copresence

In order to assess the idea that one can infer co-presence by observing interaction with an embodiedagent we assessed a behavioral measure of copresenceinterpersonal distance Building on previous work (egReeves amp Nass 1996 Krikorian Lee Chock amp Harms2000 Bailenson Blascovich Beall amp Loomis 2001Bailenson Blascovich Beall amp Loomis 2003) we hy-pothesized that copresence would be positively relatedto hesitation in approaching an embodied agent inter-personal distance and maintenance of interpersonal dis-tance

14 Cognitive Markers of Copresence

Furthermore we assessed memory for embodiedagents Previous research has found that people remem-ber more features about people in scenes than aboutnonsentient objects in scenes (New 2003) Conse-quently users should remember more informationabout embodied agents represented as sentient beings(ie human) than about nonsentient types of represen-tations Thus we expected better memory for embodied-agent features However Hoyt Blascovich and Swinth

(2003) demonstrated that the presence of an embodiedagent worsened performance on a cognitive task whenparticipants performed in front of an audience of ava-tars and that perceived copresence mediated that effectConsequently it could be argued that copresence mayin fact be indicated by worse memory for informationrelating to an embodied agent In the current study weexamined usersrsquo memory for an embodied-agent featureas a cognitive marker of copresence on an exploratorybasis without predictions concerning the direction ofmemory effects Our goal in including this cognitivemarker of copresence was to compare any memory dif-ferences observed to other more established and pre-dictable markers of copresence in order to begin to un-derstand the cognitive effects of copresence if any

15 Factors Affecting Copresence

Although a number of factors may influence co-presence two appearance and behavior of embodiedagents have received considerable attention Althoughsome work has examined the interplay between howhuman an embodied agent appears and how an embod-ied agent behaves (eg Bailenson et al 2001 Garau etal 2003) researchers typically examine each factor inisolation or confound them as a single variable

Bailenson et al (2001) for example found that mak-ing an embodied agent more realistic looking by addinga texture map from an actual photograph of a face didnot increase copresence compared with an embodiedagent that was more cartoonlike as long as both typesof embodied agents demonstrated realistic gaze behav-iors Similarly Garau et al (2003) demonstrated thatmore photographically realistic avatars elicited no morecopresence than did less realistic-looking avatars and infact found that increasing the photo-realism of an ava-tar can actually cause a decrease in copresence if the be-havioral realism is not also increased correspondingly(see also Slater amp Steed 2001 Garau 2003)

In contrast Nowak and her colleagues (Nowak2001 Nowak amp Biocca 2001) demonstrated that theextent to which embodied agents resemble real humanbeings affects peoplersquos social judgments of agents as in-teraction partners and the amount of copresence experi-

Bailenson et al 381

enced In one study interaction partners represented byanthropomorphic images were rated as more sociallyattractive and they received higher partner satisfactionratings than did partners represented either by no imageor by an image that did not appear human (Nowak2001) In addition participants indicated that they feltgreater copresence and that the communication me-dium was better able to support the social interactionwhen interactants were represented by highly anthropo-morphic images compared to when no image or an im-age low in anthropomorphism was used (Nowak ampBiocca 2001) Similarly Parise Kiesler Sproull amp Wa-ters (1996) demonstrated that more human-lookingembodied agents elicited more liking behavior and morecooperation than did less human-looking embodiedagents

One possible explanation for these seemingly contra-dictory findings concerns previous manipulations of ap-pearance Studies that have demonstrated null effectshave typically manipulated the visual fidelity (or photo-realism) of embodied agents and studies that havefound a significant effect of embodied-agent appearanceon copresence and other social outcomes have typicallymanipulated the anthropomorphism of the embodiedagent or the extent to which the embodied agent ap-pears human in form We are unaware of studies thathave systematically varied both photo-realism and an-thropomorphism and unfortunately most studies thathave been conducted on the effect of embodied-agentappearance have tended to confound photo-realism andanthropomorphism

Given that studies focusing on the photo-realism ofan embodied agent have consistently reported null ef-fects however we attempted to control for the effect ofphoto-realism in the current study by using embodiedagents of equivalent visual fidelity while simultaneouslymanipulating how much like an actual human being theembodied agent both looked and behaved We pre-dicted that embodied agents that either looked humanor acted human would elicit greater copresence thanembodied agents that did not look or act human How-ever based on the work of Blascovich and his colleagues(Blascovich 2002 Blascovich et al 2002) we alsohypothesized that embodied-agent appearance and be-

havior would interact with each other such that the be-havioral realism of an embodied agent would mattermost for non-human-looking embodied agents

16 Summary and Study Objectives

In this study we systematically examined the inde-pendent and interactive effects of both the appearanceand behavior of embodied agents on usersrsquo perceptionsand social responses within an immersive virtual envi-ronment In doing so we attempted to address a num-ber of currently unresolved issues First we examinedhow different ways of defining and operationalizingcopresence are influenced by systematic manipulationsof appearance and behavior of virtual representations Indoing so we attempted to reconcile some of the con-trasting findings that have been reported in the researchliterature Second we examined whether similar or dis-parate patterns of response obtained across differentmodalities of copresence measurement (ie self-reportvs cognitive vs behavioral dependent measures) Finallywe used four different algorithms to control the headmovements of the embodied agent some of which havenever before been empirically compared in an immersivevirtual environment

2 Method

21 Participants

One hundred and forty-six undergraduates (73males 73 females) ranging in age from 18 to 27 years(M 1963 years) recruited from the University of Cal-ifornia Santa Barbara served as participants They wereeither paid or received class credit for participation

22 Design

A two-way between-subjects factorial design wasused The first factor representation type included threelevels of embodiment human teddy bear and block-head As can be seen in Figure 1 the three embodiedagents varied in the extent to which they resembled a

382 PRESENCE VOLUME 14 NUMBER 4

human being1 For female participants the human rep-resentation created using 3dMeNow software depictedin the left panel of Figure 1 was used For male partici-pants a similar-looking male representation was usedAll three representations possessed equivalent behavioralfunctionality within the virtual environment (see below)and exhibited similar levels of photo-realism

While we attempted to ensure as much as possibleequal levels of photo-realism (by using textures of simi-lar resolution and by ensuring that the underlying 3Dmesh for each model contained similar numbers of an-chor points and structural facial features) subjectivelythere may have been minor differences in photo-realism

The second between-subjects factor behavioral real-ism included four levels of head movements Four dif-ferent algorithms were employed to control the embod-ied agentrsquos head movements In the static condition theembodied agent remained motionless In the randommovement condition the embodied agentrsquos head contin-ually moved in trajectories generated by the computertoward random points within a 60-degree arc of eachside of its head (for both pitch and yaw) In the mimiccondition the embodied agentrsquos head movements mir-rored the participantrsquos actual head movements with a

4-s delay2 In the recorded movement condition headmovements recorded from a randomly selected partici-pant in a previous session were used to control the em-bodied agentrsquos head movements In all four conditionsthe embodied agent blinked its eyes according to a pre-set algorithm but there were no facial gestures exhib-ited Complete crossing of these two independent vari-ables resulted in 12 between-subjects conditions3

22 The Virtual Environment

The virtual room in which participants were im-mersed is depicted in Figure 2 Participants sat in achair at a table facing an embodied agent Participantswere able to see the head and shoulders of the agentThe same body was used for each representation type Atag containing a string of letters and numbers later usedas part of a surprise memory task was positioned on theshirt of the embodied agent

1 There are a variety of terms that researchers use to describe howreal or how human a virtual representation looks including anthropo-morphism visual realism and photographic (or photo-) realism Wehave chosen to use the more general term representation type as a vari-able name and talk about the manipulation as mainly pertaining tophotographic realism We acknowledge that other terms such as an-thropomorphism however may also be appropriate

2 Previous research using mimicked head movements has indicatedthat users generally do not detect embodied agents that are mimickingtheir head movements with a 4-second delay (Bailenson Beall Loo-mis Blascovich amp Turk 2004 Bailenson amp Yee in press) Howeverthis study is unique in that it directly compared the effect of mimickedhead movements with other head-movement algorithms

3 There were an unequal number of females and males across the12 conditions

Figure 1 Photographically realistic human teddy bear and

blockhead embodied agents

Figure 2 A participantrsquos view of the virtual room

Bailenson et al 383

23 Equipment

Participants wore a Virtual Research V8 stereo-scopic head-mounted display (HMD see Figure 3) TheHMD featured dual 680 horizontal by 480 vertical pixelresolution LCD panels that refreshed at 72 Hz Thedisplay optics presented a visual field subtending approxi-mately 50 degrees horizontally by 38 degrees vertically

Perspectively correct stereoscopic images were ren-dered by a 450 MHz Pentium III dual-processor com-puter with an Evans amp Sutherland Tornado 3000 dualpipe graphics card and were updated at an averageframe rate of 36 Hz The simulated viewpoint was con-tinually updated as a function of the participantsrsquo headmovements The orientation of the participantrsquos headwas tracked by a three-axis orientation sensing system(Intersense IS300 update rate of 150 Hz) The systemlatency or the amount of delay between a participantrsquoshead movement and the resulting concomitant updatein the HMDrsquos visual display was 65 ms maximum Thesoftware used to assimilate the rendering and trackingwas Vizard 20 Participants used a Logitech Rumble-

Pad Pro game pad to interact with the virtual environ-ment

24 Procedure

When participants arrived at the laboratory anexperimenter sat them down and instructed them onhow to wear and adjust the HMD and how to use thegame pad depicted in Figure 3 to interact with thevirtual environment Once immersed participantsfound themselves seated at a table directly acrossfrom an embodied agent (see Figure 2) Dependingon the assigned condition the embodied agent variedin appearance and behavior Using the game pad par-ticipants scrolled through the following instructionaltext that appeared above the embodied agentrsquos head

In this laboratory we are interested in peoplersquos reac-tions to ldquoothersrdquo in virtual environments In thisstudy you will spend time with an ldquootherrdquo After ashort time we will ask you questions about theldquootherrdquo You will use the game pad to answer thequestions The questions measure your agreementThe more you agree the higher the number youshould choose on the rating scale in front of you

We carefully chose and repeatedly used the wordldquootherrdquo set off by quotation marks to avoid using sug-gestive nouns such as ldquopersonrdquo or ldquoagentrdquo Participantswere next instructed to look directly at the ldquootherrdquo for90 s Next a series of 10 copresence embarrassment andlikability self-report questionnaire items (see AppendixA) appeared one at a time over the embodied agentrsquoshead in random order Participants indicated theiragreement with these items by using the game pad toselect a response on a Likert-type rating scale that ap-peared under each item Hence ratings were providedin real time as participants experienced the embodiedagent

After completing the items participants were in-structed to approach the embodied agent for furtherexamination They used buttons on the game pad tomove toward (positive translation on the z axis) or awayfrom (negative translation on the z axis) the embodiedagent with each press of a game-pad button equivalentto a jump of 25 cm (the starting distance between the

Figure 3 A depiction of our immersive virtual environment system

The components are (1) orientation tracking sensor (2) image

generator (3) HMD and (4) game pad input device

384 PRESENCE VOLUME 14 NUMBER 4

participantrsquos nose and the embodied agentrsquos nose wasapproximately 3 m) We did not render the body of par-ticipants so there was no leaning or walking animationduring the translationmdashonly their viewpoint movedcloser or farther from the agent After participants fin-ished approaching the embodied agent and examiningit they took off the HMD and filled out a demographicquestionnaire Finally we administered a surprise mem-ory test in which participants were asked to recall therandom string of characters written on the embodiedagentrsquos chest (see Figure 2)

25 Measures

251 Perceived Copresence Perceived copres-ence was assessed via a three-item self-report measureWe created a composite copresence measure ( 71see Appendix A) by averaging the three copresencequestions Higher scores indicated higher levels ofself-reported copresence where ratings were made on aLikert-type scale with response options ranging from 0(strongly disagree) to 6 (strongly agree)

252 Embarrassment Embarrassment was as-sessed as a self-reported social response to the presenceof an embodied agent Previous research has demon-strated that willingness to perform embarrassing acts infront of a virtual representation is a worthy measure ofthat representationrsquos degree of social influence (Bailen-son et al 2003) We averaged the three embarrassmentquestions ( 72 see Appendix A) into a single com-posite measure with higher numbers indicating greaterwillingness to perform embarrassing acts in front of anembodied agent

253 Likability Likability was assessed as a sec-ond self-reported social response to the presence ofan embodied agent We averaged the four likabilityquestions ( 71 see Appendix A) into a singlecomposite measure with higher numbers indicatingthat participants liked the virtual representation more

254 Memory Memory was quantified asthe total number of correct characters (out of a possi-

ble eight) a participant recalled from the virtual rep-resentationrsquos label without penalty for incorrect an-swers

255 Interpersonal Distance Interpersonaldistance was included as a behavioral social responseto the presence of an embodied agent We assessedtwo indicators of interpersonal distanceminimumdistance and reversal count Minimum distance wasthe closest distance the participant came to the em-bodied agent reversal count referred to the numberof times a participant changed direction along thez-axis (ie went back and forth) during their exami-nation of the embodied agent We consider reversalcount to be a proxy for hesitation in approaching theembodied agent

3 Results and Discussion

31 Hypotheses

We hypothesized that greater copresence wouldbe indicated by higher perceived copresence ratingsgreater liking of the embodied agent less willingness tocommit embarrassing acts in front of the embodiedagent the maintenance of greater interpersonal dis-tance and more hesitation in approaching the embod-ied agent We did not however have any specific pre-dictions with respect to the relationship betweencopresence and memory

Based on the results from previous copresencestudies we expected greater copresence (across ourvarious copresence measures) to occur with realisti-cally behaving embodied agents (specifically our re-corded and mimic behavioral-realism conditions thatfeatured movements based on actual human motion)compared with unrealistically behaving embodiedagents (ie our static and random-movement condi-tions) In addition we expected greater copresence tooccur with more human-looking embodied agentsthan with less human-looking ones Drawing on pre-dictions derived from the Blascovich et al (2002)threshold model of social influence however we ex-pected embodied-agent appearance to moderate the

Bailenson et al 385

effect of behavioral realism such that behavioral real-ism would matter more for less human-looking em-bodied agents than for more human-looking embod-ied agents Finally we expected to demonstrateconvergence in our dependent measures such thatthey would correlate with one another and change inthe same way as a function of our independent vari-able manipulations

32 Perceived Copresence Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movement vs mimic vs recorded) analysisof covariance controlling for participant gender4 re-vealed a significant main effect of behavioral realism onself-reported perceived copresence F (3 133) 272p 05 2 06 As Figure 4 demonstrates perceivedcopresence was higher in the random-movements con-dition compared to the static condition p 01 How-ever this main effect of behavioral realism was qualifiedby a significant representation type by behavioral-realisminteraction F (6 133) 218 p 05 2 09

Simple main effects tests revealed that within theblockhead condition perceived copresence varied signif-icantly as a function of behavioral realism F (3 133)

339 p 05 2 07 Specifically perceived copres-ence was significantly higher when the blockhead repre-sentation moved according to our random-movementalgorithm compared with either the static or mimic con-ditions ps 05 Simple main effects tests also revealedthat perceived copresence varied as a function of behav-

ioral realism for participants in the teddy-bear condi-tion F (3 133) 330 p 05 2 07 Specificallyperceived copresence was significantly higher in mimiccondition than in either the static or recorded condi-tions ps 05 and higher in the random-movementscondition relative to the recorded and static conditions(ps 05 and 10 respectively)

In contrast there were no significant differences inperceived copresence between any of the four behavioral-realism conditions for participants in the humanldquootherrdquo condition This pattern of findings is consis-tent with our prediction that behavioral realismwould have less of an influence on perceived copres-ence for embodied agents that appear human ascompared to embodied agents with a nonhuman ap-pearance In addition it is interesting to note that thetype of head movement that elicited the greatest levelof perceived copresence varied as a function of the em-bodied agentrsquos appearance with participants attributinghigh copresence ratings to both mimicking and ran-domly moving teddy-bear representations but only ran-domly moving blockhead representations

4 Although both male and female participants were used in thisstudy gender was not included as a separate factor in analyses involv-ing both behavioral realism and representation type When randomlyassigning participants to conditions we did not block participants bygender and as a result when behavioral realism and representationtype were fully crossed there were not enough participants of bothgenders in all 12 experimental conditions to reliably compute a three-way analysis of variance Thus our data analytic technique involvedincluding participant gender as a covariate in our analyses involvingboth behavioral realism and representation type in order to control forthe potential effect of participant gender The results do not changeappreciably if we remove gender as a covariate in our analyses but weinclude gender as a covariate to provide the most thorough analysispossible

Figure 4 Mean perceived-copresence ratings (using gender-

covariate adjusted means) as a function of representation type and

behavioral realism Conditions that share the same letter do not differ

within that level of representation type at p 05

386 PRESENCE VOLUME 14 NUMBER 4

33 Embarrassment Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender re-vealed a significant main effect of representation type onparticipantsrsquo self-reported willingness to perform embar-rassing acts in front of the embodied agent (ie embar-rassment) F (2 133) 421 p 05 2 06 Spe-cifically participants in the human embodied-agentcondition (M 260 SD 155) were significantlyless willing to perform embarrassing acts in front ofthe ldquootherrdquo than were participants in the teddy-bear(M 351 SD 154) embodied-agent conditionp 05 Participants in the human-representationcondition were also marginally less willing to performembarrassing acts than participants in the blockhead(M 313 SD 156) condition p 10 In sumthe presence of a human-looking representationseemed to reduce participantsrsquo willingness to performin embarrassing acts Regardless of the behavioral re-alism of the representation embodied agents wholooked like humans elicited greater social influencethan did non-human-looking representations

34 Likability Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender alsorevealed a significant main effect of representation typeon self-reported likability ratings F (2 133) 543p 01 2 08 Specifically both the human-lookingrepresentation (M 250 SD 109) and the teddy-bear representation (M 228 SD 109) were ratedas significantly more likable than the blockhead repre-sentation (M 179 SD 110) ps 05 Partici-pants found the blockhead representation (which wasthe least human looking of the three representations) tobe less attractive than either of the other two represen-tation types and they liked the blockhead representationthe least

35 Correlations between PerceivedCopresence Embarrassment andLikability Ratings

In order to examine the relationships among per-ceived copresence embarrassment and likability a se-ries of Pearson correlations was computed Consistentwith out predictions results of these analyses revealed asignificant negative correlation between perceived co-presence and a participantrsquos willingness to perform em-barrassing acts in front of the embodied agent r(144) 27 p 001 as well as a significant positivecorrelation between perceived copresence and likabilityr (144) 31 p 001 Specifically we found that thehigher the perceived copresence experienced the lesswilling participants were to perform embarrassing acts infront of the embodied agent (r2 07) and the moreparticipants liked the embodied agent (r2 10) Therewas no significant correlation between likability and em-barrassment ratings

36 Memory

We computed a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender to ex-amine the effects of our independent variables on partic-ipantsrsquo memory Results of these analyses revealed a sig-nificant main effect of behavioral realism on memory F(3 133) 414 p 01 2 09 Specifically mem-ory was significantly worse in the mimic condition thanin either the static or random-movement conditionsps 05 suggesting that overall participants had theworst memory for characters worn by a representationthat mimicked their own movements with a 4-s delayAgain however this main effect was qualified by a sig-nificant representation type by behavioral-realism inter-action (see Figure 5) F (6 133) 270 p 02 2

11Simple main effects tests revealed that within the

blockhead condition memory varied significantly as afunction of behavioral realism F (3 133) 271p 05 2 06 Specifically individuals rememberedsignificantly more characters when they were in the

Bailenson et al 387

random-movement condition than in either the mimicor the recorded conditions (the two behavioral-realismconditions that appeared most humanlike) ps 05and participants in the random-movement conditionremembered marginally more than participants in thestatic condition p 10 These results converge quitewell with the perceived copresence data in which partic-ipants indicated higher copresence ratings for blockheadrepresentations only when the blockhead demonstratedrandom head movements Interestingly this suggeststhat for representations who look the least human theleast human-looking head movements were seen asmost appropriate This finding that avoiding mis-matches is the most effective manner of combining anembodied agentrsquos appearance and behavior is consistentwith previous research on visually embodied agents(Slater amp Steed 2001 Garau et al 2003) as well asvoice agents (Lee amp Nass 2003)

Furthermore the convergence of memory andperceived-copresence ratings provides some evidencethat better memory is indicative of higher copresenceIn other words the more copresence an embodiedagent elicits the more a person interacting with that

embodied agent remembers about it It is interesting tonote however that despite the similar patterns of find-ings between the perceived-copresence data and thememory data results of a Pearson correlation revealedthat there was no significant linear relationship betweenperceived copresence and peoplersquos ability to rememberinformation about an embodied agent

There was however a significant simple main effectof behavioral realism within the human-representationcondition F (3 133) 549 p 001 2 11Specifically participants in the mimic and random-movement conditions remembered significantly fewercharacters than participants in either the static or re-corded conditions ps 05 Participants in the teddy-bear condition displayed the same pattern but in thatcondition none of the differences reached statisticalsignificance

These results suggest that participants may have beendistracted by a human-looking embodied agent thatmimicked their own movements which in turn af-fected their ability to notice and recall information inthe virtual environment Moreover participants mayalso have been distracted in the random-movement con-dition since this condition represented the greatest mis-match between the appearance (human looking) andbehavior (not human looking) of the embodied agentFurther research should examine these potential expla-nations as they suggest that different processes maybe involved for different algorithms for generatingembodied-agent head movements

37 Interpersonal Distance

There were no significant effects of behavioral re-alism or representation type on minimum distance (ourfirst measure of interpersonal distance) Thus partici-pants in all conditions maintained equivalent levels ofinterpersonal distance between themselves and the em-bodied agent There was also no significant correlationbetween perceived copresence and either of our behav-ioral dependent measures (ie minimum distance andnumber of returns)

However as hypothesized results of a 3 (representa-tion type human vs teddy bear vs blockhead) 4 (be-

Figure 5 Gender-covariate adjusted mean number of correctly

recalled characters out of eight on a surprise memory test as a

function of representation type and behavioral realism Conditions that

share the same letter do not differ within that level of representation

type at p 05

388 PRESENCE VOLUME 14 NUMBER 4

havioral realism static vs random movements vs mimicvs recorded) analysis of covariance controlling for par-ticipant gender revealed a significant main effect of be-havioral realism on reversal count or the number oftimes a participant changed direction along the z-axisduring his or her examination of the embodied agentF(3 133) 306 p 05 2 07 As can be seen inFigure 6 participants in the random-movement condi-tion exhibited significantly fewer reversals than partici-pants in any of the other three behavioral-realism condi-tions Pairwise comparisons indicated that there weresignificantly fewer reversals in the random-movementcondition compared to the other three conditions(ps 05) which are not significantly different fromeach other

This finding in part confirms our hypothesis con-cerning our behavioral dependent variable For an em-bodied agent whose head moves with curves randomlygenerated by the computer participants show little tono hesitation in approaching In contrast when thehead movements were controlled by a real person (ei-ther a mimic of their own movements or the move-ments from another participant) participants showedgreater hesitation in approaching the embodied agentThus computer-generated movements generated lesshesitancy than human-generated movements

Given this explanation however one might expecteven less hesitation that is fewer reversals in approach-ing the static representation This finding did not occurand one potential explanation is that overall a ran-domly moving representation may elicit less social influ-ence and copresence than a completely static representa-tion In other words the static representation has atleast the potential to exhibit realistic movements whilethe unrealistically moving embodied agent immediatelylowers the sense of copresence In sum participantsrsquohesitancy in approaching an embodied agent was influ-enced by the behavioral realism of the embodied agentbut not its appearance

4 Conclusions

41 Summary of Findings

To summarize the appearance and behavior ofembodied agents had varying effects on the dependentvariables In many instances there was convergence inthat self-report cognitive and behavioral measureschanged in similar ways as a function of our indepen-dent variable manipulations For example the perceived-copresence likability and embarrassment measures cor-related in the predicted directions and the blockheadrepresentation elicited similar reactions across behaviortypes for both the memory measure and the perceived-copresence measure In other instances however therewas a lack of convergence where varying realism pro-duced different outcomes for the various responsemodes One inevitable challenge facing researchers whoexamine a large number of dependent variables is tounderstand the instances in which those dependentvariables do and do not converge

In the current study for some of our dependent mea-sures of copresence (willingness to commit embarrassingacts and likability) manipulation of representation typeaffected copresence responses On other dependentmeasures (self-reported copresence ratings interper-sonal distance approach reversals and memory) ma-nipulation of behavioral realism changed participantsrsquoresponses One way to interpret these results is thatself-report measures were adequate for measuring

Figure 6 Mean number of direction reversals while approaching the

embodied agent as a function of behavioral realism

Bailenson et al 389

copresence levels as affected by variation in appear-ance whereas self-report measures were not sensitiveenough to detect the subtle differences elicited by thebehavioral-realism manipulation Indeed the only reliablemain effects of the behavioral-realism manipulation oc-curred with cognitive and behavioral dependent measures

In other words self-report questionnaires are effectivewhen measuring how people perceive an embodiedagent but not necessarily how they will respond to thatembodied agent Consequently not only does it matterhow one defines copresence it also matters how oneoperationalizes copresence If the goal of a researcher isto understand the way in which people fathom embod-ied agents on a conceptual level then self-report ques-tionnaires may be the best tool However if the goal ofthe researcher is to understand the way in which peoplewill interact with learn from or behave toward anagent then he or she should augment questionnaireswith other types of measures In this paper we are notlobbying for the abandonment of copresence self-reportquestionnaires however we do urge that researchersuse these questionnaires as a complement to other typesof measures whenever possible

While this explanation that self-report measures lacksensitivity receives support from the current data itmust be qualified On most measures there was an in-teraction between behavioral realism and representationtype At times these interactions did not conform toour specific predictions and were difficult to interpretPart of the reason for this difficulty is most likely due tothe fact that we assumed a priori that certain types ofhead movements would be viewed as more realistic thanothers Specifically we assumed that the measures basedon actual human behaviors (mimic and recorded) wouldappear more realistic than either those generated by thecomputer (random) or than no movements at all Re-corded human movements applied to 3D models how-ever may not always be the most effective method ofrendering realistic behaviors and the differences be-tween our various levels of behavioral realism may havebeen too subtle in the current study As previous workhas demonstrated (Garau et al 2003) the interactionbetween appearance and behavioral realism is compli-cated and the finding that seems to be emerging is that

large disparities between the two types of realism cancause diminished levels of copresence Only when theblockhead representation the least humanlike of thethree demonstrated unrealistic random head move-ments did it elicit high levels of self-report and cognitivecopresence

Our findings with respect to the main effects of ap-pearance on self-report measures suggest that how hu-man an embodied agent looks may also be an importantprecondition for experiencing copresence This findingis unanticipated given predictions made by the Blasco-vich model and given findings from previous work(Bailenson et al 2001 Garau et al 2003) It seemsthat virtual representations that look like a realistic hu-man as opposed to either a familiar or unfamiliar non-human elicit embarrassment reactions that would beappropriate only in the presence of a real human

42 Limitations and Future Research

Though this effect surfaces for representations thatare humanlike to varying degrees how this effect of ap-pearance would apply to interactions with multiple hu-man representations of varying photographic realismremains an empirical question One key direction forfuture research is to examine how variation in photo-graphic realism (ie visual fidelity) of humanlike em-bodied agents affects copresence and how it interactswith behavioral realism Furthermore another impor-tant direction is to begin future studies with pretestedlevels of all realism variables (behavioral photographicand others) to assist in a more thorough interpretationof results Moreover the current study is limited in thatwe implemented behavioral realism solely by varyinghead movements A more thorough investigation of thisphenomenon should investigate other types of behav-iors such as facial gestures hand movements and gaitas well as more complex behaviors such as planning andcoordinated actions In addition the interaction task wechose (a purely nonverbal task) is limited and researchshould definitely examine copresence in richer social-interaction settings Finally our method of translationusing the game pad as opposed to naturally trackedwalking was not a perfect behavioral measure since

390 PRESENCE VOLUME 14 NUMBER 4

people do not translate by pushing buttons in the realworld Future work should provide a further examinationof copresence using more natural behavioral measures

Importantly had we measured only self-report re-sponses we would have demonstrated a different patternof results in the current study Attempting to augment anystudy geared toward understanding human interactionwith embodied agents or avatars with dependent measuresfrom multiple response channels certainly can make datainterpretation more complicated Patterns from self-reportdata that seem clean and consistent often are not replicatedwhen a less direct measurement paradigm is implementedFurthermore when asking peripheral implicitly relatedquestions about presence such as likability or embarrass-ment one risks measuring constructs inside the humanmind that are not ldquocentralrdquo to the definition of copres-ence Adding more implicit and disparate measures maynot paint a cleaner picture but it does paint a more com-plete one The challenge to the researcher is interpretingthese complex patterns in the data in order to assess amore thorough understanding of an extremely novel andabstract phenomenon of interacting with a virtual digitalentity

It may be the case that copresence is a purely affectiveconstruct inside the human mind alternatively copres-ence may be more of a cognitive construct In this earlystage of the research field examining copresence in vir-tual environments the theoretical and empirical work isextremely limited and does not allow one to rule outeither of these two (or many other) possibilities Deter-mining the exact conceptual nature of an abstract con-struct in the human mind is a challenge that faces manysocial scientists The measurement framework we offerallows the researcher to get a more complete picture ofthe copresence construct and allows for future researchto utilize more specific types of construct-type hypothe-ses as well as measurement strategies

In addition to the basic research implications with re-spect to the use of various modes of dependent measuresthe current study also has some important applied implica-tions By understanding the components of a virtual repre-sentation that lead to elicitations of heightened copresencewith that representation we can better develop our virtualagents and virtual worlds to immerse and engage the user

As we as well as others have demonstrated creating highlevels of copresence is a complicated process involving anappropriate fit between levels of each type of realism Aswe grow to better understand these interactions we canbetter match agent design to context and perhaps reach alevel where it becomes easy to use agents to develop thedegree of copresence usually found with virtual avatarsand perhaps in the long term the degree of copresencefound with flesh-and-blood people during face-to-faceinteraction

References

Bailenson J N Aharoni E Beall A C Guadagno R Di-mov A amp Blascovich J (2004) Comparing behavioraland self-report measures of embodied agents social presence inimmersive virtual environments Proceedings of the 7th An-nual International Workshop on Presence Valencia Spain

Bailenson J N Beall A C amp Blascovich J (2002) Mutualgaze and task performance in shared virtual environmentsJournal of Visualization and Computer Animation 13 1ndash8

Bailenson J N Beall A C Loomis J Blascovich J ampTurk M (2004) Transformed social interaction Decou-pling representation from behavior and form in collabora-tive virtual environments Presence Teleoperators and Vir-tual Environments 13(4) 428ndash441

Bailenson J N Blascovich J Beall A C amp Loomis J(2001) Equilibrium revisited Mutual gaze and personalspace in virtual environments Presence Teleoperators andVirtual Environments 10 583ndash598

Bailenson J N Blascovich J Beall A C amp Loomis J M(2003) Interpersonal distance in immersive virtual environ-ments Personality and Social Psychology Bulletin 29 1ndash 15

Bailenson J N amp Yee N (in press) Digital chameleonsAutomatic assimilation of nonverbal gestures in immersivevirtual environments Psychological Science

Beall A C Bailenson J N Loomis J Blascovich J ampRex C (2003) Non-zero-sum mutual gaze in collaborativevirtual environments Proceedings of HCI InternationalCrete

Bente G Ruggenberg S Tietz B amp Wortberg S (2004)Measuring behavioral correlates of social presence in virtualencounters Paper presented at the International Communi-cation Association Conference May 27ndash31

Biocca F (1997) Cyborgrsquos dilemma Progressive embodi-

Bailenson et al 391

teractive effects of the appearance and behavior of em-bodied agents Our multimethod approach includedself-report behavioral and cognitive dependent mea-sures

11 Defining and OperationalizingCopresence

One of the difficulties in understanding and syn-thesizing the accumulating literature on copresence isthe plethora of definitions and operalizations used byinvestigators as can be seen in the thorough discussionby Schroeder (2002) Unfortunately this difficulty hasled to inconsistent and often contradictory findings re-ported in the copresence literature Heeter (1992) forexample defined copresence (or social presence) as theextent to which other beings both living and syntheticexist in a virtual world and appear to react to humaninteractants Nowak (2001) on the other hand definedcopresence as the sense of connection with anothermind Russo (2001) argued that copresence (or whatshe called mediated presence) is the perception by a com-municator that another person in a mediated or onlineenvironment is real immediate or present ConverselySlater and his colleagues have defined copresence as thesense of being and acting with others in a virtual place(Slater Sadagic Usoh amp Schroeder 2000) FinallyBlascovich and his colleagues have defined copresence asthe extent to which individuals treat embodied agents asif they were other real human beings (Blascovich et al2002)

Although there is no consensual definition of cop-resence or agreement about how best to assess itexisting research on copresence can be roughly di-vided into two camps studies that focus on peoplersquosperceptions of embodied agents versus studies thatfocus on peoplersquos social responses to embodied agentsResearchers from the first camp (eg Heeter 1992Biocca 1997 Lombard amp Ditton 1997 GarauSlater Bee amp Sasse 2001 Lessiter Freeman Keoghamp Davidoff 2001 define copresence in terms of a us-errsquos perception or feeling that others are cosituatedwithin an interpersonal environment (Swinth amp Blas-covich 2002) Researchers in this

tradition take the view that copresence occurs when auser not only perceives or feels that he or she isldquopresentrdquo within a virtual environment but that heor she perceives or feels that others are present thereas well Because this view of copresence focuses onpeoplersquos phenomenological experiences researchersadopting this orientation frequently operationalizecopresence in terms of a userrsquos self-reported percep-tion andor feeling that they are not alone in a vir-tual environment

Another group of researchers however views copres-ence as a social task-related or physiological response toembodied agents (eg Bailenson et al 2004 Blasco-vich et al 2002 Reeves amp Nass 1996 BailensonBeall amp Blascovich 2002 Beall Bailenson LoomisBlascovich amp Rex 2003 Blascovich et al 2002 Lee ampNass 2004 Bente Ruggenberg Tietz amp Wortberg2004 Beall Bailenson Loomis Blascovich amp Rex2003 Meehan Insko Whitton amp Brooks 2004Slater 2004 Bente Ruggenberg Tietz amp Wortberg2004) Researchers adopting this orientation argue thatcopresence occurs when people treat embodied agentsas if they were other real people and they operationalizecopresence using a variety of methods including theanalysis of physiological data examination of a variety ofnonverbal and verbal behaviors or the assessment oftask performance cooperation interaction satisfactionliking or other social responses

In the current study participants were asked to exam-ine embodied agents that varied in behavior and appear-ance We then measured the extent to which the em-bodied agents were perceived and treated as if they werehuman via self-report behavioral responses and cogni-tive dependent measures Adopting this approach en-abled us to compare the two general ways of definingcopresence described above by examining both partici-pantsrsquo self-reported perceptions of embodied agents aswell as their self-reported cognitive and behavioral re-sponses to them Moreover by employing three distinctmeasurement modalities (self-report questionnaires be-havioral responses and cognitive responses) we wereable to distinguish converging and diverging patternspertinent to each of the two general ways of defining

380 PRESENCE VOLUME 14 NUMBER 4

copresence We discuss each of these response modali-ties in turn

12 Self-Report Markers of Copresence

In order to tap the phenomenological aspects ofcopresence we administered a self-report questionnairethat asked participants to report on the extent to whichthey perceived or felt like there was another social entitywith them in the virtual environment In addition weadministered two self-report measures designed to assessdifferent kinds of social responses people might have toembodied agents liking of an embodied agent and will-ingness to perform embarrassing acts in front of an em-bodied agent We hypothesized that all three indicatorsof copresence would be correlated positively for copres-ence and liking and negatively between copresenceliking and willingness to perform embarrassing acts

13 Behavioral Markers of Copresence

In order to assess the idea that one can infer co-presence by observing interaction with an embodiedagent we assessed a behavioral measure of copresenceinterpersonal distance Building on previous work (egReeves amp Nass 1996 Krikorian Lee Chock amp Harms2000 Bailenson Blascovich Beall amp Loomis 2001Bailenson Blascovich Beall amp Loomis 2003) we hy-pothesized that copresence would be positively relatedto hesitation in approaching an embodied agent inter-personal distance and maintenance of interpersonal dis-tance

14 Cognitive Markers of Copresence

Furthermore we assessed memory for embodiedagents Previous research has found that people remem-ber more features about people in scenes than aboutnonsentient objects in scenes (New 2003) Conse-quently users should remember more informationabout embodied agents represented as sentient beings(ie human) than about nonsentient types of represen-tations Thus we expected better memory for embodied-agent features However Hoyt Blascovich and Swinth

(2003) demonstrated that the presence of an embodiedagent worsened performance on a cognitive task whenparticipants performed in front of an audience of ava-tars and that perceived copresence mediated that effectConsequently it could be argued that copresence mayin fact be indicated by worse memory for informationrelating to an embodied agent In the current study weexamined usersrsquo memory for an embodied-agent featureas a cognitive marker of copresence on an exploratorybasis without predictions concerning the direction ofmemory effects Our goal in including this cognitivemarker of copresence was to compare any memory dif-ferences observed to other more established and pre-dictable markers of copresence in order to begin to un-derstand the cognitive effects of copresence if any

15 Factors Affecting Copresence

Although a number of factors may influence co-presence two appearance and behavior of embodiedagents have received considerable attention Althoughsome work has examined the interplay between howhuman an embodied agent appears and how an embod-ied agent behaves (eg Bailenson et al 2001 Garau etal 2003) researchers typically examine each factor inisolation or confound them as a single variable

Bailenson et al (2001) for example found that mak-ing an embodied agent more realistic looking by addinga texture map from an actual photograph of a face didnot increase copresence compared with an embodiedagent that was more cartoonlike as long as both typesof embodied agents demonstrated realistic gaze behav-iors Similarly Garau et al (2003) demonstrated thatmore photographically realistic avatars elicited no morecopresence than did less realistic-looking avatars and infact found that increasing the photo-realism of an ava-tar can actually cause a decrease in copresence if the be-havioral realism is not also increased correspondingly(see also Slater amp Steed 2001 Garau 2003)

In contrast Nowak and her colleagues (Nowak2001 Nowak amp Biocca 2001) demonstrated that theextent to which embodied agents resemble real humanbeings affects peoplersquos social judgments of agents as in-teraction partners and the amount of copresence experi-

Bailenson et al 381

enced In one study interaction partners represented byanthropomorphic images were rated as more sociallyattractive and they received higher partner satisfactionratings than did partners represented either by no imageor by an image that did not appear human (Nowak2001) In addition participants indicated that they feltgreater copresence and that the communication me-dium was better able to support the social interactionwhen interactants were represented by highly anthropo-morphic images compared to when no image or an im-age low in anthropomorphism was used (Nowak ampBiocca 2001) Similarly Parise Kiesler Sproull amp Wa-ters (1996) demonstrated that more human-lookingembodied agents elicited more liking behavior and morecooperation than did less human-looking embodiedagents

One possible explanation for these seemingly contra-dictory findings concerns previous manipulations of ap-pearance Studies that have demonstrated null effectshave typically manipulated the visual fidelity (or photo-realism) of embodied agents and studies that havefound a significant effect of embodied-agent appearanceon copresence and other social outcomes have typicallymanipulated the anthropomorphism of the embodiedagent or the extent to which the embodied agent ap-pears human in form We are unaware of studies thathave systematically varied both photo-realism and an-thropomorphism and unfortunately most studies thathave been conducted on the effect of embodied-agentappearance have tended to confound photo-realism andanthropomorphism

Given that studies focusing on the photo-realism ofan embodied agent have consistently reported null ef-fects however we attempted to control for the effect ofphoto-realism in the current study by using embodiedagents of equivalent visual fidelity while simultaneouslymanipulating how much like an actual human being theembodied agent both looked and behaved We pre-dicted that embodied agents that either looked humanor acted human would elicit greater copresence thanembodied agents that did not look or act human How-ever based on the work of Blascovich and his colleagues(Blascovich 2002 Blascovich et al 2002) we alsohypothesized that embodied-agent appearance and be-

havior would interact with each other such that the be-havioral realism of an embodied agent would mattermost for non-human-looking embodied agents

16 Summary and Study Objectives

In this study we systematically examined the inde-pendent and interactive effects of both the appearanceand behavior of embodied agents on usersrsquo perceptionsand social responses within an immersive virtual envi-ronment In doing so we attempted to address a num-ber of currently unresolved issues First we examinedhow different ways of defining and operationalizingcopresence are influenced by systematic manipulationsof appearance and behavior of virtual representations Indoing so we attempted to reconcile some of the con-trasting findings that have been reported in the researchliterature Second we examined whether similar or dis-parate patterns of response obtained across differentmodalities of copresence measurement (ie self-reportvs cognitive vs behavioral dependent measures) Finallywe used four different algorithms to control the headmovements of the embodied agent some of which havenever before been empirically compared in an immersivevirtual environment

2 Method

21 Participants

One hundred and forty-six undergraduates (73males 73 females) ranging in age from 18 to 27 years(M 1963 years) recruited from the University of Cal-ifornia Santa Barbara served as participants They wereeither paid or received class credit for participation

22 Design

A two-way between-subjects factorial design wasused The first factor representation type included threelevels of embodiment human teddy bear and block-head As can be seen in Figure 1 the three embodiedagents varied in the extent to which they resembled a

382 PRESENCE VOLUME 14 NUMBER 4

human being1 For female participants the human rep-resentation created using 3dMeNow software depictedin the left panel of Figure 1 was used For male partici-pants a similar-looking male representation was usedAll three representations possessed equivalent behavioralfunctionality within the virtual environment (see below)and exhibited similar levels of photo-realism

While we attempted to ensure as much as possibleequal levels of photo-realism (by using textures of simi-lar resolution and by ensuring that the underlying 3Dmesh for each model contained similar numbers of an-chor points and structural facial features) subjectivelythere may have been minor differences in photo-realism

The second between-subjects factor behavioral real-ism included four levels of head movements Four dif-ferent algorithms were employed to control the embod-ied agentrsquos head movements In the static condition theembodied agent remained motionless In the randommovement condition the embodied agentrsquos head contin-ually moved in trajectories generated by the computertoward random points within a 60-degree arc of eachside of its head (for both pitch and yaw) In the mimiccondition the embodied agentrsquos head movements mir-rored the participantrsquos actual head movements with a

4-s delay2 In the recorded movement condition headmovements recorded from a randomly selected partici-pant in a previous session were used to control the em-bodied agentrsquos head movements In all four conditionsthe embodied agent blinked its eyes according to a pre-set algorithm but there were no facial gestures exhib-ited Complete crossing of these two independent vari-ables resulted in 12 between-subjects conditions3

22 The Virtual Environment

The virtual room in which participants were im-mersed is depicted in Figure 2 Participants sat in achair at a table facing an embodied agent Participantswere able to see the head and shoulders of the agentThe same body was used for each representation type Atag containing a string of letters and numbers later usedas part of a surprise memory task was positioned on theshirt of the embodied agent

1 There are a variety of terms that researchers use to describe howreal or how human a virtual representation looks including anthropo-morphism visual realism and photographic (or photo-) realism Wehave chosen to use the more general term representation type as a vari-able name and talk about the manipulation as mainly pertaining tophotographic realism We acknowledge that other terms such as an-thropomorphism however may also be appropriate

2 Previous research using mimicked head movements has indicatedthat users generally do not detect embodied agents that are mimickingtheir head movements with a 4-second delay (Bailenson Beall Loo-mis Blascovich amp Turk 2004 Bailenson amp Yee in press) Howeverthis study is unique in that it directly compared the effect of mimickedhead movements with other head-movement algorithms

3 There were an unequal number of females and males across the12 conditions

Figure 1 Photographically realistic human teddy bear and

blockhead embodied agents

Figure 2 A participantrsquos view of the virtual room

Bailenson et al 383

23 Equipment

Participants wore a Virtual Research V8 stereo-scopic head-mounted display (HMD see Figure 3) TheHMD featured dual 680 horizontal by 480 vertical pixelresolution LCD panels that refreshed at 72 Hz Thedisplay optics presented a visual field subtending approxi-mately 50 degrees horizontally by 38 degrees vertically

Perspectively correct stereoscopic images were ren-dered by a 450 MHz Pentium III dual-processor com-puter with an Evans amp Sutherland Tornado 3000 dualpipe graphics card and were updated at an averageframe rate of 36 Hz The simulated viewpoint was con-tinually updated as a function of the participantsrsquo headmovements The orientation of the participantrsquos headwas tracked by a three-axis orientation sensing system(Intersense IS300 update rate of 150 Hz) The systemlatency or the amount of delay between a participantrsquoshead movement and the resulting concomitant updatein the HMDrsquos visual display was 65 ms maximum Thesoftware used to assimilate the rendering and trackingwas Vizard 20 Participants used a Logitech Rumble-

Pad Pro game pad to interact with the virtual environ-ment

24 Procedure

When participants arrived at the laboratory anexperimenter sat them down and instructed them onhow to wear and adjust the HMD and how to use thegame pad depicted in Figure 3 to interact with thevirtual environment Once immersed participantsfound themselves seated at a table directly acrossfrom an embodied agent (see Figure 2) Dependingon the assigned condition the embodied agent variedin appearance and behavior Using the game pad par-ticipants scrolled through the following instructionaltext that appeared above the embodied agentrsquos head

In this laboratory we are interested in peoplersquos reac-tions to ldquoothersrdquo in virtual environments In thisstudy you will spend time with an ldquootherrdquo After ashort time we will ask you questions about theldquootherrdquo You will use the game pad to answer thequestions The questions measure your agreementThe more you agree the higher the number youshould choose on the rating scale in front of you

We carefully chose and repeatedly used the wordldquootherrdquo set off by quotation marks to avoid using sug-gestive nouns such as ldquopersonrdquo or ldquoagentrdquo Participantswere next instructed to look directly at the ldquootherrdquo for90 s Next a series of 10 copresence embarrassment andlikability self-report questionnaire items (see AppendixA) appeared one at a time over the embodied agentrsquoshead in random order Participants indicated theiragreement with these items by using the game pad toselect a response on a Likert-type rating scale that ap-peared under each item Hence ratings were providedin real time as participants experienced the embodiedagent

After completing the items participants were in-structed to approach the embodied agent for furtherexamination They used buttons on the game pad tomove toward (positive translation on the z axis) or awayfrom (negative translation on the z axis) the embodiedagent with each press of a game-pad button equivalentto a jump of 25 cm (the starting distance between the

Figure 3 A depiction of our immersive virtual environment system

The components are (1) orientation tracking sensor (2) image

generator (3) HMD and (4) game pad input device

384 PRESENCE VOLUME 14 NUMBER 4

participantrsquos nose and the embodied agentrsquos nose wasapproximately 3 m) We did not render the body of par-ticipants so there was no leaning or walking animationduring the translationmdashonly their viewpoint movedcloser or farther from the agent After participants fin-ished approaching the embodied agent and examiningit they took off the HMD and filled out a demographicquestionnaire Finally we administered a surprise mem-ory test in which participants were asked to recall therandom string of characters written on the embodiedagentrsquos chest (see Figure 2)

25 Measures

251 Perceived Copresence Perceived copres-ence was assessed via a three-item self-report measureWe created a composite copresence measure ( 71see Appendix A) by averaging the three copresencequestions Higher scores indicated higher levels ofself-reported copresence where ratings were made on aLikert-type scale with response options ranging from 0(strongly disagree) to 6 (strongly agree)

252 Embarrassment Embarrassment was as-sessed as a self-reported social response to the presenceof an embodied agent Previous research has demon-strated that willingness to perform embarrassing acts infront of a virtual representation is a worthy measure ofthat representationrsquos degree of social influence (Bailen-son et al 2003) We averaged the three embarrassmentquestions ( 72 see Appendix A) into a single com-posite measure with higher numbers indicating greaterwillingness to perform embarrassing acts in front of anembodied agent

253 Likability Likability was assessed as a sec-ond self-reported social response to the presence ofan embodied agent We averaged the four likabilityquestions ( 71 see Appendix A) into a singlecomposite measure with higher numbers indicatingthat participants liked the virtual representation more

254 Memory Memory was quantified asthe total number of correct characters (out of a possi-

ble eight) a participant recalled from the virtual rep-resentationrsquos label without penalty for incorrect an-swers

255 Interpersonal Distance Interpersonaldistance was included as a behavioral social responseto the presence of an embodied agent We assessedtwo indicators of interpersonal distanceminimumdistance and reversal count Minimum distance wasthe closest distance the participant came to the em-bodied agent reversal count referred to the numberof times a participant changed direction along thez-axis (ie went back and forth) during their exami-nation of the embodied agent We consider reversalcount to be a proxy for hesitation in approaching theembodied agent

3 Results and Discussion

31 Hypotheses

We hypothesized that greater copresence wouldbe indicated by higher perceived copresence ratingsgreater liking of the embodied agent less willingness tocommit embarrassing acts in front of the embodiedagent the maintenance of greater interpersonal dis-tance and more hesitation in approaching the embod-ied agent We did not however have any specific pre-dictions with respect to the relationship betweencopresence and memory

Based on the results from previous copresencestudies we expected greater copresence (across ourvarious copresence measures) to occur with realisti-cally behaving embodied agents (specifically our re-corded and mimic behavioral-realism conditions thatfeatured movements based on actual human motion)compared with unrealistically behaving embodiedagents (ie our static and random-movement condi-tions) In addition we expected greater copresence tooccur with more human-looking embodied agentsthan with less human-looking ones Drawing on pre-dictions derived from the Blascovich et al (2002)threshold model of social influence however we ex-pected embodied-agent appearance to moderate the

Bailenson et al 385

effect of behavioral realism such that behavioral real-ism would matter more for less human-looking em-bodied agents than for more human-looking embod-ied agents Finally we expected to demonstrateconvergence in our dependent measures such thatthey would correlate with one another and change inthe same way as a function of our independent vari-able manipulations

32 Perceived Copresence Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movement vs mimic vs recorded) analysisof covariance controlling for participant gender4 re-vealed a significant main effect of behavioral realism onself-reported perceived copresence F (3 133) 272p 05 2 06 As Figure 4 demonstrates perceivedcopresence was higher in the random-movements con-dition compared to the static condition p 01 How-ever this main effect of behavioral realism was qualifiedby a significant representation type by behavioral-realisminteraction F (6 133) 218 p 05 2 09

Simple main effects tests revealed that within theblockhead condition perceived copresence varied signif-icantly as a function of behavioral realism F (3 133)

339 p 05 2 07 Specifically perceived copres-ence was significantly higher when the blockhead repre-sentation moved according to our random-movementalgorithm compared with either the static or mimic con-ditions ps 05 Simple main effects tests also revealedthat perceived copresence varied as a function of behav-

ioral realism for participants in the teddy-bear condi-tion F (3 133) 330 p 05 2 07 Specificallyperceived copresence was significantly higher in mimiccondition than in either the static or recorded condi-tions ps 05 and higher in the random-movementscondition relative to the recorded and static conditions(ps 05 and 10 respectively)

In contrast there were no significant differences inperceived copresence between any of the four behavioral-realism conditions for participants in the humanldquootherrdquo condition This pattern of findings is consis-tent with our prediction that behavioral realismwould have less of an influence on perceived copres-ence for embodied agents that appear human ascompared to embodied agents with a nonhuman ap-pearance In addition it is interesting to note that thetype of head movement that elicited the greatest levelof perceived copresence varied as a function of the em-bodied agentrsquos appearance with participants attributinghigh copresence ratings to both mimicking and ran-domly moving teddy-bear representations but only ran-domly moving blockhead representations

4 Although both male and female participants were used in thisstudy gender was not included as a separate factor in analyses involv-ing both behavioral realism and representation type When randomlyassigning participants to conditions we did not block participants bygender and as a result when behavioral realism and representationtype were fully crossed there were not enough participants of bothgenders in all 12 experimental conditions to reliably compute a three-way analysis of variance Thus our data analytic technique involvedincluding participant gender as a covariate in our analyses involvingboth behavioral realism and representation type in order to control forthe potential effect of participant gender The results do not changeappreciably if we remove gender as a covariate in our analyses but weinclude gender as a covariate to provide the most thorough analysispossible

Figure 4 Mean perceived-copresence ratings (using gender-

covariate adjusted means) as a function of representation type and

behavioral realism Conditions that share the same letter do not differ

within that level of representation type at p 05

386 PRESENCE VOLUME 14 NUMBER 4

33 Embarrassment Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender re-vealed a significant main effect of representation type onparticipantsrsquo self-reported willingness to perform embar-rassing acts in front of the embodied agent (ie embar-rassment) F (2 133) 421 p 05 2 06 Spe-cifically participants in the human embodied-agentcondition (M 260 SD 155) were significantlyless willing to perform embarrassing acts in front ofthe ldquootherrdquo than were participants in the teddy-bear(M 351 SD 154) embodied-agent conditionp 05 Participants in the human-representationcondition were also marginally less willing to performembarrassing acts than participants in the blockhead(M 313 SD 156) condition p 10 In sumthe presence of a human-looking representationseemed to reduce participantsrsquo willingness to performin embarrassing acts Regardless of the behavioral re-alism of the representation embodied agents wholooked like humans elicited greater social influencethan did non-human-looking representations

34 Likability Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender alsorevealed a significant main effect of representation typeon self-reported likability ratings F (2 133) 543p 01 2 08 Specifically both the human-lookingrepresentation (M 250 SD 109) and the teddy-bear representation (M 228 SD 109) were ratedas significantly more likable than the blockhead repre-sentation (M 179 SD 110) ps 05 Partici-pants found the blockhead representation (which wasthe least human looking of the three representations) tobe less attractive than either of the other two represen-tation types and they liked the blockhead representationthe least

35 Correlations between PerceivedCopresence Embarrassment andLikability Ratings

In order to examine the relationships among per-ceived copresence embarrassment and likability a se-ries of Pearson correlations was computed Consistentwith out predictions results of these analyses revealed asignificant negative correlation between perceived co-presence and a participantrsquos willingness to perform em-barrassing acts in front of the embodied agent r(144) 27 p 001 as well as a significant positivecorrelation between perceived copresence and likabilityr (144) 31 p 001 Specifically we found that thehigher the perceived copresence experienced the lesswilling participants were to perform embarrassing acts infront of the embodied agent (r2 07) and the moreparticipants liked the embodied agent (r2 10) Therewas no significant correlation between likability and em-barrassment ratings

36 Memory

We computed a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender to ex-amine the effects of our independent variables on partic-ipantsrsquo memory Results of these analyses revealed a sig-nificant main effect of behavioral realism on memory F(3 133) 414 p 01 2 09 Specifically mem-ory was significantly worse in the mimic condition thanin either the static or random-movement conditionsps 05 suggesting that overall participants had theworst memory for characters worn by a representationthat mimicked their own movements with a 4-s delayAgain however this main effect was qualified by a sig-nificant representation type by behavioral-realism inter-action (see Figure 5) F (6 133) 270 p 02 2

11Simple main effects tests revealed that within the

blockhead condition memory varied significantly as afunction of behavioral realism F (3 133) 271p 05 2 06 Specifically individuals rememberedsignificantly more characters when they were in the

Bailenson et al 387

random-movement condition than in either the mimicor the recorded conditions (the two behavioral-realismconditions that appeared most humanlike) ps 05and participants in the random-movement conditionremembered marginally more than participants in thestatic condition p 10 These results converge quitewell with the perceived copresence data in which partic-ipants indicated higher copresence ratings for blockheadrepresentations only when the blockhead demonstratedrandom head movements Interestingly this suggeststhat for representations who look the least human theleast human-looking head movements were seen asmost appropriate This finding that avoiding mis-matches is the most effective manner of combining anembodied agentrsquos appearance and behavior is consistentwith previous research on visually embodied agents(Slater amp Steed 2001 Garau et al 2003) as well asvoice agents (Lee amp Nass 2003)

Furthermore the convergence of memory andperceived-copresence ratings provides some evidencethat better memory is indicative of higher copresenceIn other words the more copresence an embodiedagent elicits the more a person interacting with that

embodied agent remembers about it It is interesting tonote however that despite the similar patterns of find-ings between the perceived-copresence data and thememory data results of a Pearson correlation revealedthat there was no significant linear relationship betweenperceived copresence and peoplersquos ability to rememberinformation about an embodied agent

There was however a significant simple main effectof behavioral realism within the human-representationcondition F (3 133) 549 p 001 2 11Specifically participants in the mimic and random-movement conditions remembered significantly fewercharacters than participants in either the static or re-corded conditions ps 05 Participants in the teddy-bear condition displayed the same pattern but in thatcondition none of the differences reached statisticalsignificance

These results suggest that participants may have beendistracted by a human-looking embodied agent thatmimicked their own movements which in turn af-fected their ability to notice and recall information inthe virtual environment Moreover participants mayalso have been distracted in the random-movement con-dition since this condition represented the greatest mis-match between the appearance (human looking) andbehavior (not human looking) of the embodied agentFurther research should examine these potential expla-nations as they suggest that different processes maybe involved for different algorithms for generatingembodied-agent head movements

37 Interpersonal Distance

There were no significant effects of behavioral re-alism or representation type on minimum distance (ourfirst measure of interpersonal distance) Thus partici-pants in all conditions maintained equivalent levels ofinterpersonal distance between themselves and the em-bodied agent There was also no significant correlationbetween perceived copresence and either of our behav-ioral dependent measures (ie minimum distance andnumber of returns)

However as hypothesized results of a 3 (representa-tion type human vs teddy bear vs blockhead) 4 (be-

Figure 5 Gender-covariate adjusted mean number of correctly

recalled characters out of eight on a surprise memory test as a

function of representation type and behavioral realism Conditions that

share the same letter do not differ within that level of representation

type at p 05

388 PRESENCE VOLUME 14 NUMBER 4

havioral realism static vs random movements vs mimicvs recorded) analysis of covariance controlling for par-ticipant gender revealed a significant main effect of be-havioral realism on reversal count or the number oftimes a participant changed direction along the z-axisduring his or her examination of the embodied agentF(3 133) 306 p 05 2 07 As can be seen inFigure 6 participants in the random-movement condi-tion exhibited significantly fewer reversals than partici-pants in any of the other three behavioral-realism condi-tions Pairwise comparisons indicated that there weresignificantly fewer reversals in the random-movementcondition compared to the other three conditions(ps 05) which are not significantly different fromeach other

This finding in part confirms our hypothesis con-cerning our behavioral dependent variable For an em-bodied agent whose head moves with curves randomlygenerated by the computer participants show little tono hesitation in approaching In contrast when thehead movements were controlled by a real person (ei-ther a mimic of their own movements or the move-ments from another participant) participants showedgreater hesitation in approaching the embodied agentThus computer-generated movements generated lesshesitancy than human-generated movements

Given this explanation however one might expecteven less hesitation that is fewer reversals in approach-ing the static representation This finding did not occurand one potential explanation is that overall a ran-domly moving representation may elicit less social influ-ence and copresence than a completely static representa-tion In other words the static representation has atleast the potential to exhibit realistic movements whilethe unrealistically moving embodied agent immediatelylowers the sense of copresence In sum participantsrsquohesitancy in approaching an embodied agent was influ-enced by the behavioral realism of the embodied agentbut not its appearance

4 Conclusions

41 Summary of Findings

To summarize the appearance and behavior ofembodied agents had varying effects on the dependentvariables In many instances there was convergence inthat self-report cognitive and behavioral measureschanged in similar ways as a function of our indepen-dent variable manipulations For example the perceived-copresence likability and embarrassment measures cor-related in the predicted directions and the blockheadrepresentation elicited similar reactions across behaviortypes for both the memory measure and the perceived-copresence measure In other instances however therewas a lack of convergence where varying realism pro-duced different outcomes for the various responsemodes One inevitable challenge facing researchers whoexamine a large number of dependent variables is tounderstand the instances in which those dependentvariables do and do not converge

In the current study for some of our dependent mea-sures of copresence (willingness to commit embarrassingacts and likability) manipulation of representation typeaffected copresence responses On other dependentmeasures (self-reported copresence ratings interper-sonal distance approach reversals and memory) ma-nipulation of behavioral realism changed participantsrsquoresponses One way to interpret these results is thatself-report measures were adequate for measuring

Figure 6 Mean number of direction reversals while approaching the

embodied agent as a function of behavioral realism

Bailenson et al 389

copresence levels as affected by variation in appear-ance whereas self-report measures were not sensitiveenough to detect the subtle differences elicited by thebehavioral-realism manipulation Indeed the only reliablemain effects of the behavioral-realism manipulation oc-curred with cognitive and behavioral dependent measures

In other words self-report questionnaires are effectivewhen measuring how people perceive an embodiedagent but not necessarily how they will respond to thatembodied agent Consequently not only does it matterhow one defines copresence it also matters how oneoperationalizes copresence If the goal of a researcher isto understand the way in which people fathom embod-ied agents on a conceptual level then self-report ques-tionnaires may be the best tool However if the goal ofthe researcher is to understand the way in which peoplewill interact with learn from or behave toward anagent then he or she should augment questionnaireswith other types of measures In this paper we are notlobbying for the abandonment of copresence self-reportquestionnaires however we do urge that researchersuse these questionnaires as a complement to other typesof measures whenever possible

While this explanation that self-report measures lacksensitivity receives support from the current data itmust be qualified On most measures there was an in-teraction between behavioral realism and representationtype At times these interactions did not conform toour specific predictions and were difficult to interpretPart of the reason for this difficulty is most likely due tothe fact that we assumed a priori that certain types ofhead movements would be viewed as more realistic thanothers Specifically we assumed that the measures basedon actual human behaviors (mimic and recorded) wouldappear more realistic than either those generated by thecomputer (random) or than no movements at all Re-corded human movements applied to 3D models how-ever may not always be the most effective method ofrendering realistic behaviors and the differences be-tween our various levels of behavioral realism may havebeen too subtle in the current study As previous workhas demonstrated (Garau et al 2003) the interactionbetween appearance and behavioral realism is compli-cated and the finding that seems to be emerging is that

large disparities between the two types of realism cancause diminished levels of copresence Only when theblockhead representation the least humanlike of thethree demonstrated unrealistic random head move-ments did it elicit high levels of self-report and cognitivecopresence

Our findings with respect to the main effects of ap-pearance on self-report measures suggest that how hu-man an embodied agent looks may also be an importantprecondition for experiencing copresence This findingis unanticipated given predictions made by the Blasco-vich model and given findings from previous work(Bailenson et al 2001 Garau et al 2003) It seemsthat virtual representations that look like a realistic hu-man as opposed to either a familiar or unfamiliar non-human elicit embarrassment reactions that would beappropriate only in the presence of a real human

42 Limitations and Future Research

Though this effect surfaces for representations thatare humanlike to varying degrees how this effect of ap-pearance would apply to interactions with multiple hu-man representations of varying photographic realismremains an empirical question One key direction forfuture research is to examine how variation in photo-graphic realism (ie visual fidelity) of humanlike em-bodied agents affects copresence and how it interactswith behavioral realism Furthermore another impor-tant direction is to begin future studies with pretestedlevels of all realism variables (behavioral photographicand others) to assist in a more thorough interpretationof results Moreover the current study is limited in thatwe implemented behavioral realism solely by varyinghead movements A more thorough investigation of thisphenomenon should investigate other types of behav-iors such as facial gestures hand movements and gaitas well as more complex behaviors such as planning andcoordinated actions In addition the interaction task wechose (a purely nonverbal task) is limited and researchshould definitely examine copresence in richer social-interaction settings Finally our method of translationusing the game pad as opposed to naturally trackedwalking was not a perfect behavioral measure since

390 PRESENCE VOLUME 14 NUMBER 4

people do not translate by pushing buttons in the realworld Future work should provide a further examinationof copresence using more natural behavioral measures

Importantly had we measured only self-report re-sponses we would have demonstrated a different patternof results in the current study Attempting to augment anystudy geared toward understanding human interactionwith embodied agents or avatars with dependent measuresfrom multiple response channels certainly can make datainterpretation more complicated Patterns from self-reportdata that seem clean and consistent often are not replicatedwhen a less direct measurement paradigm is implementedFurthermore when asking peripheral implicitly relatedquestions about presence such as likability or embarrass-ment one risks measuring constructs inside the humanmind that are not ldquocentralrdquo to the definition of copres-ence Adding more implicit and disparate measures maynot paint a cleaner picture but it does paint a more com-plete one The challenge to the researcher is interpretingthese complex patterns in the data in order to assess amore thorough understanding of an extremely novel andabstract phenomenon of interacting with a virtual digitalentity

It may be the case that copresence is a purely affectiveconstruct inside the human mind alternatively copres-ence may be more of a cognitive construct In this earlystage of the research field examining copresence in vir-tual environments the theoretical and empirical work isextremely limited and does not allow one to rule outeither of these two (or many other) possibilities Deter-mining the exact conceptual nature of an abstract con-struct in the human mind is a challenge that faces manysocial scientists The measurement framework we offerallows the researcher to get a more complete picture ofthe copresence construct and allows for future researchto utilize more specific types of construct-type hypothe-ses as well as measurement strategies

In addition to the basic research implications with re-spect to the use of various modes of dependent measuresthe current study also has some important applied implica-tions By understanding the components of a virtual repre-sentation that lead to elicitations of heightened copresencewith that representation we can better develop our virtualagents and virtual worlds to immerse and engage the user

As we as well as others have demonstrated creating highlevels of copresence is a complicated process involving anappropriate fit between levels of each type of realism Aswe grow to better understand these interactions we canbetter match agent design to context and perhaps reach alevel where it becomes easy to use agents to develop thedegree of copresence usually found with virtual avatarsand perhaps in the long term the degree of copresencefound with flesh-and-blood people during face-to-faceinteraction

References

Bailenson J N Aharoni E Beall A C Guadagno R Di-mov A amp Blascovich J (2004) Comparing behavioraland self-report measures of embodied agents social presence inimmersive virtual environments Proceedings of the 7th An-nual International Workshop on Presence Valencia Spain

Bailenson J N Beall A C amp Blascovich J (2002) Mutualgaze and task performance in shared virtual environmentsJournal of Visualization and Computer Animation 13 1ndash8

Bailenson J N Beall A C Loomis J Blascovich J ampTurk M (2004) Transformed social interaction Decou-pling representation from behavior and form in collabora-tive virtual environments Presence Teleoperators and Vir-tual Environments 13(4) 428ndash441

Bailenson J N Blascovich J Beall A C amp Loomis J(2001) Equilibrium revisited Mutual gaze and personalspace in virtual environments Presence Teleoperators andVirtual Environments 10 583ndash598

Bailenson J N Blascovich J Beall A C amp Loomis J M(2003) Interpersonal distance in immersive virtual environ-ments Personality and Social Psychology Bulletin 29 1ndash 15

Bailenson J N amp Yee N (in press) Digital chameleonsAutomatic assimilation of nonverbal gestures in immersivevirtual environments Psychological Science

Beall A C Bailenson J N Loomis J Blascovich J ampRex C (2003) Non-zero-sum mutual gaze in collaborativevirtual environments Proceedings of HCI InternationalCrete

Bente G Ruggenberg S Tietz B amp Wortberg S (2004)Measuring behavioral correlates of social presence in virtualencounters Paper presented at the International Communi-cation Association Conference May 27ndash31

Biocca F (1997) Cyborgrsquos dilemma Progressive embodi-

Bailenson et al 391

copresence We discuss each of these response modali-ties in turn

12 Self-Report Markers of Copresence

In order to tap the phenomenological aspects ofcopresence we administered a self-report questionnairethat asked participants to report on the extent to whichthey perceived or felt like there was another social entitywith them in the virtual environment In addition weadministered two self-report measures designed to assessdifferent kinds of social responses people might have toembodied agents liking of an embodied agent and will-ingness to perform embarrassing acts in front of an em-bodied agent We hypothesized that all three indicatorsof copresence would be correlated positively for copres-ence and liking and negatively between copresenceliking and willingness to perform embarrassing acts

13 Behavioral Markers of Copresence

In order to assess the idea that one can infer co-presence by observing interaction with an embodiedagent we assessed a behavioral measure of copresenceinterpersonal distance Building on previous work (egReeves amp Nass 1996 Krikorian Lee Chock amp Harms2000 Bailenson Blascovich Beall amp Loomis 2001Bailenson Blascovich Beall amp Loomis 2003) we hy-pothesized that copresence would be positively relatedto hesitation in approaching an embodied agent inter-personal distance and maintenance of interpersonal dis-tance

14 Cognitive Markers of Copresence

Furthermore we assessed memory for embodiedagents Previous research has found that people remem-ber more features about people in scenes than aboutnonsentient objects in scenes (New 2003) Conse-quently users should remember more informationabout embodied agents represented as sentient beings(ie human) than about nonsentient types of represen-tations Thus we expected better memory for embodied-agent features However Hoyt Blascovich and Swinth

(2003) demonstrated that the presence of an embodiedagent worsened performance on a cognitive task whenparticipants performed in front of an audience of ava-tars and that perceived copresence mediated that effectConsequently it could be argued that copresence mayin fact be indicated by worse memory for informationrelating to an embodied agent In the current study weexamined usersrsquo memory for an embodied-agent featureas a cognitive marker of copresence on an exploratorybasis without predictions concerning the direction ofmemory effects Our goal in including this cognitivemarker of copresence was to compare any memory dif-ferences observed to other more established and pre-dictable markers of copresence in order to begin to un-derstand the cognitive effects of copresence if any

15 Factors Affecting Copresence

Although a number of factors may influence co-presence two appearance and behavior of embodiedagents have received considerable attention Althoughsome work has examined the interplay between howhuman an embodied agent appears and how an embod-ied agent behaves (eg Bailenson et al 2001 Garau etal 2003) researchers typically examine each factor inisolation or confound them as a single variable

Bailenson et al (2001) for example found that mak-ing an embodied agent more realistic looking by addinga texture map from an actual photograph of a face didnot increase copresence compared with an embodiedagent that was more cartoonlike as long as both typesof embodied agents demonstrated realistic gaze behav-iors Similarly Garau et al (2003) demonstrated thatmore photographically realistic avatars elicited no morecopresence than did less realistic-looking avatars and infact found that increasing the photo-realism of an ava-tar can actually cause a decrease in copresence if the be-havioral realism is not also increased correspondingly(see also Slater amp Steed 2001 Garau 2003)

In contrast Nowak and her colleagues (Nowak2001 Nowak amp Biocca 2001) demonstrated that theextent to which embodied agents resemble real humanbeings affects peoplersquos social judgments of agents as in-teraction partners and the amount of copresence experi-

Bailenson et al 381

enced In one study interaction partners represented byanthropomorphic images were rated as more sociallyattractive and they received higher partner satisfactionratings than did partners represented either by no imageor by an image that did not appear human (Nowak2001) In addition participants indicated that they feltgreater copresence and that the communication me-dium was better able to support the social interactionwhen interactants were represented by highly anthropo-morphic images compared to when no image or an im-age low in anthropomorphism was used (Nowak ampBiocca 2001) Similarly Parise Kiesler Sproull amp Wa-ters (1996) demonstrated that more human-lookingembodied agents elicited more liking behavior and morecooperation than did less human-looking embodiedagents

One possible explanation for these seemingly contra-dictory findings concerns previous manipulations of ap-pearance Studies that have demonstrated null effectshave typically manipulated the visual fidelity (or photo-realism) of embodied agents and studies that havefound a significant effect of embodied-agent appearanceon copresence and other social outcomes have typicallymanipulated the anthropomorphism of the embodiedagent or the extent to which the embodied agent ap-pears human in form We are unaware of studies thathave systematically varied both photo-realism and an-thropomorphism and unfortunately most studies thathave been conducted on the effect of embodied-agentappearance have tended to confound photo-realism andanthropomorphism

Given that studies focusing on the photo-realism ofan embodied agent have consistently reported null ef-fects however we attempted to control for the effect ofphoto-realism in the current study by using embodiedagents of equivalent visual fidelity while simultaneouslymanipulating how much like an actual human being theembodied agent both looked and behaved We pre-dicted that embodied agents that either looked humanor acted human would elicit greater copresence thanembodied agents that did not look or act human How-ever based on the work of Blascovich and his colleagues(Blascovich 2002 Blascovich et al 2002) we alsohypothesized that embodied-agent appearance and be-

havior would interact with each other such that the be-havioral realism of an embodied agent would mattermost for non-human-looking embodied agents

16 Summary and Study Objectives

In this study we systematically examined the inde-pendent and interactive effects of both the appearanceand behavior of embodied agents on usersrsquo perceptionsand social responses within an immersive virtual envi-ronment In doing so we attempted to address a num-ber of currently unresolved issues First we examinedhow different ways of defining and operationalizingcopresence are influenced by systematic manipulationsof appearance and behavior of virtual representations Indoing so we attempted to reconcile some of the con-trasting findings that have been reported in the researchliterature Second we examined whether similar or dis-parate patterns of response obtained across differentmodalities of copresence measurement (ie self-reportvs cognitive vs behavioral dependent measures) Finallywe used four different algorithms to control the headmovements of the embodied agent some of which havenever before been empirically compared in an immersivevirtual environment

2 Method

21 Participants

One hundred and forty-six undergraduates (73males 73 females) ranging in age from 18 to 27 years(M 1963 years) recruited from the University of Cal-ifornia Santa Barbara served as participants They wereeither paid or received class credit for participation

22 Design

A two-way between-subjects factorial design wasused The first factor representation type included threelevels of embodiment human teddy bear and block-head As can be seen in Figure 1 the three embodiedagents varied in the extent to which they resembled a

382 PRESENCE VOLUME 14 NUMBER 4

human being1 For female participants the human rep-resentation created using 3dMeNow software depictedin the left panel of Figure 1 was used For male partici-pants a similar-looking male representation was usedAll three representations possessed equivalent behavioralfunctionality within the virtual environment (see below)and exhibited similar levels of photo-realism

While we attempted to ensure as much as possibleequal levels of photo-realism (by using textures of simi-lar resolution and by ensuring that the underlying 3Dmesh for each model contained similar numbers of an-chor points and structural facial features) subjectivelythere may have been minor differences in photo-realism

The second between-subjects factor behavioral real-ism included four levels of head movements Four dif-ferent algorithms were employed to control the embod-ied agentrsquos head movements In the static condition theembodied agent remained motionless In the randommovement condition the embodied agentrsquos head contin-ually moved in trajectories generated by the computertoward random points within a 60-degree arc of eachside of its head (for both pitch and yaw) In the mimiccondition the embodied agentrsquos head movements mir-rored the participantrsquos actual head movements with a

4-s delay2 In the recorded movement condition headmovements recorded from a randomly selected partici-pant in a previous session were used to control the em-bodied agentrsquos head movements In all four conditionsthe embodied agent blinked its eyes according to a pre-set algorithm but there were no facial gestures exhib-ited Complete crossing of these two independent vari-ables resulted in 12 between-subjects conditions3

22 The Virtual Environment

The virtual room in which participants were im-mersed is depicted in Figure 2 Participants sat in achair at a table facing an embodied agent Participantswere able to see the head and shoulders of the agentThe same body was used for each representation type Atag containing a string of letters and numbers later usedas part of a surprise memory task was positioned on theshirt of the embodied agent

1 There are a variety of terms that researchers use to describe howreal or how human a virtual representation looks including anthropo-morphism visual realism and photographic (or photo-) realism Wehave chosen to use the more general term representation type as a vari-able name and talk about the manipulation as mainly pertaining tophotographic realism We acknowledge that other terms such as an-thropomorphism however may also be appropriate

2 Previous research using mimicked head movements has indicatedthat users generally do not detect embodied agents that are mimickingtheir head movements with a 4-second delay (Bailenson Beall Loo-mis Blascovich amp Turk 2004 Bailenson amp Yee in press) Howeverthis study is unique in that it directly compared the effect of mimickedhead movements with other head-movement algorithms

3 There were an unequal number of females and males across the12 conditions

Figure 1 Photographically realistic human teddy bear and

blockhead embodied agents

Figure 2 A participantrsquos view of the virtual room

Bailenson et al 383

23 Equipment

Participants wore a Virtual Research V8 stereo-scopic head-mounted display (HMD see Figure 3) TheHMD featured dual 680 horizontal by 480 vertical pixelresolution LCD panels that refreshed at 72 Hz Thedisplay optics presented a visual field subtending approxi-mately 50 degrees horizontally by 38 degrees vertically

Perspectively correct stereoscopic images were ren-dered by a 450 MHz Pentium III dual-processor com-puter with an Evans amp Sutherland Tornado 3000 dualpipe graphics card and were updated at an averageframe rate of 36 Hz The simulated viewpoint was con-tinually updated as a function of the participantsrsquo headmovements The orientation of the participantrsquos headwas tracked by a three-axis orientation sensing system(Intersense IS300 update rate of 150 Hz) The systemlatency or the amount of delay between a participantrsquoshead movement and the resulting concomitant updatein the HMDrsquos visual display was 65 ms maximum Thesoftware used to assimilate the rendering and trackingwas Vizard 20 Participants used a Logitech Rumble-

Pad Pro game pad to interact with the virtual environ-ment

24 Procedure

When participants arrived at the laboratory anexperimenter sat them down and instructed them onhow to wear and adjust the HMD and how to use thegame pad depicted in Figure 3 to interact with thevirtual environment Once immersed participantsfound themselves seated at a table directly acrossfrom an embodied agent (see Figure 2) Dependingon the assigned condition the embodied agent variedin appearance and behavior Using the game pad par-ticipants scrolled through the following instructionaltext that appeared above the embodied agentrsquos head

In this laboratory we are interested in peoplersquos reac-tions to ldquoothersrdquo in virtual environments In thisstudy you will spend time with an ldquootherrdquo After ashort time we will ask you questions about theldquootherrdquo You will use the game pad to answer thequestions The questions measure your agreementThe more you agree the higher the number youshould choose on the rating scale in front of you

We carefully chose and repeatedly used the wordldquootherrdquo set off by quotation marks to avoid using sug-gestive nouns such as ldquopersonrdquo or ldquoagentrdquo Participantswere next instructed to look directly at the ldquootherrdquo for90 s Next a series of 10 copresence embarrassment andlikability self-report questionnaire items (see AppendixA) appeared one at a time over the embodied agentrsquoshead in random order Participants indicated theiragreement with these items by using the game pad toselect a response on a Likert-type rating scale that ap-peared under each item Hence ratings were providedin real time as participants experienced the embodiedagent

After completing the items participants were in-structed to approach the embodied agent for furtherexamination They used buttons on the game pad tomove toward (positive translation on the z axis) or awayfrom (negative translation on the z axis) the embodiedagent with each press of a game-pad button equivalentto a jump of 25 cm (the starting distance between the

Figure 3 A depiction of our immersive virtual environment system

The components are (1) orientation tracking sensor (2) image

generator (3) HMD and (4) game pad input device

384 PRESENCE VOLUME 14 NUMBER 4

participantrsquos nose and the embodied agentrsquos nose wasapproximately 3 m) We did not render the body of par-ticipants so there was no leaning or walking animationduring the translationmdashonly their viewpoint movedcloser or farther from the agent After participants fin-ished approaching the embodied agent and examiningit they took off the HMD and filled out a demographicquestionnaire Finally we administered a surprise mem-ory test in which participants were asked to recall therandom string of characters written on the embodiedagentrsquos chest (see Figure 2)

25 Measures

251 Perceived Copresence Perceived copres-ence was assessed via a three-item self-report measureWe created a composite copresence measure ( 71see Appendix A) by averaging the three copresencequestions Higher scores indicated higher levels ofself-reported copresence where ratings were made on aLikert-type scale with response options ranging from 0(strongly disagree) to 6 (strongly agree)

252 Embarrassment Embarrassment was as-sessed as a self-reported social response to the presenceof an embodied agent Previous research has demon-strated that willingness to perform embarrassing acts infront of a virtual representation is a worthy measure ofthat representationrsquos degree of social influence (Bailen-son et al 2003) We averaged the three embarrassmentquestions ( 72 see Appendix A) into a single com-posite measure with higher numbers indicating greaterwillingness to perform embarrassing acts in front of anembodied agent

253 Likability Likability was assessed as a sec-ond self-reported social response to the presence ofan embodied agent We averaged the four likabilityquestions ( 71 see Appendix A) into a singlecomposite measure with higher numbers indicatingthat participants liked the virtual representation more

254 Memory Memory was quantified asthe total number of correct characters (out of a possi-

ble eight) a participant recalled from the virtual rep-resentationrsquos label without penalty for incorrect an-swers

255 Interpersonal Distance Interpersonaldistance was included as a behavioral social responseto the presence of an embodied agent We assessedtwo indicators of interpersonal distanceminimumdistance and reversal count Minimum distance wasthe closest distance the participant came to the em-bodied agent reversal count referred to the numberof times a participant changed direction along thez-axis (ie went back and forth) during their exami-nation of the embodied agent We consider reversalcount to be a proxy for hesitation in approaching theembodied agent

3 Results and Discussion

31 Hypotheses

We hypothesized that greater copresence wouldbe indicated by higher perceived copresence ratingsgreater liking of the embodied agent less willingness tocommit embarrassing acts in front of the embodiedagent the maintenance of greater interpersonal dis-tance and more hesitation in approaching the embod-ied agent We did not however have any specific pre-dictions with respect to the relationship betweencopresence and memory

Based on the results from previous copresencestudies we expected greater copresence (across ourvarious copresence measures) to occur with realisti-cally behaving embodied agents (specifically our re-corded and mimic behavioral-realism conditions thatfeatured movements based on actual human motion)compared with unrealistically behaving embodiedagents (ie our static and random-movement condi-tions) In addition we expected greater copresence tooccur with more human-looking embodied agentsthan with less human-looking ones Drawing on pre-dictions derived from the Blascovich et al (2002)threshold model of social influence however we ex-pected embodied-agent appearance to moderate the

Bailenson et al 385

effect of behavioral realism such that behavioral real-ism would matter more for less human-looking em-bodied agents than for more human-looking embod-ied agents Finally we expected to demonstrateconvergence in our dependent measures such thatthey would correlate with one another and change inthe same way as a function of our independent vari-able manipulations

32 Perceived Copresence Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movement vs mimic vs recorded) analysisof covariance controlling for participant gender4 re-vealed a significant main effect of behavioral realism onself-reported perceived copresence F (3 133) 272p 05 2 06 As Figure 4 demonstrates perceivedcopresence was higher in the random-movements con-dition compared to the static condition p 01 How-ever this main effect of behavioral realism was qualifiedby a significant representation type by behavioral-realisminteraction F (6 133) 218 p 05 2 09

Simple main effects tests revealed that within theblockhead condition perceived copresence varied signif-icantly as a function of behavioral realism F (3 133)

339 p 05 2 07 Specifically perceived copres-ence was significantly higher when the blockhead repre-sentation moved according to our random-movementalgorithm compared with either the static or mimic con-ditions ps 05 Simple main effects tests also revealedthat perceived copresence varied as a function of behav-

ioral realism for participants in the teddy-bear condi-tion F (3 133) 330 p 05 2 07 Specificallyperceived copresence was significantly higher in mimiccondition than in either the static or recorded condi-tions ps 05 and higher in the random-movementscondition relative to the recorded and static conditions(ps 05 and 10 respectively)

In contrast there were no significant differences inperceived copresence between any of the four behavioral-realism conditions for participants in the humanldquootherrdquo condition This pattern of findings is consis-tent with our prediction that behavioral realismwould have less of an influence on perceived copres-ence for embodied agents that appear human ascompared to embodied agents with a nonhuman ap-pearance In addition it is interesting to note that thetype of head movement that elicited the greatest levelof perceived copresence varied as a function of the em-bodied agentrsquos appearance with participants attributinghigh copresence ratings to both mimicking and ran-domly moving teddy-bear representations but only ran-domly moving blockhead representations

4 Although both male and female participants were used in thisstudy gender was not included as a separate factor in analyses involv-ing both behavioral realism and representation type When randomlyassigning participants to conditions we did not block participants bygender and as a result when behavioral realism and representationtype were fully crossed there were not enough participants of bothgenders in all 12 experimental conditions to reliably compute a three-way analysis of variance Thus our data analytic technique involvedincluding participant gender as a covariate in our analyses involvingboth behavioral realism and representation type in order to control forthe potential effect of participant gender The results do not changeappreciably if we remove gender as a covariate in our analyses but weinclude gender as a covariate to provide the most thorough analysispossible

Figure 4 Mean perceived-copresence ratings (using gender-

covariate adjusted means) as a function of representation type and

behavioral realism Conditions that share the same letter do not differ

within that level of representation type at p 05

386 PRESENCE VOLUME 14 NUMBER 4

33 Embarrassment Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender re-vealed a significant main effect of representation type onparticipantsrsquo self-reported willingness to perform embar-rassing acts in front of the embodied agent (ie embar-rassment) F (2 133) 421 p 05 2 06 Spe-cifically participants in the human embodied-agentcondition (M 260 SD 155) were significantlyless willing to perform embarrassing acts in front ofthe ldquootherrdquo than were participants in the teddy-bear(M 351 SD 154) embodied-agent conditionp 05 Participants in the human-representationcondition were also marginally less willing to performembarrassing acts than participants in the blockhead(M 313 SD 156) condition p 10 In sumthe presence of a human-looking representationseemed to reduce participantsrsquo willingness to performin embarrassing acts Regardless of the behavioral re-alism of the representation embodied agents wholooked like humans elicited greater social influencethan did non-human-looking representations

34 Likability Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender alsorevealed a significant main effect of representation typeon self-reported likability ratings F (2 133) 543p 01 2 08 Specifically both the human-lookingrepresentation (M 250 SD 109) and the teddy-bear representation (M 228 SD 109) were ratedas significantly more likable than the blockhead repre-sentation (M 179 SD 110) ps 05 Partici-pants found the blockhead representation (which wasthe least human looking of the three representations) tobe less attractive than either of the other two represen-tation types and they liked the blockhead representationthe least

35 Correlations between PerceivedCopresence Embarrassment andLikability Ratings

In order to examine the relationships among per-ceived copresence embarrassment and likability a se-ries of Pearson correlations was computed Consistentwith out predictions results of these analyses revealed asignificant negative correlation between perceived co-presence and a participantrsquos willingness to perform em-barrassing acts in front of the embodied agent r(144) 27 p 001 as well as a significant positivecorrelation between perceived copresence and likabilityr (144) 31 p 001 Specifically we found that thehigher the perceived copresence experienced the lesswilling participants were to perform embarrassing acts infront of the embodied agent (r2 07) and the moreparticipants liked the embodied agent (r2 10) Therewas no significant correlation between likability and em-barrassment ratings

36 Memory

We computed a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender to ex-amine the effects of our independent variables on partic-ipantsrsquo memory Results of these analyses revealed a sig-nificant main effect of behavioral realism on memory F(3 133) 414 p 01 2 09 Specifically mem-ory was significantly worse in the mimic condition thanin either the static or random-movement conditionsps 05 suggesting that overall participants had theworst memory for characters worn by a representationthat mimicked their own movements with a 4-s delayAgain however this main effect was qualified by a sig-nificant representation type by behavioral-realism inter-action (see Figure 5) F (6 133) 270 p 02 2

11Simple main effects tests revealed that within the

blockhead condition memory varied significantly as afunction of behavioral realism F (3 133) 271p 05 2 06 Specifically individuals rememberedsignificantly more characters when they were in the

Bailenson et al 387

random-movement condition than in either the mimicor the recorded conditions (the two behavioral-realismconditions that appeared most humanlike) ps 05and participants in the random-movement conditionremembered marginally more than participants in thestatic condition p 10 These results converge quitewell with the perceived copresence data in which partic-ipants indicated higher copresence ratings for blockheadrepresentations only when the blockhead demonstratedrandom head movements Interestingly this suggeststhat for representations who look the least human theleast human-looking head movements were seen asmost appropriate This finding that avoiding mis-matches is the most effective manner of combining anembodied agentrsquos appearance and behavior is consistentwith previous research on visually embodied agents(Slater amp Steed 2001 Garau et al 2003) as well asvoice agents (Lee amp Nass 2003)

Furthermore the convergence of memory andperceived-copresence ratings provides some evidencethat better memory is indicative of higher copresenceIn other words the more copresence an embodiedagent elicits the more a person interacting with that

embodied agent remembers about it It is interesting tonote however that despite the similar patterns of find-ings between the perceived-copresence data and thememory data results of a Pearson correlation revealedthat there was no significant linear relationship betweenperceived copresence and peoplersquos ability to rememberinformation about an embodied agent

There was however a significant simple main effectof behavioral realism within the human-representationcondition F (3 133) 549 p 001 2 11Specifically participants in the mimic and random-movement conditions remembered significantly fewercharacters than participants in either the static or re-corded conditions ps 05 Participants in the teddy-bear condition displayed the same pattern but in thatcondition none of the differences reached statisticalsignificance

These results suggest that participants may have beendistracted by a human-looking embodied agent thatmimicked their own movements which in turn af-fected their ability to notice and recall information inthe virtual environment Moreover participants mayalso have been distracted in the random-movement con-dition since this condition represented the greatest mis-match between the appearance (human looking) andbehavior (not human looking) of the embodied agentFurther research should examine these potential expla-nations as they suggest that different processes maybe involved for different algorithms for generatingembodied-agent head movements

37 Interpersonal Distance

There were no significant effects of behavioral re-alism or representation type on minimum distance (ourfirst measure of interpersonal distance) Thus partici-pants in all conditions maintained equivalent levels ofinterpersonal distance between themselves and the em-bodied agent There was also no significant correlationbetween perceived copresence and either of our behav-ioral dependent measures (ie minimum distance andnumber of returns)

However as hypothesized results of a 3 (representa-tion type human vs teddy bear vs blockhead) 4 (be-

Figure 5 Gender-covariate adjusted mean number of correctly

recalled characters out of eight on a surprise memory test as a

function of representation type and behavioral realism Conditions that

share the same letter do not differ within that level of representation

type at p 05

388 PRESENCE VOLUME 14 NUMBER 4

havioral realism static vs random movements vs mimicvs recorded) analysis of covariance controlling for par-ticipant gender revealed a significant main effect of be-havioral realism on reversal count or the number oftimes a participant changed direction along the z-axisduring his or her examination of the embodied agentF(3 133) 306 p 05 2 07 As can be seen inFigure 6 participants in the random-movement condi-tion exhibited significantly fewer reversals than partici-pants in any of the other three behavioral-realism condi-tions Pairwise comparisons indicated that there weresignificantly fewer reversals in the random-movementcondition compared to the other three conditions(ps 05) which are not significantly different fromeach other

This finding in part confirms our hypothesis con-cerning our behavioral dependent variable For an em-bodied agent whose head moves with curves randomlygenerated by the computer participants show little tono hesitation in approaching In contrast when thehead movements were controlled by a real person (ei-ther a mimic of their own movements or the move-ments from another participant) participants showedgreater hesitation in approaching the embodied agentThus computer-generated movements generated lesshesitancy than human-generated movements

Given this explanation however one might expecteven less hesitation that is fewer reversals in approach-ing the static representation This finding did not occurand one potential explanation is that overall a ran-domly moving representation may elicit less social influ-ence and copresence than a completely static representa-tion In other words the static representation has atleast the potential to exhibit realistic movements whilethe unrealistically moving embodied agent immediatelylowers the sense of copresence In sum participantsrsquohesitancy in approaching an embodied agent was influ-enced by the behavioral realism of the embodied agentbut not its appearance

4 Conclusions

41 Summary of Findings

To summarize the appearance and behavior ofembodied agents had varying effects on the dependentvariables In many instances there was convergence inthat self-report cognitive and behavioral measureschanged in similar ways as a function of our indepen-dent variable manipulations For example the perceived-copresence likability and embarrassment measures cor-related in the predicted directions and the blockheadrepresentation elicited similar reactions across behaviortypes for both the memory measure and the perceived-copresence measure In other instances however therewas a lack of convergence where varying realism pro-duced different outcomes for the various responsemodes One inevitable challenge facing researchers whoexamine a large number of dependent variables is tounderstand the instances in which those dependentvariables do and do not converge

In the current study for some of our dependent mea-sures of copresence (willingness to commit embarrassingacts and likability) manipulation of representation typeaffected copresence responses On other dependentmeasures (self-reported copresence ratings interper-sonal distance approach reversals and memory) ma-nipulation of behavioral realism changed participantsrsquoresponses One way to interpret these results is thatself-report measures were adequate for measuring

Figure 6 Mean number of direction reversals while approaching the

embodied agent as a function of behavioral realism

Bailenson et al 389

copresence levels as affected by variation in appear-ance whereas self-report measures were not sensitiveenough to detect the subtle differences elicited by thebehavioral-realism manipulation Indeed the only reliablemain effects of the behavioral-realism manipulation oc-curred with cognitive and behavioral dependent measures

In other words self-report questionnaires are effectivewhen measuring how people perceive an embodiedagent but not necessarily how they will respond to thatembodied agent Consequently not only does it matterhow one defines copresence it also matters how oneoperationalizes copresence If the goal of a researcher isto understand the way in which people fathom embod-ied agents on a conceptual level then self-report ques-tionnaires may be the best tool However if the goal ofthe researcher is to understand the way in which peoplewill interact with learn from or behave toward anagent then he or she should augment questionnaireswith other types of measures In this paper we are notlobbying for the abandonment of copresence self-reportquestionnaires however we do urge that researchersuse these questionnaires as a complement to other typesof measures whenever possible

While this explanation that self-report measures lacksensitivity receives support from the current data itmust be qualified On most measures there was an in-teraction between behavioral realism and representationtype At times these interactions did not conform toour specific predictions and were difficult to interpretPart of the reason for this difficulty is most likely due tothe fact that we assumed a priori that certain types ofhead movements would be viewed as more realistic thanothers Specifically we assumed that the measures basedon actual human behaviors (mimic and recorded) wouldappear more realistic than either those generated by thecomputer (random) or than no movements at all Re-corded human movements applied to 3D models how-ever may not always be the most effective method ofrendering realistic behaviors and the differences be-tween our various levels of behavioral realism may havebeen too subtle in the current study As previous workhas demonstrated (Garau et al 2003) the interactionbetween appearance and behavioral realism is compli-cated and the finding that seems to be emerging is that

large disparities between the two types of realism cancause diminished levels of copresence Only when theblockhead representation the least humanlike of thethree demonstrated unrealistic random head move-ments did it elicit high levels of self-report and cognitivecopresence

Our findings with respect to the main effects of ap-pearance on self-report measures suggest that how hu-man an embodied agent looks may also be an importantprecondition for experiencing copresence This findingis unanticipated given predictions made by the Blasco-vich model and given findings from previous work(Bailenson et al 2001 Garau et al 2003) It seemsthat virtual representations that look like a realistic hu-man as opposed to either a familiar or unfamiliar non-human elicit embarrassment reactions that would beappropriate only in the presence of a real human

42 Limitations and Future Research

Though this effect surfaces for representations thatare humanlike to varying degrees how this effect of ap-pearance would apply to interactions with multiple hu-man representations of varying photographic realismremains an empirical question One key direction forfuture research is to examine how variation in photo-graphic realism (ie visual fidelity) of humanlike em-bodied agents affects copresence and how it interactswith behavioral realism Furthermore another impor-tant direction is to begin future studies with pretestedlevels of all realism variables (behavioral photographicand others) to assist in a more thorough interpretationof results Moreover the current study is limited in thatwe implemented behavioral realism solely by varyinghead movements A more thorough investigation of thisphenomenon should investigate other types of behav-iors such as facial gestures hand movements and gaitas well as more complex behaviors such as planning andcoordinated actions In addition the interaction task wechose (a purely nonverbal task) is limited and researchshould definitely examine copresence in richer social-interaction settings Finally our method of translationusing the game pad as opposed to naturally trackedwalking was not a perfect behavioral measure since

390 PRESENCE VOLUME 14 NUMBER 4

people do not translate by pushing buttons in the realworld Future work should provide a further examinationof copresence using more natural behavioral measures

Importantly had we measured only self-report re-sponses we would have demonstrated a different patternof results in the current study Attempting to augment anystudy geared toward understanding human interactionwith embodied agents or avatars with dependent measuresfrom multiple response channels certainly can make datainterpretation more complicated Patterns from self-reportdata that seem clean and consistent often are not replicatedwhen a less direct measurement paradigm is implementedFurthermore when asking peripheral implicitly relatedquestions about presence such as likability or embarrass-ment one risks measuring constructs inside the humanmind that are not ldquocentralrdquo to the definition of copres-ence Adding more implicit and disparate measures maynot paint a cleaner picture but it does paint a more com-plete one The challenge to the researcher is interpretingthese complex patterns in the data in order to assess amore thorough understanding of an extremely novel andabstract phenomenon of interacting with a virtual digitalentity

It may be the case that copresence is a purely affectiveconstruct inside the human mind alternatively copres-ence may be more of a cognitive construct In this earlystage of the research field examining copresence in vir-tual environments the theoretical and empirical work isextremely limited and does not allow one to rule outeither of these two (or many other) possibilities Deter-mining the exact conceptual nature of an abstract con-struct in the human mind is a challenge that faces manysocial scientists The measurement framework we offerallows the researcher to get a more complete picture ofthe copresence construct and allows for future researchto utilize more specific types of construct-type hypothe-ses as well as measurement strategies

In addition to the basic research implications with re-spect to the use of various modes of dependent measuresthe current study also has some important applied implica-tions By understanding the components of a virtual repre-sentation that lead to elicitations of heightened copresencewith that representation we can better develop our virtualagents and virtual worlds to immerse and engage the user

As we as well as others have demonstrated creating highlevels of copresence is a complicated process involving anappropriate fit between levels of each type of realism Aswe grow to better understand these interactions we canbetter match agent design to context and perhaps reach alevel where it becomes easy to use agents to develop thedegree of copresence usually found with virtual avatarsand perhaps in the long term the degree of copresencefound with flesh-and-blood people during face-to-faceinteraction

References

Bailenson J N Aharoni E Beall A C Guadagno R Di-mov A amp Blascovich J (2004) Comparing behavioraland self-report measures of embodied agents social presence inimmersive virtual environments Proceedings of the 7th An-nual International Workshop on Presence Valencia Spain

Bailenson J N Beall A C amp Blascovich J (2002) Mutualgaze and task performance in shared virtual environmentsJournal of Visualization and Computer Animation 13 1ndash8

Bailenson J N Beall A C Loomis J Blascovich J ampTurk M (2004) Transformed social interaction Decou-pling representation from behavior and form in collabora-tive virtual environments Presence Teleoperators and Vir-tual Environments 13(4) 428ndash441

Bailenson J N Blascovich J Beall A C amp Loomis J(2001) Equilibrium revisited Mutual gaze and personalspace in virtual environments Presence Teleoperators andVirtual Environments 10 583ndash598

Bailenson J N Blascovich J Beall A C amp Loomis J M(2003) Interpersonal distance in immersive virtual environ-ments Personality and Social Psychology Bulletin 29 1ndash 15

Bailenson J N amp Yee N (in press) Digital chameleonsAutomatic assimilation of nonverbal gestures in immersivevirtual environments Psychological Science

Beall A C Bailenson J N Loomis J Blascovich J ampRex C (2003) Non-zero-sum mutual gaze in collaborativevirtual environments Proceedings of HCI InternationalCrete

Bente G Ruggenberg S Tietz B amp Wortberg S (2004)Measuring behavioral correlates of social presence in virtualencounters Paper presented at the International Communi-cation Association Conference May 27ndash31

Biocca F (1997) Cyborgrsquos dilemma Progressive embodi-

Bailenson et al 391

enced In one study interaction partners represented byanthropomorphic images were rated as more sociallyattractive and they received higher partner satisfactionratings than did partners represented either by no imageor by an image that did not appear human (Nowak2001) In addition participants indicated that they feltgreater copresence and that the communication me-dium was better able to support the social interactionwhen interactants were represented by highly anthropo-morphic images compared to when no image or an im-age low in anthropomorphism was used (Nowak ampBiocca 2001) Similarly Parise Kiesler Sproull amp Wa-ters (1996) demonstrated that more human-lookingembodied agents elicited more liking behavior and morecooperation than did less human-looking embodiedagents

One possible explanation for these seemingly contra-dictory findings concerns previous manipulations of ap-pearance Studies that have demonstrated null effectshave typically manipulated the visual fidelity (or photo-realism) of embodied agents and studies that havefound a significant effect of embodied-agent appearanceon copresence and other social outcomes have typicallymanipulated the anthropomorphism of the embodiedagent or the extent to which the embodied agent ap-pears human in form We are unaware of studies thathave systematically varied both photo-realism and an-thropomorphism and unfortunately most studies thathave been conducted on the effect of embodied-agentappearance have tended to confound photo-realism andanthropomorphism

Given that studies focusing on the photo-realism ofan embodied agent have consistently reported null ef-fects however we attempted to control for the effect ofphoto-realism in the current study by using embodiedagents of equivalent visual fidelity while simultaneouslymanipulating how much like an actual human being theembodied agent both looked and behaved We pre-dicted that embodied agents that either looked humanor acted human would elicit greater copresence thanembodied agents that did not look or act human How-ever based on the work of Blascovich and his colleagues(Blascovich 2002 Blascovich et al 2002) we alsohypothesized that embodied-agent appearance and be-

havior would interact with each other such that the be-havioral realism of an embodied agent would mattermost for non-human-looking embodied agents

16 Summary and Study Objectives

In this study we systematically examined the inde-pendent and interactive effects of both the appearanceand behavior of embodied agents on usersrsquo perceptionsand social responses within an immersive virtual envi-ronment In doing so we attempted to address a num-ber of currently unresolved issues First we examinedhow different ways of defining and operationalizingcopresence are influenced by systematic manipulationsof appearance and behavior of virtual representations Indoing so we attempted to reconcile some of the con-trasting findings that have been reported in the researchliterature Second we examined whether similar or dis-parate patterns of response obtained across differentmodalities of copresence measurement (ie self-reportvs cognitive vs behavioral dependent measures) Finallywe used four different algorithms to control the headmovements of the embodied agent some of which havenever before been empirically compared in an immersivevirtual environment

2 Method

21 Participants

One hundred and forty-six undergraduates (73males 73 females) ranging in age from 18 to 27 years(M 1963 years) recruited from the University of Cal-ifornia Santa Barbara served as participants They wereeither paid or received class credit for participation

22 Design

A two-way between-subjects factorial design wasused The first factor representation type included threelevels of embodiment human teddy bear and block-head As can be seen in Figure 1 the three embodiedagents varied in the extent to which they resembled a

382 PRESENCE VOLUME 14 NUMBER 4

human being1 For female participants the human rep-resentation created using 3dMeNow software depictedin the left panel of Figure 1 was used For male partici-pants a similar-looking male representation was usedAll three representations possessed equivalent behavioralfunctionality within the virtual environment (see below)and exhibited similar levels of photo-realism

While we attempted to ensure as much as possibleequal levels of photo-realism (by using textures of simi-lar resolution and by ensuring that the underlying 3Dmesh for each model contained similar numbers of an-chor points and structural facial features) subjectivelythere may have been minor differences in photo-realism

The second between-subjects factor behavioral real-ism included four levels of head movements Four dif-ferent algorithms were employed to control the embod-ied agentrsquos head movements In the static condition theembodied agent remained motionless In the randommovement condition the embodied agentrsquos head contin-ually moved in trajectories generated by the computertoward random points within a 60-degree arc of eachside of its head (for both pitch and yaw) In the mimiccondition the embodied agentrsquos head movements mir-rored the participantrsquos actual head movements with a

4-s delay2 In the recorded movement condition headmovements recorded from a randomly selected partici-pant in a previous session were used to control the em-bodied agentrsquos head movements In all four conditionsthe embodied agent blinked its eyes according to a pre-set algorithm but there were no facial gestures exhib-ited Complete crossing of these two independent vari-ables resulted in 12 between-subjects conditions3

22 The Virtual Environment

The virtual room in which participants were im-mersed is depicted in Figure 2 Participants sat in achair at a table facing an embodied agent Participantswere able to see the head and shoulders of the agentThe same body was used for each representation type Atag containing a string of letters and numbers later usedas part of a surprise memory task was positioned on theshirt of the embodied agent

1 There are a variety of terms that researchers use to describe howreal or how human a virtual representation looks including anthropo-morphism visual realism and photographic (or photo-) realism Wehave chosen to use the more general term representation type as a vari-able name and talk about the manipulation as mainly pertaining tophotographic realism We acknowledge that other terms such as an-thropomorphism however may also be appropriate

2 Previous research using mimicked head movements has indicatedthat users generally do not detect embodied agents that are mimickingtheir head movements with a 4-second delay (Bailenson Beall Loo-mis Blascovich amp Turk 2004 Bailenson amp Yee in press) Howeverthis study is unique in that it directly compared the effect of mimickedhead movements with other head-movement algorithms

3 There were an unequal number of females and males across the12 conditions

Figure 1 Photographically realistic human teddy bear and

blockhead embodied agents

Figure 2 A participantrsquos view of the virtual room

Bailenson et al 383

23 Equipment

Participants wore a Virtual Research V8 stereo-scopic head-mounted display (HMD see Figure 3) TheHMD featured dual 680 horizontal by 480 vertical pixelresolution LCD panels that refreshed at 72 Hz Thedisplay optics presented a visual field subtending approxi-mately 50 degrees horizontally by 38 degrees vertically

Perspectively correct stereoscopic images were ren-dered by a 450 MHz Pentium III dual-processor com-puter with an Evans amp Sutherland Tornado 3000 dualpipe graphics card and were updated at an averageframe rate of 36 Hz The simulated viewpoint was con-tinually updated as a function of the participantsrsquo headmovements The orientation of the participantrsquos headwas tracked by a three-axis orientation sensing system(Intersense IS300 update rate of 150 Hz) The systemlatency or the amount of delay between a participantrsquoshead movement and the resulting concomitant updatein the HMDrsquos visual display was 65 ms maximum Thesoftware used to assimilate the rendering and trackingwas Vizard 20 Participants used a Logitech Rumble-

Pad Pro game pad to interact with the virtual environ-ment

24 Procedure

When participants arrived at the laboratory anexperimenter sat them down and instructed them onhow to wear and adjust the HMD and how to use thegame pad depicted in Figure 3 to interact with thevirtual environment Once immersed participantsfound themselves seated at a table directly acrossfrom an embodied agent (see Figure 2) Dependingon the assigned condition the embodied agent variedin appearance and behavior Using the game pad par-ticipants scrolled through the following instructionaltext that appeared above the embodied agentrsquos head

In this laboratory we are interested in peoplersquos reac-tions to ldquoothersrdquo in virtual environments In thisstudy you will spend time with an ldquootherrdquo After ashort time we will ask you questions about theldquootherrdquo You will use the game pad to answer thequestions The questions measure your agreementThe more you agree the higher the number youshould choose on the rating scale in front of you

We carefully chose and repeatedly used the wordldquootherrdquo set off by quotation marks to avoid using sug-gestive nouns such as ldquopersonrdquo or ldquoagentrdquo Participantswere next instructed to look directly at the ldquootherrdquo for90 s Next a series of 10 copresence embarrassment andlikability self-report questionnaire items (see AppendixA) appeared one at a time over the embodied agentrsquoshead in random order Participants indicated theiragreement with these items by using the game pad toselect a response on a Likert-type rating scale that ap-peared under each item Hence ratings were providedin real time as participants experienced the embodiedagent

After completing the items participants were in-structed to approach the embodied agent for furtherexamination They used buttons on the game pad tomove toward (positive translation on the z axis) or awayfrom (negative translation on the z axis) the embodiedagent with each press of a game-pad button equivalentto a jump of 25 cm (the starting distance between the

Figure 3 A depiction of our immersive virtual environment system

The components are (1) orientation tracking sensor (2) image

generator (3) HMD and (4) game pad input device

384 PRESENCE VOLUME 14 NUMBER 4

participantrsquos nose and the embodied agentrsquos nose wasapproximately 3 m) We did not render the body of par-ticipants so there was no leaning or walking animationduring the translationmdashonly their viewpoint movedcloser or farther from the agent After participants fin-ished approaching the embodied agent and examiningit they took off the HMD and filled out a demographicquestionnaire Finally we administered a surprise mem-ory test in which participants were asked to recall therandom string of characters written on the embodiedagentrsquos chest (see Figure 2)

25 Measures

251 Perceived Copresence Perceived copres-ence was assessed via a three-item self-report measureWe created a composite copresence measure ( 71see Appendix A) by averaging the three copresencequestions Higher scores indicated higher levels ofself-reported copresence where ratings were made on aLikert-type scale with response options ranging from 0(strongly disagree) to 6 (strongly agree)

252 Embarrassment Embarrassment was as-sessed as a self-reported social response to the presenceof an embodied agent Previous research has demon-strated that willingness to perform embarrassing acts infront of a virtual representation is a worthy measure ofthat representationrsquos degree of social influence (Bailen-son et al 2003) We averaged the three embarrassmentquestions ( 72 see Appendix A) into a single com-posite measure with higher numbers indicating greaterwillingness to perform embarrassing acts in front of anembodied agent

253 Likability Likability was assessed as a sec-ond self-reported social response to the presence ofan embodied agent We averaged the four likabilityquestions ( 71 see Appendix A) into a singlecomposite measure with higher numbers indicatingthat participants liked the virtual representation more

254 Memory Memory was quantified asthe total number of correct characters (out of a possi-

ble eight) a participant recalled from the virtual rep-resentationrsquos label without penalty for incorrect an-swers

255 Interpersonal Distance Interpersonaldistance was included as a behavioral social responseto the presence of an embodied agent We assessedtwo indicators of interpersonal distanceminimumdistance and reversal count Minimum distance wasthe closest distance the participant came to the em-bodied agent reversal count referred to the numberof times a participant changed direction along thez-axis (ie went back and forth) during their exami-nation of the embodied agent We consider reversalcount to be a proxy for hesitation in approaching theembodied agent

3 Results and Discussion

31 Hypotheses

We hypothesized that greater copresence wouldbe indicated by higher perceived copresence ratingsgreater liking of the embodied agent less willingness tocommit embarrassing acts in front of the embodiedagent the maintenance of greater interpersonal dis-tance and more hesitation in approaching the embod-ied agent We did not however have any specific pre-dictions with respect to the relationship betweencopresence and memory

Based on the results from previous copresencestudies we expected greater copresence (across ourvarious copresence measures) to occur with realisti-cally behaving embodied agents (specifically our re-corded and mimic behavioral-realism conditions thatfeatured movements based on actual human motion)compared with unrealistically behaving embodiedagents (ie our static and random-movement condi-tions) In addition we expected greater copresence tooccur with more human-looking embodied agentsthan with less human-looking ones Drawing on pre-dictions derived from the Blascovich et al (2002)threshold model of social influence however we ex-pected embodied-agent appearance to moderate the

Bailenson et al 385

effect of behavioral realism such that behavioral real-ism would matter more for less human-looking em-bodied agents than for more human-looking embod-ied agents Finally we expected to demonstrateconvergence in our dependent measures such thatthey would correlate with one another and change inthe same way as a function of our independent vari-able manipulations

32 Perceived Copresence Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movement vs mimic vs recorded) analysisof covariance controlling for participant gender4 re-vealed a significant main effect of behavioral realism onself-reported perceived copresence F (3 133) 272p 05 2 06 As Figure 4 demonstrates perceivedcopresence was higher in the random-movements con-dition compared to the static condition p 01 How-ever this main effect of behavioral realism was qualifiedby a significant representation type by behavioral-realisminteraction F (6 133) 218 p 05 2 09

Simple main effects tests revealed that within theblockhead condition perceived copresence varied signif-icantly as a function of behavioral realism F (3 133)

339 p 05 2 07 Specifically perceived copres-ence was significantly higher when the blockhead repre-sentation moved according to our random-movementalgorithm compared with either the static or mimic con-ditions ps 05 Simple main effects tests also revealedthat perceived copresence varied as a function of behav-

ioral realism for participants in the teddy-bear condi-tion F (3 133) 330 p 05 2 07 Specificallyperceived copresence was significantly higher in mimiccondition than in either the static or recorded condi-tions ps 05 and higher in the random-movementscondition relative to the recorded and static conditions(ps 05 and 10 respectively)

In contrast there were no significant differences inperceived copresence between any of the four behavioral-realism conditions for participants in the humanldquootherrdquo condition This pattern of findings is consis-tent with our prediction that behavioral realismwould have less of an influence on perceived copres-ence for embodied agents that appear human ascompared to embodied agents with a nonhuman ap-pearance In addition it is interesting to note that thetype of head movement that elicited the greatest levelof perceived copresence varied as a function of the em-bodied agentrsquos appearance with participants attributinghigh copresence ratings to both mimicking and ran-domly moving teddy-bear representations but only ran-domly moving blockhead representations

4 Although both male and female participants were used in thisstudy gender was not included as a separate factor in analyses involv-ing both behavioral realism and representation type When randomlyassigning participants to conditions we did not block participants bygender and as a result when behavioral realism and representationtype were fully crossed there were not enough participants of bothgenders in all 12 experimental conditions to reliably compute a three-way analysis of variance Thus our data analytic technique involvedincluding participant gender as a covariate in our analyses involvingboth behavioral realism and representation type in order to control forthe potential effect of participant gender The results do not changeappreciably if we remove gender as a covariate in our analyses but weinclude gender as a covariate to provide the most thorough analysispossible

Figure 4 Mean perceived-copresence ratings (using gender-

covariate adjusted means) as a function of representation type and

behavioral realism Conditions that share the same letter do not differ

within that level of representation type at p 05

386 PRESENCE VOLUME 14 NUMBER 4

33 Embarrassment Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender re-vealed a significant main effect of representation type onparticipantsrsquo self-reported willingness to perform embar-rassing acts in front of the embodied agent (ie embar-rassment) F (2 133) 421 p 05 2 06 Spe-cifically participants in the human embodied-agentcondition (M 260 SD 155) were significantlyless willing to perform embarrassing acts in front ofthe ldquootherrdquo than were participants in the teddy-bear(M 351 SD 154) embodied-agent conditionp 05 Participants in the human-representationcondition were also marginally less willing to performembarrassing acts than participants in the blockhead(M 313 SD 156) condition p 10 In sumthe presence of a human-looking representationseemed to reduce participantsrsquo willingness to performin embarrassing acts Regardless of the behavioral re-alism of the representation embodied agents wholooked like humans elicited greater social influencethan did non-human-looking representations

34 Likability Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender alsorevealed a significant main effect of representation typeon self-reported likability ratings F (2 133) 543p 01 2 08 Specifically both the human-lookingrepresentation (M 250 SD 109) and the teddy-bear representation (M 228 SD 109) were ratedas significantly more likable than the blockhead repre-sentation (M 179 SD 110) ps 05 Partici-pants found the blockhead representation (which wasthe least human looking of the three representations) tobe less attractive than either of the other two represen-tation types and they liked the blockhead representationthe least

35 Correlations between PerceivedCopresence Embarrassment andLikability Ratings

In order to examine the relationships among per-ceived copresence embarrassment and likability a se-ries of Pearson correlations was computed Consistentwith out predictions results of these analyses revealed asignificant negative correlation between perceived co-presence and a participantrsquos willingness to perform em-barrassing acts in front of the embodied agent r(144) 27 p 001 as well as a significant positivecorrelation between perceived copresence and likabilityr (144) 31 p 001 Specifically we found that thehigher the perceived copresence experienced the lesswilling participants were to perform embarrassing acts infront of the embodied agent (r2 07) and the moreparticipants liked the embodied agent (r2 10) Therewas no significant correlation between likability and em-barrassment ratings

36 Memory

We computed a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender to ex-amine the effects of our independent variables on partic-ipantsrsquo memory Results of these analyses revealed a sig-nificant main effect of behavioral realism on memory F(3 133) 414 p 01 2 09 Specifically mem-ory was significantly worse in the mimic condition thanin either the static or random-movement conditionsps 05 suggesting that overall participants had theworst memory for characters worn by a representationthat mimicked their own movements with a 4-s delayAgain however this main effect was qualified by a sig-nificant representation type by behavioral-realism inter-action (see Figure 5) F (6 133) 270 p 02 2

11Simple main effects tests revealed that within the

blockhead condition memory varied significantly as afunction of behavioral realism F (3 133) 271p 05 2 06 Specifically individuals rememberedsignificantly more characters when they were in the

Bailenson et al 387

random-movement condition than in either the mimicor the recorded conditions (the two behavioral-realismconditions that appeared most humanlike) ps 05and participants in the random-movement conditionremembered marginally more than participants in thestatic condition p 10 These results converge quitewell with the perceived copresence data in which partic-ipants indicated higher copresence ratings for blockheadrepresentations only when the blockhead demonstratedrandom head movements Interestingly this suggeststhat for representations who look the least human theleast human-looking head movements were seen asmost appropriate This finding that avoiding mis-matches is the most effective manner of combining anembodied agentrsquos appearance and behavior is consistentwith previous research on visually embodied agents(Slater amp Steed 2001 Garau et al 2003) as well asvoice agents (Lee amp Nass 2003)

Furthermore the convergence of memory andperceived-copresence ratings provides some evidencethat better memory is indicative of higher copresenceIn other words the more copresence an embodiedagent elicits the more a person interacting with that

embodied agent remembers about it It is interesting tonote however that despite the similar patterns of find-ings between the perceived-copresence data and thememory data results of a Pearson correlation revealedthat there was no significant linear relationship betweenperceived copresence and peoplersquos ability to rememberinformation about an embodied agent

There was however a significant simple main effectof behavioral realism within the human-representationcondition F (3 133) 549 p 001 2 11Specifically participants in the mimic and random-movement conditions remembered significantly fewercharacters than participants in either the static or re-corded conditions ps 05 Participants in the teddy-bear condition displayed the same pattern but in thatcondition none of the differences reached statisticalsignificance

These results suggest that participants may have beendistracted by a human-looking embodied agent thatmimicked their own movements which in turn af-fected their ability to notice and recall information inthe virtual environment Moreover participants mayalso have been distracted in the random-movement con-dition since this condition represented the greatest mis-match between the appearance (human looking) andbehavior (not human looking) of the embodied agentFurther research should examine these potential expla-nations as they suggest that different processes maybe involved for different algorithms for generatingembodied-agent head movements

37 Interpersonal Distance

There were no significant effects of behavioral re-alism or representation type on minimum distance (ourfirst measure of interpersonal distance) Thus partici-pants in all conditions maintained equivalent levels ofinterpersonal distance between themselves and the em-bodied agent There was also no significant correlationbetween perceived copresence and either of our behav-ioral dependent measures (ie minimum distance andnumber of returns)

However as hypothesized results of a 3 (representa-tion type human vs teddy bear vs blockhead) 4 (be-

Figure 5 Gender-covariate adjusted mean number of correctly

recalled characters out of eight on a surprise memory test as a

function of representation type and behavioral realism Conditions that

share the same letter do not differ within that level of representation

type at p 05

388 PRESENCE VOLUME 14 NUMBER 4

havioral realism static vs random movements vs mimicvs recorded) analysis of covariance controlling for par-ticipant gender revealed a significant main effect of be-havioral realism on reversal count or the number oftimes a participant changed direction along the z-axisduring his or her examination of the embodied agentF(3 133) 306 p 05 2 07 As can be seen inFigure 6 participants in the random-movement condi-tion exhibited significantly fewer reversals than partici-pants in any of the other three behavioral-realism condi-tions Pairwise comparisons indicated that there weresignificantly fewer reversals in the random-movementcondition compared to the other three conditions(ps 05) which are not significantly different fromeach other

This finding in part confirms our hypothesis con-cerning our behavioral dependent variable For an em-bodied agent whose head moves with curves randomlygenerated by the computer participants show little tono hesitation in approaching In contrast when thehead movements were controlled by a real person (ei-ther a mimic of their own movements or the move-ments from another participant) participants showedgreater hesitation in approaching the embodied agentThus computer-generated movements generated lesshesitancy than human-generated movements

Given this explanation however one might expecteven less hesitation that is fewer reversals in approach-ing the static representation This finding did not occurand one potential explanation is that overall a ran-domly moving representation may elicit less social influ-ence and copresence than a completely static representa-tion In other words the static representation has atleast the potential to exhibit realistic movements whilethe unrealistically moving embodied agent immediatelylowers the sense of copresence In sum participantsrsquohesitancy in approaching an embodied agent was influ-enced by the behavioral realism of the embodied agentbut not its appearance

4 Conclusions

41 Summary of Findings

To summarize the appearance and behavior ofembodied agents had varying effects on the dependentvariables In many instances there was convergence inthat self-report cognitive and behavioral measureschanged in similar ways as a function of our indepen-dent variable manipulations For example the perceived-copresence likability and embarrassment measures cor-related in the predicted directions and the blockheadrepresentation elicited similar reactions across behaviortypes for both the memory measure and the perceived-copresence measure In other instances however therewas a lack of convergence where varying realism pro-duced different outcomes for the various responsemodes One inevitable challenge facing researchers whoexamine a large number of dependent variables is tounderstand the instances in which those dependentvariables do and do not converge

In the current study for some of our dependent mea-sures of copresence (willingness to commit embarrassingacts and likability) manipulation of representation typeaffected copresence responses On other dependentmeasures (self-reported copresence ratings interper-sonal distance approach reversals and memory) ma-nipulation of behavioral realism changed participantsrsquoresponses One way to interpret these results is thatself-report measures were adequate for measuring

Figure 6 Mean number of direction reversals while approaching the

embodied agent as a function of behavioral realism

Bailenson et al 389

copresence levels as affected by variation in appear-ance whereas self-report measures were not sensitiveenough to detect the subtle differences elicited by thebehavioral-realism manipulation Indeed the only reliablemain effects of the behavioral-realism manipulation oc-curred with cognitive and behavioral dependent measures

In other words self-report questionnaires are effectivewhen measuring how people perceive an embodiedagent but not necessarily how they will respond to thatembodied agent Consequently not only does it matterhow one defines copresence it also matters how oneoperationalizes copresence If the goal of a researcher isto understand the way in which people fathom embod-ied agents on a conceptual level then self-report ques-tionnaires may be the best tool However if the goal ofthe researcher is to understand the way in which peoplewill interact with learn from or behave toward anagent then he or she should augment questionnaireswith other types of measures In this paper we are notlobbying for the abandonment of copresence self-reportquestionnaires however we do urge that researchersuse these questionnaires as a complement to other typesof measures whenever possible

While this explanation that self-report measures lacksensitivity receives support from the current data itmust be qualified On most measures there was an in-teraction between behavioral realism and representationtype At times these interactions did not conform toour specific predictions and were difficult to interpretPart of the reason for this difficulty is most likely due tothe fact that we assumed a priori that certain types ofhead movements would be viewed as more realistic thanothers Specifically we assumed that the measures basedon actual human behaviors (mimic and recorded) wouldappear more realistic than either those generated by thecomputer (random) or than no movements at all Re-corded human movements applied to 3D models how-ever may not always be the most effective method ofrendering realistic behaviors and the differences be-tween our various levels of behavioral realism may havebeen too subtle in the current study As previous workhas demonstrated (Garau et al 2003) the interactionbetween appearance and behavioral realism is compli-cated and the finding that seems to be emerging is that

large disparities between the two types of realism cancause diminished levels of copresence Only when theblockhead representation the least humanlike of thethree demonstrated unrealistic random head move-ments did it elicit high levels of self-report and cognitivecopresence

Our findings with respect to the main effects of ap-pearance on self-report measures suggest that how hu-man an embodied agent looks may also be an importantprecondition for experiencing copresence This findingis unanticipated given predictions made by the Blasco-vich model and given findings from previous work(Bailenson et al 2001 Garau et al 2003) It seemsthat virtual representations that look like a realistic hu-man as opposed to either a familiar or unfamiliar non-human elicit embarrassment reactions that would beappropriate only in the presence of a real human

42 Limitations and Future Research

Though this effect surfaces for representations thatare humanlike to varying degrees how this effect of ap-pearance would apply to interactions with multiple hu-man representations of varying photographic realismremains an empirical question One key direction forfuture research is to examine how variation in photo-graphic realism (ie visual fidelity) of humanlike em-bodied agents affects copresence and how it interactswith behavioral realism Furthermore another impor-tant direction is to begin future studies with pretestedlevels of all realism variables (behavioral photographicand others) to assist in a more thorough interpretationof results Moreover the current study is limited in thatwe implemented behavioral realism solely by varyinghead movements A more thorough investigation of thisphenomenon should investigate other types of behav-iors such as facial gestures hand movements and gaitas well as more complex behaviors such as planning andcoordinated actions In addition the interaction task wechose (a purely nonverbal task) is limited and researchshould definitely examine copresence in richer social-interaction settings Finally our method of translationusing the game pad as opposed to naturally trackedwalking was not a perfect behavioral measure since

390 PRESENCE VOLUME 14 NUMBER 4

people do not translate by pushing buttons in the realworld Future work should provide a further examinationof copresence using more natural behavioral measures

Importantly had we measured only self-report re-sponses we would have demonstrated a different patternof results in the current study Attempting to augment anystudy geared toward understanding human interactionwith embodied agents or avatars with dependent measuresfrom multiple response channels certainly can make datainterpretation more complicated Patterns from self-reportdata that seem clean and consistent often are not replicatedwhen a less direct measurement paradigm is implementedFurthermore when asking peripheral implicitly relatedquestions about presence such as likability or embarrass-ment one risks measuring constructs inside the humanmind that are not ldquocentralrdquo to the definition of copres-ence Adding more implicit and disparate measures maynot paint a cleaner picture but it does paint a more com-plete one The challenge to the researcher is interpretingthese complex patterns in the data in order to assess amore thorough understanding of an extremely novel andabstract phenomenon of interacting with a virtual digitalentity

It may be the case that copresence is a purely affectiveconstruct inside the human mind alternatively copres-ence may be more of a cognitive construct In this earlystage of the research field examining copresence in vir-tual environments the theoretical and empirical work isextremely limited and does not allow one to rule outeither of these two (or many other) possibilities Deter-mining the exact conceptual nature of an abstract con-struct in the human mind is a challenge that faces manysocial scientists The measurement framework we offerallows the researcher to get a more complete picture ofthe copresence construct and allows for future researchto utilize more specific types of construct-type hypothe-ses as well as measurement strategies

In addition to the basic research implications with re-spect to the use of various modes of dependent measuresthe current study also has some important applied implica-tions By understanding the components of a virtual repre-sentation that lead to elicitations of heightened copresencewith that representation we can better develop our virtualagents and virtual worlds to immerse and engage the user

As we as well as others have demonstrated creating highlevels of copresence is a complicated process involving anappropriate fit between levels of each type of realism Aswe grow to better understand these interactions we canbetter match agent design to context and perhaps reach alevel where it becomes easy to use agents to develop thedegree of copresence usually found with virtual avatarsand perhaps in the long term the degree of copresencefound with flesh-and-blood people during face-to-faceinteraction

References

Bailenson J N Aharoni E Beall A C Guadagno R Di-mov A amp Blascovich J (2004) Comparing behavioraland self-report measures of embodied agents social presence inimmersive virtual environments Proceedings of the 7th An-nual International Workshop on Presence Valencia Spain

Bailenson J N Beall A C amp Blascovich J (2002) Mutualgaze and task performance in shared virtual environmentsJournal of Visualization and Computer Animation 13 1ndash8

Bailenson J N Beall A C Loomis J Blascovich J ampTurk M (2004) Transformed social interaction Decou-pling representation from behavior and form in collabora-tive virtual environments Presence Teleoperators and Vir-tual Environments 13(4) 428ndash441

Bailenson J N Blascovich J Beall A C amp Loomis J(2001) Equilibrium revisited Mutual gaze and personalspace in virtual environments Presence Teleoperators andVirtual Environments 10 583ndash598

Bailenson J N Blascovich J Beall A C amp Loomis J M(2003) Interpersonal distance in immersive virtual environ-ments Personality and Social Psychology Bulletin 29 1ndash 15

Bailenson J N amp Yee N (in press) Digital chameleonsAutomatic assimilation of nonverbal gestures in immersivevirtual environments Psychological Science

Beall A C Bailenson J N Loomis J Blascovich J ampRex C (2003) Non-zero-sum mutual gaze in collaborativevirtual environments Proceedings of HCI InternationalCrete

Bente G Ruggenberg S Tietz B amp Wortberg S (2004)Measuring behavioral correlates of social presence in virtualencounters Paper presented at the International Communi-cation Association Conference May 27ndash31

Biocca F (1997) Cyborgrsquos dilemma Progressive embodi-

Bailenson et al 391

human being1 For female participants the human rep-resentation created using 3dMeNow software depictedin the left panel of Figure 1 was used For male partici-pants a similar-looking male representation was usedAll three representations possessed equivalent behavioralfunctionality within the virtual environment (see below)and exhibited similar levels of photo-realism

While we attempted to ensure as much as possibleequal levels of photo-realism (by using textures of simi-lar resolution and by ensuring that the underlying 3Dmesh for each model contained similar numbers of an-chor points and structural facial features) subjectivelythere may have been minor differences in photo-realism

The second between-subjects factor behavioral real-ism included four levels of head movements Four dif-ferent algorithms were employed to control the embod-ied agentrsquos head movements In the static condition theembodied agent remained motionless In the randommovement condition the embodied agentrsquos head contin-ually moved in trajectories generated by the computertoward random points within a 60-degree arc of eachside of its head (for both pitch and yaw) In the mimiccondition the embodied agentrsquos head movements mir-rored the participantrsquos actual head movements with a

4-s delay2 In the recorded movement condition headmovements recorded from a randomly selected partici-pant in a previous session were used to control the em-bodied agentrsquos head movements In all four conditionsthe embodied agent blinked its eyes according to a pre-set algorithm but there were no facial gestures exhib-ited Complete crossing of these two independent vari-ables resulted in 12 between-subjects conditions3

22 The Virtual Environment

The virtual room in which participants were im-mersed is depicted in Figure 2 Participants sat in achair at a table facing an embodied agent Participantswere able to see the head and shoulders of the agentThe same body was used for each representation type Atag containing a string of letters and numbers later usedas part of a surprise memory task was positioned on theshirt of the embodied agent

1 There are a variety of terms that researchers use to describe howreal or how human a virtual representation looks including anthropo-morphism visual realism and photographic (or photo-) realism Wehave chosen to use the more general term representation type as a vari-able name and talk about the manipulation as mainly pertaining tophotographic realism We acknowledge that other terms such as an-thropomorphism however may also be appropriate

2 Previous research using mimicked head movements has indicatedthat users generally do not detect embodied agents that are mimickingtheir head movements with a 4-second delay (Bailenson Beall Loo-mis Blascovich amp Turk 2004 Bailenson amp Yee in press) Howeverthis study is unique in that it directly compared the effect of mimickedhead movements with other head-movement algorithms

3 There were an unequal number of females and males across the12 conditions

Figure 1 Photographically realistic human teddy bear and

blockhead embodied agents

Figure 2 A participantrsquos view of the virtual room

Bailenson et al 383

23 Equipment

Participants wore a Virtual Research V8 stereo-scopic head-mounted display (HMD see Figure 3) TheHMD featured dual 680 horizontal by 480 vertical pixelresolution LCD panels that refreshed at 72 Hz Thedisplay optics presented a visual field subtending approxi-mately 50 degrees horizontally by 38 degrees vertically

Perspectively correct stereoscopic images were ren-dered by a 450 MHz Pentium III dual-processor com-puter with an Evans amp Sutherland Tornado 3000 dualpipe graphics card and were updated at an averageframe rate of 36 Hz The simulated viewpoint was con-tinually updated as a function of the participantsrsquo headmovements The orientation of the participantrsquos headwas tracked by a three-axis orientation sensing system(Intersense IS300 update rate of 150 Hz) The systemlatency or the amount of delay between a participantrsquoshead movement and the resulting concomitant updatein the HMDrsquos visual display was 65 ms maximum Thesoftware used to assimilate the rendering and trackingwas Vizard 20 Participants used a Logitech Rumble-

Pad Pro game pad to interact with the virtual environ-ment

24 Procedure

When participants arrived at the laboratory anexperimenter sat them down and instructed them onhow to wear and adjust the HMD and how to use thegame pad depicted in Figure 3 to interact with thevirtual environment Once immersed participantsfound themselves seated at a table directly acrossfrom an embodied agent (see Figure 2) Dependingon the assigned condition the embodied agent variedin appearance and behavior Using the game pad par-ticipants scrolled through the following instructionaltext that appeared above the embodied agentrsquos head

In this laboratory we are interested in peoplersquos reac-tions to ldquoothersrdquo in virtual environments In thisstudy you will spend time with an ldquootherrdquo After ashort time we will ask you questions about theldquootherrdquo You will use the game pad to answer thequestions The questions measure your agreementThe more you agree the higher the number youshould choose on the rating scale in front of you

We carefully chose and repeatedly used the wordldquootherrdquo set off by quotation marks to avoid using sug-gestive nouns such as ldquopersonrdquo or ldquoagentrdquo Participantswere next instructed to look directly at the ldquootherrdquo for90 s Next a series of 10 copresence embarrassment andlikability self-report questionnaire items (see AppendixA) appeared one at a time over the embodied agentrsquoshead in random order Participants indicated theiragreement with these items by using the game pad toselect a response on a Likert-type rating scale that ap-peared under each item Hence ratings were providedin real time as participants experienced the embodiedagent

After completing the items participants were in-structed to approach the embodied agent for furtherexamination They used buttons on the game pad tomove toward (positive translation on the z axis) or awayfrom (negative translation on the z axis) the embodiedagent with each press of a game-pad button equivalentto a jump of 25 cm (the starting distance between the

Figure 3 A depiction of our immersive virtual environment system

The components are (1) orientation tracking sensor (2) image

generator (3) HMD and (4) game pad input device

384 PRESENCE VOLUME 14 NUMBER 4

participantrsquos nose and the embodied agentrsquos nose wasapproximately 3 m) We did not render the body of par-ticipants so there was no leaning or walking animationduring the translationmdashonly their viewpoint movedcloser or farther from the agent After participants fin-ished approaching the embodied agent and examiningit they took off the HMD and filled out a demographicquestionnaire Finally we administered a surprise mem-ory test in which participants were asked to recall therandom string of characters written on the embodiedagentrsquos chest (see Figure 2)

25 Measures

251 Perceived Copresence Perceived copres-ence was assessed via a three-item self-report measureWe created a composite copresence measure ( 71see Appendix A) by averaging the three copresencequestions Higher scores indicated higher levels ofself-reported copresence where ratings were made on aLikert-type scale with response options ranging from 0(strongly disagree) to 6 (strongly agree)

252 Embarrassment Embarrassment was as-sessed as a self-reported social response to the presenceof an embodied agent Previous research has demon-strated that willingness to perform embarrassing acts infront of a virtual representation is a worthy measure ofthat representationrsquos degree of social influence (Bailen-son et al 2003) We averaged the three embarrassmentquestions ( 72 see Appendix A) into a single com-posite measure with higher numbers indicating greaterwillingness to perform embarrassing acts in front of anembodied agent

253 Likability Likability was assessed as a sec-ond self-reported social response to the presence ofan embodied agent We averaged the four likabilityquestions ( 71 see Appendix A) into a singlecomposite measure with higher numbers indicatingthat participants liked the virtual representation more

254 Memory Memory was quantified asthe total number of correct characters (out of a possi-

ble eight) a participant recalled from the virtual rep-resentationrsquos label without penalty for incorrect an-swers

255 Interpersonal Distance Interpersonaldistance was included as a behavioral social responseto the presence of an embodied agent We assessedtwo indicators of interpersonal distanceminimumdistance and reversal count Minimum distance wasthe closest distance the participant came to the em-bodied agent reversal count referred to the numberof times a participant changed direction along thez-axis (ie went back and forth) during their exami-nation of the embodied agent We consider reversalcount to be a proxy for hesitation in approaching theembodied agent

3 Results and Discussion

31 Hypotheses

We hypothesized that greater copresence wouldbe indicated by higher perceived copresence ratingsgreater liking of the embodied agent less willingness tocommit embarrassing acts in front of the embodiedagent the maintenance of greater interpersonal dis-tance and more hesitation in approaching the embod-ied agent We did not however have any specific pre-dictions with respect to the relationship betweencopresence and memory

Based on the results from previous copresencestudies we expected greater copresence (across ourvarious copresence measures) to occur with realisti-cally behaving embodied agents (specifically our re-corded and mimic behavioral-realism conditions thatfeatured movements based on actual human motion)compared with unrealistically behaving embodiedagents (ie our static and random-movement condi-tions) In addition we expected greater copresence tooccur with more human-looking embodied agentsthan with less human-looking ones Drawing on pre-dictions derived from the Blascovich et al (2002)threshold model of social influence however we ex-pected embodied-agent appearance to moderate the

Bailenson et al 385

effect of behavioral realism such that behavioral real-ism would matter more for less human-looking em-bodied agents than for more human-looking embod-ied agents Finally we expected to demonstrateconvergence in our dependent measures such thatthey would correlate with one another and change inthe same way as a function of our independent vari-able manipulations

32 Perceived Copresence Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movement vs mimic vs recorded) analysisof covariance controlling for participant gender4 re-vealed a significant main effect of behavioral realism onself-reported perceived copresence F (3 133) 272p 05 2 06 As Figure 4 demonstrates perceivedcopresence was higher in the random-movements con-dition compared to the static condition p 01 How-ever this main effect of behavioral realism was qualifiedby a significant representation type by behavioral-realisminteraction F (6 133) 218 p 05 2 09

Simple main effects tests revealed that within theblockhead condition perceived copresence varied signif-icantly as a function of behavioral realism F (3 133)

339 p 05 2 07 Specifically perceived copres-ence was significantly higher when the blockhead repre-sentation moved according to our random-movementalgorithm compared with either the static or mimic con-ditions ps 05 Simple main effects tests also revealedthat perceived copresence varied as a function of behav-

ioral realism for participants in the teddy-bear condi-tion F (3 133) 330 p 05 2 07 Specificallyperceived copresence was significantly higher in mimiccondition than in either the static or recorded condi-tions ps 05 and higher in the random-movementscondition relative to the recorded and static conditions(ps 05 and 10 respectively)

In contrast there were no significant differences inperceived copresence between any of the four behavioral-realism conditions for participants in the humanldquootherrdquo condition This pattern of findings is consis-tent with our prediction that behavioral realismwould have less of an influence on perceived copres-ence for embodied agents that appear human ascompared to embodied agents with a nonhuman ap-pearance In addition it is interesting to note that thetype of head movement that elicited the greatest levelof perceived copresence varied as a function of the em-bodied agentrsquos appearance with participants attributinghigh copresence ratings to both mimicking and ran-domly moving teddy-bear representations but only ran-domly moving blockhead representations

4 Although both male and female participants were used in thisstudy gender was not included as a separate factor in analyses involv-ing both behavioral realism and representation type When randomlyassigning participants to conditions we did not block participants bygender and as a result when behavioral realism and representationtype were fully crossed there were not enough participants of bothgenders in all 12 experimental conditions to reliably compute a three-way analysis of variance Thus our data analytic technique involvedincluding participant gender as a covariate in our analyses involvingboth behavioral realism and representation type in order to control forthe potential effect of participant gender The results do not changeappreciably if we remove gender as a covariate in our analyses but weinclude gender as a covariate to provide the most thorough analysispossible

Figure 4 Mean perceived-copresence ratings (using gender-

covariate adjusted means) as a function of representation type and

behavioral realism Conditions that share the same letter do not differ

within that level of representation type at p 05

386 PRESENCE VOLUME 14 NUMBER 4

33 Embarrassment Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender re-vealed a significant main effect of representation type onparticipantsrsquo self-reported willingness to perform embar-rassing acts in front of the embodied agent (ie embar-rassment) F (2 133) 421 p 05 2 06 Spe-cifically participants in the human embodied-agentcondition (M 260 SD 155) were significantlyless willing to perform embarrassing acts in front ofthe ldquootherrdquo than were participants in the teddy-bear(M 351 SD 154) embodied-agent conditionp 05 Participants in the human-representationcondition were also marginally less willing to performembarrassing acts than participants in the blockhead(M 313 SD 156) condition p 10 In sumthe presence of a human-looking representationseemed to reduce participantsrsquo willingness to performin embarrassing acts Regardless of the behavioral re-alism of the representation embodied agents wholooked like humans elicited greater social influencethan did non-human-looking representations

34 Likability Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender alsorevealed a significant main effect of representation typeon self-reported likability ratings F (2 133) 543p 01 2 08 Specifically both the human-lookingrepresentation (M 250 SD 109) and the teddy-bear representation (M 228 SD 109) were ratedas significantly more likable than the blockhead repre-sentation (M 179 SD 110) ps 05 Partici-pants found the blockhead representation (which wasthe least human looking of the three representations) tobe less attractive than either of the other two represen-tation types and they liked the blockhead representationthe least

35 Correlations between PerceivedCopresence Embarrassment andLikability Ratings

In order to examine the relationships among per-ceived copresence embarrassment and likability a se-ries of Pearson correlations was computed Consistentwith out predictions results of these analyses revealed asignificant negative correlation between perceived co-presence and a participantrsquos willingness to perform em-barrassing acts in front of the embodied agent r(144) 27 p 001 as well as a significant positivecorrelation between perceived copresence and likabilityr (144) 31 p 001 Specifically we found that thehigher the perceived copresence experienced the lesswilling participants were to perform embarrassing acts infront of the embodied agent (r2 07) and the moreparticipants liked the embodied agent (r2 10) Therewas no significant correlation between likability and em-barrassment ratings

36 Memory

We computed a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender to ex-amine the effects of our independent variables on partic-ipantsrsquo memory Results of these analyses revealed a sig-nificant main effect of behavioral realism on memory F(3 133) 414 p 01 2 09 Specifically mem-ory was significantly worse in the mimic condition thanin either the static or random-movement conditionsps 05 suggesting that overall participants had theworst memory for characters worn by a representationthat mimicked their own movements with a 4-s delayAgain however this main effect was qualified by a sig-nificant representation type by behavioral-realism inter-action (see Figure 5) F (6 133) 270 p 02 2

11Simple main effects tests revealed that within the

blockhead condition memory varied significantly as afunction of behavioral realism F (3 133) 271p 05 2 06 Specifically individuals rememberedsignificantly more characters when they were in the

Bailenson et al 387

random-movement condition than in either the mimicor the recorded conditions (the two behavioral-realismconditions that appeared most humanlike) ps 05and participants in the random-movement conditionremembered marginally more than participants in thestatic condition p 10 These results converge quitewell with the perceived copresence data in which partic-ipants indicated higher copresence ratings for blockheadrepresentations only when the blockhead demonstratedrandom head movements Interestingly this suggeststhat for representations who look the least human theleast human-looking head movements were seen asmost appropriate This finding that avoiding mis-matches is the most effective manner of combining anembodied agentrsquos appearance and behavior is consistentwith previous research on visually embodied agents(Slater amp Steed 2001 Garau et al 2003) as well asvoice agents (Lee amp Nass 2003)

Furthermore the convergence of memory andperceived-copresence ratings provides some evidencethat better memory is indicative of higher copresenceIn other words the more copresence an embodiedagent elicits the more a person interacting with that

embodied agent remembers about it It is interesting tonote however that despite the similar patterns of find-ings between the perceived-copresence data and thememory data results of a Pearson correlation revealedthat there was no significant linear relationship betweenperceived copresence and peoplersquos ability to rememberinformation about an embodied agent

There was however a significant simple main effectof behavioral realism within the human-representationcondition F (3 133) 549 p 001 2 11Specifically participants in the mimic and random-movement conditions remembered significantly fewercharacters than participants in either the static or re-corded conditions ps 05 Participants in the teddy-bear condition displayed the same pattern but in thatcondition none of the differences reached statisticalsignificance

These results suggest that participants may have beendistracted by a human-looking embodied agent thatmimicked their own movements which in turn af-fected their ability to notice and recall information inthe virtual environment Moreover participants mayalso have been distracted in the random-movement con-dition since this condition represented the greatest mis-match between the appearance (human looking) andbehavior (not human looking) of the embodied agentFurther research should examine these potential expla-nations as they suggest that different processes maybe involved for different algorithms for generatingembodied-agent head movements

37 Interpersonal Distance

There were no significant effects of behavioral re-alism or representation type on minimum distance (ourfirst measure of interpersonal distance) Thus partici-pants in all conditions maintained equivalent levels ofinterpersonal distance between themselves and the em-bodied agent There was also no significant correlationbetween perceived copresence and either of our behav-ioral dependent measures (ie minimum distance andnumber of returns)

However as hypothesized results of a 3 (representa-tion type human vs teddy bear vs blockhead) 4 (be-

Figure 5 Gender-covariate adjusted mean number of correctly

recalled characters out of eight on a surprise memory test as a

function of representation type and behavioral realism Conditions that

share the same letter do not differ within that level of representation

type at p 05

388 PRESENCE VOLUME 14 NUMBER 4

havioral realism static vs random movements vs mimicvs recorded) analysis of covariance controlling for par-ticipant gender revealed a significant main effect of be-havioral realism on reversal count or the number oftimes a participant changed direction along the z-axisduring his or her examination of the embodied agentF(3 133) 306 p 05 2 07 As can be seen inFigure 6 participants in the random-movement condi-tion exhibited significantly fewer reversals than partici-pants in any of the other three behavioral-realism condi-tions Pairwise comparisons indicated that there weresignificantly fewer reversals in the random-movementcondition compared to the other three conditions(ps 05) which are not significantly different fromeach other

This finding in part confirms our hypothesis con-cerning our behavioral dependent variable For an em-bodied agent whose head moves with curves randomlygenerated by the computer participants show little tono hesitation in approaching In contrast when thehead movements were controlled by a real person (ei-ther a mimic of their own movements or the move-ments from another participant) participants showedgreater hesitation in approaching the embodied agentThus computer-generated movements generated lesshesitancy than human-generated movements

Given this explanation however one might expecteven less hesitation that is fewer reversals in approach-ing the static representation This finding did not occurand one potential explanation is that overall a ran-domly moving representation may elicit less social influ-ence and copresence than a completely static representa-tion In other words the static representation has atleast the potential to exhibit realistic movements whilethe unrealistically moving embodied agent immediatelylowers the sense of copresence In sum participantsrsquohesitancy in approaching an embodied agent was influ-enced by the behavioral realism of the embodied agentbut not its appearance

4 Conclusions

41 Summary of Findings

To summarize the appearance and behavior ofembodied agents had varying effects on the dependentvariables In many instances there was convergence inthat self-report cognitive and behavioral measureschanged in similar ways as a function of our indepen-dent variable manipulations For example the perceived-copresence likability and embarrassment measures cor-related in the predicted directions and the blockheadrepresentation elicited similar reactions across behaviortypes for both the memory measure and the perceived-copresence measure In other instances however therewas a lack of convergence where varying realism pro-duced different outcomes for the various responsemodes One inevitable challenge facing researchers whoexamine a large number of dependent variables is tounderstand the instances in which those dependentvariables do and do not converge

In the current study for some of our dependent mea-sures of copresence (willingness to commit embarrassingacts and likability) manipulation of representation typeaffected copresence responses On other dependentmeasures (self-reported copresence ratings interper-sonal distance approach reversals and memory) ma-nipulation of behavioral realism changed participantsrsquoresponses One way to interpret these results is thatself-report measures were adequate for measuring

Figure 6 Mean number of direction reversals while approaching the

embodied agent as a function of behavioral realism

Bailenson et al 389

copresence levels as affected by variation in appear-ance whereas self-report measures were not sensitiveenough to detect the subtle differences elicited by thebehavioral-realism manipulation Indeed the only reliablemain effects of the behavioral-realism manipulation oc-curred with cognitive and behavioral dependent measures

In other words self-report questionnaires are effectivewhen measuring how people perceive an embodiedagent but not necessarily how they will respond to thatembodied agent Consequently not only does it matterhow one defines copresence it also matters how oneoperationalizes copresence If the goal of a researcher isto understand the way in which people fathom embod-ied agents on a conceptual level then self-report ques-tionnaires may be the best tool However if the goal ofthe researcher is to understand the way in which peoplewill interact with learn from or behave toward anagent then he or she should augment questionnaireswith other types of measures In this paper we are notlobbying for the abandonment of copresence self-reportquestionnaires however we do urge that researchersuse these questionnaires as a complement to other typesof measures whenever possible

While this explanation that self-report measures lacksensitivity receives support from the current data itmust be qualified On most measures there was an in-teraction between behavioral realism and representationtype At times these interactions did not conform toour specific predictions and were difficult to interpretPart of the reason for this difficulty is most likely due tothe fact that we assumed a priori that certain types ofhead movements would be viewed as more realistic thanothers Specifically we assumed that the measures basedon actual human behaviors (mimic and recorded) wouldappear more realistic than either those generated by thecomputer (random) or than no movements at all Re-corded human movements applied to 3D models how-ever may not always be the most effective method ofrendering realistic behaviors and the differences be-tween our various levels of behavioral realism may havebeen too subtle in the current study As previous workhas demonstrated (Garau et al 2003) the interactionbetween appearance and behavioral realism is compli-cated and the finding that seems to be emerging is that

large disparities between the two types of realism cancause diminished levels of copresence Only when theblockhead representation the least humanlike of thethree demonstrated unrealistic random head move-ments did it elicit high levels of self-report and cognitivecopresence

Our findings with respect to the main effects of ap-pearance on self-report measures suggest that how hu-man an embodied agent looks may also be an importantprecondition for experiencing copresence This findingis unanticipated given predictions made by the Blasco-vich model and given findings from previous work(Bailenson et al 2001 Garau et al 2003) It seemsthat virtual representations that look like a realistic hu-man as opposed to either a familiar or unfamiliar non-human elicit embarrassment reactions that would beappropriate only in the presence of a real human

42 Limitations and Future Research

Though this effect surfaces for representations thatare humanlike to varying degrees how this effect of ap-pearance would apply to interactions with multiple hu-man representations of varying photographic realismremains an empirical question One key direction forfuture research is to examine how variation in photo-graphic realism (ie visual fidelity) of humanlike em-bodied agents affects copresence and how it interactswith behavioral realism Furthermore another impor-tant direction is to begin future studies with pretestedlevels of all realism variables (behavioral photographicand others) to assist in a more thorough interpretationof results Moreover the current study is limited in thatwe implemented behavioral realism solely by varyinghead movements A more thorough investigation of thisphenomenon should investigate other types of behav-iors such as facial gestures hand movements and gaitas well as more complex behaviors such as planning andcoordinated actions In addition the interaction task wechose (a purely nonverbal task) is limited and researchshould definitely examine copresence in richer social-interaction settings Finally our method of translationusing the game pad as opposed to naturally trackedwalking was not a perfect behavioral measure since

390 PRESENCE VOLUME 14 NUMBER 4

people do not translate by pushing buttons in the realworld Future work should provide a further examinationof copresence using more natural behavioral measures

Importantly had we measured only self-report re-sponses we would have demonstrated a different patternof results in the current study Attempting to augment anystudy geared toward understanding human interactionwith embodied agents or avatars with dependent measuresfrom multiple response channels certainly can make datainterpretation more complicated Patterns from self-reportdata that seem clean and consistent often are not replicatedwhen a less direct measurement paradigm is implementedFurthermore when asking peripheral implicitly relatedquestions about presence such as likability or embarrass-ment one risks measuring constructs inside the humanmind that are not ldquocentralrdquo to the definition of copres-ence Adding more implicit and disparate measures maynot paint a cleaner picture but it does paint a more com-plete one The challenge to the researcher is interpretingthese complex patterns in the data in order to assess amore thorough understanding of an extremely novel andabstract phenomenon of interacting with a virtual digitalentity

It may be the case that copresence is a purely affectiveconstruct inside the human mind alternatively copres-ence may be more of a cognitive construct In this earlystage of the research field examining copresence in vir-tual environments the theoretical and empirical work isextremely limited and does not allow one to rule outeither of these two (or many other) possibilities Deter-mining the exact conceptual nature of an abstract con-struct in the human mind is a challenge that faces manysocial scientists The measurement framework we offerallows the researcher to get a more complete picture ofthe copresence construct and allows for future researchto utilize more specific types of construct-type hypothe-ses as well as measurement strategies

In addition to the basic research implications with re-spect to the use of various modes of dependent measuresthe current study also has some important applied implica-tions By understanding the components of a virtual repre-sentation that lead to elicitations of heightened copresencewith that representation we can better develop our virtualagents and virtual worlds to immerse and engage the user

As we as well as others have demonstrated creating highlevels of copresence is a complicated process involving anappropriate fit between levels of each type of realism Aswe grow to better understand these interactions we canbetter match agent design to context and perhaps reach alevel where it becomes easy to use agents to develop thedegree of copresence usually found with virtual avatarsand perhaps in the long term the degree of copresencefound with flesh-and-blood people during face-to-faceinteraction

References

Bailenson J N Aharoni E Beall A C Guadagno R Di-mov A amp Blascovich J (2004) Comparing behavioraland self-report measures of embodied agents social presence inimmersive virtual environments Proceedings of the 7th An-nual International Workshop on Presence Valencia Spain

Bailenson J N Beall A C amp Blascovich J (2002) Mutualgaze and task performance in shared virtual environmentsJournal of Visualization and Computer Animation 13 1ndash8

Bailenson J N Beall A C Loomis J Blascovich J ampTurk M (2004) Transformed social interaction Decou-pling representation from behavior and form in collabora-tive virtual environments Presence Teleoperators and Vir-tual Environments 13(4) 428ndash441

Bailenson J N Blascovich J Beall A C amp Loomis J(2001) Equilibrium revisited Mutual gaze and personalspace in virtual environments Presence Teleoperators andVirtual Environments 10 583ndash598

Bailenson J N Blascovich J Beall A C amp Loomis J M(2003) Interpersonal distance in immersive virtual environ-ments Personality and Social Psychology Bulletin 29 1ndash 15

Bailenson J N amp Yee N (in press) Digital chameleonsAutomatic assimilation of nonverbal gestures in immersivevirtual environments Psychological Science

Beall A C Bailenson J N Loomis J Blascovich J ampRex C (2003) Non-zero-sum mutual gaze in collaborativevirtual environments Proceedings of HCI InternationalCrete

Bente G Ruggenberg S Tietz B amp Wortberg S (2004)Measuring behavioral correlates of social presence in virtualencounters Paper presented at the International Communi-cation Association Conference May 27ndash31

Biocca F (1997) Cyborgrsquos dilemma Progressive embodi-

Bailenson et al 391

23 Equipment

Participants wore a Virtual Research V8 stereo-scopic head-mounted display (HMD see Figure 3) TheHMD featured dual 680 horizontal by 480 vertical pixelresolution LCD panels that refreshed at 72 Hz Thedisplay optics presented a visual field subtending approxi-mately 50 degrees horizontally by 38 degrees vertically

Perspectively correct stereoscopic images were ren-dered by a 450 MHz Pentium III dual-processor com-puter with an Evans amp Sutherland Tornado 3000 dualpipe graphics card and were updated at an averageframe rate of 36 Hz The simulated viewpoint was con-tinually updated as a function of the participantsrsquo headmovements The orientation of the participantrsquos headwas tracked by a three-axis orientation sensing system(Intersense IS300 update rate of 150 Hz) The systemlatency or the amount of delay between a participantrsquoshead movement and the resulting concomitant updatein the HMDrsquos visual display was 65 ms maximum Thesoftware used to assimilate the rendering and trackingwas Vizard 20 Participants used a Logitech Rumble-

Pad Pro game pad to interact with the virtual environ-ment

24 Procedure

When participants arrived at the laboratory anexperimenter sat them down and instructed them onhow to wear and adjust the HMD and how to use thegame pad depicted in Figure 3 to interact with thevirtual environment Once immersed participantsfound themselves seated at a table directly acrossfrom an embodied agent (see Figure 2) Dependingon the assigned condition the embodied agent variedin appearance and behavior Using the game pad par-ticipants scrolled through the following instructionaltext that appeared above the embodied agentrsquos head

In this laboratory we are interested in peoplersquos reac-tions to ldquoothersrdquo in virtual environments In thisstudy you will spend time with an ldquootherrdquo After ashort time we will ask you questions about theldquootherrdquo You will use the game pad to answer thequestions The questions measure your agreementThe more you agree the higher the number youshould choose on the rating scale in front of you

We carefully chose and repeatedly used the wordldquootherrdquo set off by quotation marks to avoid using sug-gestive nouns such as ldquopersonrdquo or ldquoagentrdquo Participantswere next instructed to look directly at the ldquootherrdquo for90 s Next a series of 10 copresence embarrassment andlikability self-report questionnaire items (see AppendixA) appeared one at a time over the embodied agentrsquoshead in random order Participants indicated theiragreement with these items by using the game pad toselect a response on a Likert-type rating scale that ap-peared under each item Hence ratings were providedin real time as participants experienced the embodiedagent

After completing the items participants were in-structed to approach the embodied agent for furtherexamination They used buttons on the game pad tomove toward (positive translation on the z axis) or awayfrom (negative translation on the z axis) the embodiedagent with each press of a game-pad button equivalentto a jump of 25 cm (the starting distance between the

Figure 3 A depiction of our immersive virtual environment system

The components are (1) orientation tracking sensor (2) image

generator (3) HMD and (4) game pad input device

384 PRESENCE VOLUME 14 NUMBER 4

participantrsquos nose and the embodied agentrsquos nose wasapproximately 3 m) We did not render the body of par-ticipants so there was no leaning or walking animationduring the translationmdashonly their viewpoint movedcloser or farther from the agent After participants fin-ished approaching the embodied agent and examiningit they took off the HMD and filled out a demographicquestionnaire Finally we administered a surprise mem-ory test in which participants were asked to recall therandom string of characters written on the embodiedagentrsquos chest (see Figure 2)

25 Measures

251 Perceived Copresence Perceived copres-ence was assessed via a three-item self-report measureWe created a composite copresence measure ( 71see Appendix A) by averaging the three copresencequestions Higher scores indicated higher levels ofself-reported copresence where ratings were made on aLikert-type scale with response options ranging from 0(strongly disagree) to 6 (strongly agree)

252 Embarrassment Embarrassment was as-sessed as a self-reported social response to the presenceof an embodied agent Previous research has demon-strated that willingness to perform embarrassing acts infront of a virtual representation is a worthy measure ofthat representationrsquos degree of social influence (Bailen-son et al 2003) We averaged the three embarrassmentquestions ( 72 see Appendix A) into a single com-posite measure with higher numbers indicating greaterwillingness to perform embarrassing acts in front of anembodied agent

253 Likability Likability was assessed as a sec-ond self-reported social response to the presence ofan embodied agent We averaged the four likabilityquestions ( 71 see Appendix A) into a singlecomposite measure with higher numbers indicatingthat participants liked the virtual representation more

254 Memory Memory was quantified asthe total number of correct characters (out of a possi-

ble eight) a participant recalled from the virtual rep-resentationrsquos label without penalty for incorrect an-swers

255 Interpersonal Distance Interpersonaldistance was included as a behavioral social responseto the presence of an embodied agent We assessedtwo indicators of interpersonal distanceminimumdistance and reversal count Minimum distance wasthe closest distance the participant came to the em-bodied agent reversal count referred to the numberof times a participant changed direction along thez-axis (ie went back and forth) during their exami-nation of the embodied agent We consider reversalcount to be a proxy for hesitation in approaching theembodied agent

3 Results and Discussion

31 Hypotheses

We hypothesized that greater copresence wouldbe indicated by higher perceived copresence ratingsgreater liking of the embodied agent less willingness tocommit embarrassing acts in front of the embodiedagent the maintenance of greater interpersonal dis-tance and more hesitation in approaching the embod-ied agent We did not however have any specific pre-dictions with respect to the relationship betweencopresence and memory

Based on the results from previous copresencestudies we expected greater copresence (across ourvarious copresence measures) to occur with realisti-cally behaving embodied agents (specifically our re-corded and mimic behavioral-realism conditions thatfeatured movements based on actual human motion)compared with unrealistically behaving embodiedagents (ie our static and random-movement condi-tions) In addition we expected greater copresence tooccur with more human-looking embodied agentsthan with less human-looking ones Drawing on pre-dictions derived from the Blascovich et al (2002)threshold model of social influence however we ex-pected embodied-agent appearance to moderate the

Bailenson et al 385

effect of behavioral realism such that behavioral real-ism would matter more for less human-looking em-bodied agents than for more human-looking embod-ied agents Finally we expected to demonstrateconvergence in our dependent measures such thatthey would correlate with one another and change inthe same way as a function of our independent vari-able manipulations

32 Perceived Copresence Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movement vs mimic vs recorded) analysisof covariance controlling for participant gender4 re-vealed a significant main effect of behavioral realism onself-reported perceived copresence F (3 133) 272p 05 2 06 As Figure 4 demonstrates perceivedcopresence was higher in the random-movements con-dition compared to the static condition p 01 How-ever this main effect of behavioral realism was qualifiedby a significant representation type by behavioral-realisminteraction F (6 133) 218 p 05 2 09

Simple main effects tests revealed that within theblockhead condition perceived copresence varied signif-icantly as a function of behavioral realism F (3 133)

339 p 05 2 07 Specifically perceived copres-ence was significantly higher when the blockhead repre-sentation moved according to our random-movementalgorithm compared with either the static or mimic con-ditions ps 05 Simple main effects tests also revealedthat perceived copresence varied as a function of behav-

ioral realism for participants in the teddy-bear condi-tion F (3 133) 330 p 05 2 07 Specificallyperceived copresence was significantly higher in mimiccondition than in either the static or recorded condi-tions ps 05 and higher in the random-movementscondition relative to the recorded and static conditions(ps 05 and 10 respectively)

In contrast there were no significant differences inperceived copresence between any of the four behavioral-realism conditions for participants in the humanldquootherrdquo condition This pattern of findings is consis-tent with our prediction that behavioral realismwould have less of an influence on perceived copres-ence for embodied agents that appear human ascompared to embodied agents with a nonhuman ap-pearance In addition it is interesting to note that thetype of head movement that elicited the greatest levelof perceived copresence varied as a function of the em-bodied agentrsquos appearance with participants attributinghigh copresence ratings to both mimicking and ran-domly moving teddy-bear representations but only ran-domly moving blockhead representations

4 Although both male and female participants were used in thisstudy gender was not included as a separate factor in analyses involv-ing both behavioral realism and representation type When randomlyassigning participants to conditions we did not block participants bygender and as a result when behavioral realism and representationtype were fully crossed there were not enough participants of bothgenders in all 12 experimental conditions to reliably compute a three-way analysis of variance Thus our data analytic technique involvedincluding participant gender as a covariate in our analyses involvingboth behavioral realism and representation type in order to control forthe potential effect of participant gender The results do not changeappreciably if we remove gender as a covariate in our analyses but weinclude gender as a covariate to provide the most thorough analysispossible

Figure 4 Mean perceived-copresence ratings (using gender-

covariate adjusted means) as a function of representation type and

behavioral realism Conditions that share the same letter do not differ

within that level of representation type at p 05

386 PRESENCE VOLUME 14 NUMBER 4

33 Embarrassment Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender re-vealed a significant main effect of representation type onparticipantsrsquo self-reported willingness to perform embar-rassing acts in front of the embodied agent (ie embar-rassment) F (2 133) 421 p 05 2 06 Spe-cifically participants in the human embodied-agentcondition (M 260 SD 155) were significantlyless willing to perform embarrassing acts in front ofthe ldquootherrdquo than were participants in the teddy-bear(M 351 SD 154) embodied-agent conditionp 05 Participants in the human-representationcondition were also marginally less willing to performembarrassing acts than participants in the blockhead(M 313 SD 156) condition p 10 In sumthe presence of a human-looking representationseemed to reduce participantsrsquo willingness to performin embarrassing acts Regardless of the behavioral re-alism of the representation embodied agents wholooked like humans elicited greater social influencethan did non-human-looking representations

34 Likability Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender alsorevealed a significant main effect of representation typeon self-reported likability ratings F (2 133) 543p 01 2 08 Specifically both the human-lookingrepresentation (M 250 SD 109) and the teddy-bear representation (M 228 SD 109) were ratedas significantly more likable than the blockhead repre-sentation (M 179 SD 110) ps 05 Partici-pants found the blockhead representation (which wasthe least human looking of the three representations) tobe less attractive than either of the other two represen-tation types and they liked the blockhead representationthe least

35 Correlations between PerceivedCopresence Embarrassment andLikability Ratings

In order to examine the relationships among per-ceived copresence embarrassment and likability a se-ries of Pearson correlations was computed Consistentwith out predictions results of these analyses revealed asignificant negative correlation between perceived co-presence and a participantrsquos willingness to perform em-barrassing acts in front of the embodied agent r(144) 27 p 001 as well as a significant positivecorrelation between perceived copresence and likabilityr (144) 31 p 001 Specifically we found that thehigher the perceived copresence experienced the lesswilling participants were to perform embarrassing acts infront of the embodied agent (r2 07) and the moreparticipants liked the embodied agent (r2 10) Therewas no significant correlation between likability and em-barrassment ratings

36 Memory

We computed a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender to ex-amine the effects of our independent variables on partic-ipantsrsquo memory Results of these analyses revealed a sig-nificant main effect of behavioral realism on memory F(3 133) 414 p 01 2 09 Specifically mem-ory was significantly worse in the mimic condition thanin either the static or random-movement conditionsps 05 suggesting that overall participants had theworst memory for characters worn by a representationthat mimicked their own movements with a 4-s delayAgain however this main effect was qualified by a sig-nificant representation type by behavioral-realism inter-action (see Figure 5) F (6 133) 270 p 02 2

11Simple main effects tests revealed that within the

blockhead condition memory varied significantly as afunction of behavioral realism F (3 133) 271p 05 2 06 Specifically individuals rememberedsignificantly more characters when they were in the

Bailenson et al 387

random-movement condition than in either the mimicor the recorded conditions (the two behavioral-realismconditions that appeared most humanlike) ps 05and participants in the random-movement conditionremembered marginally more than participants in thestatic condition p 10 These results converge quitewell with the perceived copresence data in which partic-ipants indicated higher copresence ratings for blockheadrepresentations only when the blockhead demonstratedrandom head movements Interestingly this suggeststhat for representations who look the least human theleast human-looking head movements were seen asmost appropriate This finding that avoiding mis-matches is the most effective manner of combining anembodied agentrsquos appearance and behavior is consistentwith previous research on visually embodied agents(Slater amp Steed 2001 Garau et al 2003) as well asvoice agents (Lee amp Nass 2003)

Furthermore the convergence of memory andperceived-copresence ratings provides some evidencethat better memory is indicative of higher copresenceIn other words the more copresence an embodiedagent elicits the more a person interacting with that

embodied agent remembers about it It is interesting tonote however that despite the similar patterns of find-ings between the perceived-copresence data and thememory data results of a Pearson correlation revealedthat there was no significant linear relationship betweenperceived copresence and peoplersquos ability to rememberinformation about an embodied agent

There was however a significant simple main effectof behavioral realism within the human-representationcondition F (3 133) 549 p 001 2 11Specifically participants in the mimic and random-movement conditions remembered significantly fewercharacters than participants in either the static or re-corded conditions ps 05 Participants in the teddy-bear condition displayed the same pattern but in thatcondition none of the differences reached statisticalsignificance

These results suggest that participants may have beendistracted by a human-looking embodied agent thatmimicked their own movements which in turn af-fected their ability to notice and recall information inthe virtual environment Moreover participants mayalso have been distracted in the random-movement con-dition since this condition represented the greatest mis-match between the appearance (human looking) andbehavior (not human looking) of the embodied agentFurther research should examine these potential expla-nations as they suggest that different processes maybe involved for different algorithms for generatingembodied-agent head movements

37 Interpersonal Distance

There were no significant effects of behavioral re-alism or representation type on minimum distance (ourfirst measure of interpersonal distance) Thus partici-pants in all conditions maintained equivalent levels ofinterpersonal distance between themselves and the em-bodied agent There was also no significant correlationbetween perceived copresence and either of our behav-ioral dependent measures (ie minimum distance andnumber of returns)

However as hypothesized results of a 3 (representa-tion type human vs teddy bear vs blockhead) 4 (be-

Figure 5 Gender-covariate adjusted mean number of correctly

recalled characters out of eight on a surprise memory test as a

function of representation type and behavioral realism Conditions that

share the same letter do not differ within that level of representation

type at p 05

388 PRESENCE VOLUME 14 NUMBER 4

havioral realism static vs random movements vs mimicvs recorded) analysis of covariance controlling for par-ticipant gender revealed a significant main effect of be-havioral realism on reversal count or the number oftimes a participant changed direction along the z-axisduring his or her examination of the embodied agentF(3 133) 306 p 05 2 07 As can be seen inFigure 6 participants in the random-movement condi-tion exhibited significantly fewer reversals than partici-pants in any of the other three behavioral-realism condi-tions Pairwise comparisons indicated that there weresignificantly fewer reversals in the random-movementcondition compared to the other three conditions(ps 05) which are not significantly different fromeach other

This finding in part confirms our hypothesis con-cerning our behavioral dependent variable For an em-bodied agent whose head moves with curves randomlygenerated by the computer participants show little tono hesitation in approaching In contrast when thehead movements were controlled by a real person (ei-ther a mimic of their own movements or the move-ments from another participant) participants showedgreater hesitation in approaching the embodied agentThus computer-generated movements generated lesshesitancy than human-generated movements

Given this explanation however one might expecteven less hesitation that is fewer reversals in approach-ing the static representation This finding did not occurand one potential explanation is that overall a ran-domly moving representation may elicit less social influ-ence and copresence than a completely static representa-tion In other words the static representation has atleast the potential to exhibit realistic movements whilethe unrealistically moving embodied agent immediatelylowers the sense of copresence In sum participantsrsquohesitancy in approaching an embodied agent was influ-enced by the behavioral realism of the embodied agentbut not its appearance

4 Conclusions

41 Summary of Findings

To summarize the appearance and behavior ofembodied agents had varying effects on the dependentvariables In many instances there was convergence inthat self-report cognitive and behavioral measureschanged in similar ways as a function of our indepen-dent variable manipulations For example the perceived-copresence likability and embarrassment measures cor-related in the predicted directions and the blockheadrepresentation elicited similar reactions across behaviortypes for both the memory measure and the perceived-copresence measure In other instances however therewas a lack of convergence where varying realism pro-duced different outcomes for the various responsemodes One inevitable challenge facing researchers whoexamine a large number of dependent variables is tounderstand the instances in which those dependentvariables do and do not converge

In the current study for some of our dependent mea-sures of copresence (willingness to commit embarrassingacts and likability) manipulation of representation typeaffected copresence responses On other dependentmeasures (self-reported copresence ratings interper-sonal distance approach reversals and memory) ma-nipulation of behavioral realism changed participantsrsquoresponses One way to interpret these results is thatself-report measures were adequate for measuring

Figure 6 Mean number of direction reversals while approaching the

embodied agent as a function of behavioral realism

Bailenson et al 389

copresence levels as affected by variation in appear-ance whereas self-report measures were not sensitiveenough to detect the subtle differences elicited by thebehavioral-realism manipulation Indeed the only reliablemain effects of the behavioral-realism manipulation oc-curred with cognitive and behavioral dependent measures

In other words self-report questionnaires are effectivewhen measuring how people perceive an embodiedagent but not necessarily how they will respond to thatembodied agent Consequently not only does it matterhow one defines copresence it also matters how oneoperationalizes copresence If the goal of a researcher isto understand the way in which people fathom embod-ied agents on a conceptual level then self-report ques-tionnaires may be the best tool However if the goal ofthe researcher is to understand the way in which peoplewill interact with learn from or behave toward anagent then he or she should augment questionnaireswith other types of measures In this paper we are notlobbying for the abandonment of copresence self-reportquestionnaires however we do urge that researchersuse these questionnaires as a complement to other typesof measures whenever possible

While this explanation that self-report measures lacksensitivity receives support from the current data itmust be qualified On most measures there was an in-teraction between behavioral realism and representationtype At times these interactions did not conform toour specific predictions and were difficult to interpretPart of the reason for this difficulty is most likely due tothe fact that we assumed a priori that certain types ofhead movements would be viewed as more realistic thanothers Specifically we assumed that the measures basedon actual human behaviors (mimic and recorded) wouldappear more realistic than either those generated by thecomputer (random) or than no movements at all Re-corded human movements applied to 3D models how-ever may not always be the most effective method ofrendering realistic behaviors and the differences be-tween our various levels of behavioral realism may havebeen too subtle in the current study As previous workhas demonstrated (Garau et al 2003) the interactionbetween appearance and behavioral realism is compli-cated and the finding that seems to be emerging is that

large disparities between the two types of realism cancause diminished levels of copresence Only when theblockhead representation the least humanlike of thethree demonstrated unrealistic random head move-ments did it elicit high levels of self-report and cognitivecopresence

Our findings with respect to the main effects of ap-pearance on self-report measures suggest that how hu-man an embodied agent looks may also be an importantprecondition for experiencing copresence This findingis unanticipated given predictions made by the Blasco-vich model and given findings from previous work(Bailenson et al 2001 Garau et al 2003) It seemsthat virtual representations that look like a realistic hu-man as opposed to either a familiar or unfamiliar non-human elicit embarrassment reactions that would beappropriate only in the presence of a real human

42 Limitations and Future Research

Though this effect surfaces for representations thatare humanlike to varying degrees how this effect of ap-pearance would apply to interactions with multiple hu-man representations of varying photographic realismremains an empirical question One key direction forfuture research is to examine how variation in photo-graphic realism (ie visual fidelity) of humanlike em-bodied agents affects copresence and how it interactswith behavioral realism Furthermore another impor-tant direction is to begin future studies with pretestedlevels of all realism variables (behavioral photographicand others) to assist in a more thorough interpretationof results Moreover the current study is limited in thatwe implemented behavioral realism solely by varyinghead movements A more thorough investigation of thisphenomenon should investigate other types of behav-iors such as facial gestures hand movements and gaitas well as more complex behaviors such as planning andcoordinated actions In addition the interaction task wechose (a purely nonverbal task) is limited and researchshould definitely examine copresence in richer social-interaction settings Finally our method of translationusing the game pad as opposed to naturally trackedwalking was not a perfect behavioral measure since

390 PRESENCE VOLUME 14 NUMBER 4

people do not translate by pushing buttons in the realworld Future work should provide a further examinationof copresence using more natural behavioral measures

Importantly had we measured only self-report re-sponses we would have demonstrated a different patternof results in the current study Attempting to augment anystudy geared toward understanding human interactionwith embodied agents or avatars with dependent measuresfrom multiple response channels certainly can make datainterpretation more complicated Patterns from self-reportdata that seem clean and consistent often are not replicatedwhen a less direct measurement paradigm is implementedFurthermore when asking peripheral implicitly relatedquestions about presence such as likability or embarrass-ment one risks measuring constructs inside the humanmind that are not ldquocentralrdquo to the definition of copres-ence Adding more implicit and disparate measures maynot paint a cleaner picture but it does paint a more com-plete one The challenge to the researcher is interpretingthese complex patterns in the data in order to assess amore thorough understanding of an extremely novel andabstract phenomenon of interacting with a virtual digitalentity

It may be the case that copresence is a purely affectiveconstruct inside the human mind alternatively copres-ence may be more of a cognitive construct In this earlystage of the research field examining copresence in vir-tual environments the theoretical and empirical work isextremely limited and does not allow one to rule outeither of these two (or many other) possibilities Deter-mining the exact conceptual nature of an abstract con-struct in the human mind is a challenge that faces manysocial scientists The measurement framework we offerallows the researcher to get a more complete picture ofthe copresence construct and allows for future researchto utilize more specific types of construct-type hypothe-ses as well as measurement strategies

In addition to the basic research implications with re-spect to the use of various modes of dependent measuresthe current study also has some important applied implica-tions By understanding the components of a virtual repre-sentation that lead to elicitations of heightened copresencewith that representation we can better develop our virtualagents and virtual worlds to immerse and engage the user

As we as well as others have demonstrated creating highlevels of copresence is a complicated process involving anappropriate fit between levels of each type of realism Aswe grow to better understand these interactions we canbetter match agent design to context and perhaps reach alevel where it becomes easy to use agents to develop thedegree of copresence usually found with virtual avatarsand perhaps in the long term the degree of copresencefound with flesh-and-blood people during face-to-faceinteraction

References

Bailenson J N Aharoni E Beall A C Guadagno R Di-mov A amp Blascovich J (2004) Comparing behavioraland self-report measures of embodied agents social presence inimmersive virtual environments Proceedings of the 7th An-nual International Workshop on Presence Valencia Spain

Bailenson J N Beall A C amp Blascovich J (2002) Mutualgaze and task performance in shared virtual environmentsJournal of Visualization and Computer Animation 13 1ndash8

Bailenson J N Beall A C Loomis J Blascovich J ampTurk M (2004) Transformed social interaction Decou-pling representation from behavior and form in collabora-tive virtual environments Presence Teleoperators and Vir-tual Environments 13(4) 428ndash441

Bailenson J N Blascovich J Beall A C amp Loomis J(2001) Equilibrium revisited Mutual gaze and personalspace in virtual environments Presence Teleoperators andVirtual Environments 10 583ndash598

Bailenson J N Blascovich J Beall A C amp Loomis J M(2003) Interpersonal distance in immersive virtual environ-ments Personality and Social Psychology Bulletin 29 1ndash 15

Bailenson J N amp Yee N (in press) Digital chameleonsAutomatic assimilation of nonverbal gestures in immersivevirtual environments Psychological Science

Beall A C Bailenson J N Loomis J Blascovich J ampRex C (2003) Non-zero-sum mutual gaze in collaborativevirtual environments Proceedings of HCI InternationalCrete

Bente G Ruggenberg S Tietz B amp Wortberg S (2004)Measuring behavioral correlates of social presence in virtualencounters Paper presented at the International Communi-cation Association Conference May 27ndash31

Biocca F (1997) Cyborgrsquos dilemma Progressive embodi-

Bailenson et al 391

participantrsquos nose and the embodied agentrsquos nose wasapproximately 3 m) We did not render the body of par-ticipants so there was no leaning or walking animationduring the translationmdashonly their viewpoint movedcloser or farther from the agent After participants fin-ished approaching the embodied agent and examiningit they took off the HMD and filled out a demographicquestionnaire Finally we administered a surprise mem-ory test in which participants were asked to recall therandom string of characters written on the embodiedagentrsquos chest (see Figure 2)

25 Measures

251 Perceived Copresence Perceived copres-ence was assessed via a three-item self-report measureWe created a composite copresence measure ( 71see Appendix A) by averaging the three copresencequestions Higher scores indicated higher levels ofself-reported copresence where ratings were made on aLikert-type scale with response options ranging from 0(strongly disagree) to 6 (strongly agree)

252 Embarrassment Embarrassment was as-sessed as a self-reported social response to the presenceof an embodied agent Previous research has demon-strated that willingness to perform embarrassing acts infront of a virtual representation is a worthy measure ofthat representationrsquos degree of social influence (Bailen-son et al 2003) We averaged the three embarrassmentquestions ( 72 see Appendix A) into a single com-posite measure with higher numbers indicating greaterwillingness to perform embarrassing acts in front of anembodied agent

253 Likability Likability was assessed as a sec-ond self-reported social response to the presence ofan embodied agent We averaged the four likabilityquestions ( 71 see Appendix A) into a singlecomposite measure with higher numbers indicatingthat participants liked the virtual representation more

254 Memory Memory was quantified asthe total number of correct characters (out of a possi-

ble eight) a participant recalled from the virtual rep-resentationrsquos label without penalty for incorrect an-swers

255 Interpersonal Distance Interpersonaldistance was included as a behavioral social responseto the presence of an embodied agent We assessedtwo indicators of interpersonal distanceminimumdistance and reversal count Minimum distance wasthe closest distance the participant came to the em-bodied agent reversal count referred to the numberof times a participant changed direction along thez-axis (ie went back and forth) during their exami-nation of the embodied agent We consider reversalcount to be a proxy for hesitation in approaching theembodied agent

3 Results and Discussion

31 Hypotheses

We hypothesized that greater copresence wouldbe indicated by higher perceived copresence ratingsgreater liking of the embodied agent less willingness tocommit embarrassing acts in front of the embodiedagent the maintenance of greater interpersonal dis-tance and more hesitation in approaching the embod-ied agent We did not however have any specific pre-dictions with respect to the relationship betweencopresence and memory

Based on the results from previous copresencestudies we expected greater copresence (across ourvarious copresence measures) to occur with realisti-cally behaving embodied agents (specifically our re-corded and mimic behavioral-realism conditions thatfeatured movements based on actual human motion)compared with unrealistically behaving embodiedagents (ie our static and random-movement condi-tions) In addition we expected greater copresence tooccur with more human-looking embodied agentsthan with less human-looking ones Drawing on pre-dictions derived from the Blascovich et al (2002)threshold model of social influence however we ex-pected embodied-agent appearance to moderate the

Bailenson et al 385

effect of behavioral realism such that behavioral real-ism would matter more for less human-looking em-bodied agents than for more human-looking embod-ied agents Finally we expected to demonstrateconvergence in our dependent measures such thatthey would correlate with one another and change inthe same way as a function of our independent vari-able manipulations

32 Perceived Copresence Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movement vs mimic vs recorded) analysisof covariance controlling for participant gender4 re-vealed a significant main effect of behavioral realism onself-reported perceived copresence F (3 133) 272p 05 2 06 As Figure 4 demonstrates perceivedcopresence was higher in the random-movements con-dition compared to the static condition p 01 How-ever this main effect of behavioral realism was qualifiedby a significant representation type by behavioral-realisminteraction F (6 133) 218 p 05 2 09

Simple main effects tests revealed that within theblockhead condition perceived copresence varied signif-icantly as a function of behavioral realism F (3 133)

339 p 05 2 07 Specifically perceived copres-ence was significantly higher when the blockhead repre-sentation moved according to our random-movementalgorithm compared with either the static or mimic con-ditions ps 05 Simple main effects tests also revealedthat perceived copresence varied as a function of behav-

ioral realism for participants in the teddy-bear condi-tion F (3 133) 330 p 05 2 07 Specificallyperceived copresence was significantly higher in mimiccondition than in either the static or recorded condi-tions ps 05 and higher in the random-movementscondition relative to the recorded and static conditions(ps 05 and 10 respectively)

In contrast there were no significant differences inperceived copresence between any of the four behavioral-realism conditions for participants in the humanldquootherrdquo condition This pattern of findings is consis-tent with our prediction that behavioral realismwould have less of an influence on perceived copres-ence for embodied agents that appear human ascompared to embodied agents with a nonhuman ap-pearance In addition it is interesting to note that thetype of head movement that elicited the greatest levelof perceived copresence varied as a function of the em-bodied agentrsquos appearance with participants attributinghigh copresence ratings to both mimicking and ran-domly moving teddy-bear representations but only ran-domly moving blockhead representations

4 Although both male and female participants were used in thisstudy gender was not included as a separate factor in analyses involv-ing both behavioral realism and representation type When randomlyassigning participants to conditions we did not block participants bygender and as a result when behavioral realism and representationtype were fully crossed there were not enough participants of bothgenders in all 12 experimental conditions to reliably compute a three-way analysis of variance Thus our data analytic technique involvedincluding participant gender as a covariate in our analyses involvingboth behavioral realism and representation type in order to control forthe potential effect of participant gender The results do not changeappreciably if we remove gender as a covariate in our analyses but weinclude gender as a covariate to provide the most thorough analysispossible

Figure 4 Mean perceived-copresence ratings (using gender-

covariate adjusted means) as a function of representation type and

behavioral realism Conditions that share the same letter do not differ

within that level of representation type at p 05

386 PRESENCE VOLUME 14 NUMBER 4

33 Embarrassment Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender re-vealed a significant main effect of representation type onparticipantsrsquo self-reported willingness to perform embar-rassing acts in front of the embodied agent (ie embar-rassment) F (2 133) 421 p 05 2 06 Spe-cifically participants in the human embodied-agentcondition (M 260 SD 155) were significantlyless willing to perform embarrassing acts in front ofthe ldquootherrdquo than were participants in the teddy-bear(M 351 SD 154) embodied-agent conditionp 05 Participants in the human-representationcondition were also marginally less willing to performembarrassing acts than participants in the blockhead(M 313 SD 156) condition p 10 In sumthe presence of a human-looking representationseemed to reduce participantsrsquo willingness to performin embarrassing acts Regardless of the behavioral re-alism of the representation embodied agents wholooked like humans elicited greater social influencethan did non-human-looking representations

34 Likability Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender alsorevealed a significant main effect of representation typeon self-reported likability ratings F (2 133) 543p 01 2 08 Specifically both the human-lookingrepresentation (M 250 SD 109) and the teddy-bear representation (M 228 SD 109) were ratedas significantly more likable than the blockhead repre-sentation (M 179 SD 110) ps 05 Partici-pants found the blockhead representation (which wasthe least human looking of the three representations) tobe less attractive than either of the other two represen-tation types and they liked the blockhead representationthe least

35 Correlations between PerceivedCopresence Embarrassment andLikability Ratings

In order to examine the relationships among per-ceived copresence embarrassment and likability a se-ries of Pearson correlations was computed Consistentwith out predictions results of these analyses revealed asignificant negative correlation between perceived co-presence and a participantrsquos willingness to perform em-barrassing acts in front of the embodied agent r(144) 27 p 001 as well as a significant positivecorrelation between perceived copresence and likabilityr (144) 31 p 001 Specifically we found that thehigher the perceived copresence experienced the lesswilling participants were to perform embarrassing acts infront of the embodied agent (r2 07) and the moreparticipants liked the embodied agent (r2 10) Therewas no significant correlation between likability and em-barrassment ratings

36 Memory

We computed a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender to ex-amine the effects of our independent variables on partic-ipantsrsquo memory Results of these analyses revealed a sig-nificant main effect of behavioral realism on memory F(3 133) 414 p 01 2 09 Specifically mem-ory was significantly worse in the mimic condition thanin either the static or random-movement conditionsps 05 suggesting that overall participants had theworst memory for characters worn by a representationthat mimicked their own movements with a 4-s delayAgain however this main effect was qualified by a sig-nificant representation type by behavioral-realism inter-action (see Figure 5) F (6 133) 270 p 02 2

11Simple main effects tests revealed that within the

blockhead condition memory varied significantly as afunction of behavioral realism F (3 133) 271p 05 2 06 Specifically individuals rememberedsignificantly more characters when they were in the

Bailenson et al 387

random-movement condition than in either the mimicor the recorded conditions (the two behavioral-realismconditions that appeared most humanlike) ps 05and participants in the random-movement conditionremembered marginally more than participants in thestatic condition p 10 These results converge quitewell with the perceived copresence data in which partic-ipants indicated higher copresence ratings for blockheadrepresentations only when the blockhead demonstratedrandom head movements Interestingly this suggeststhat for representations who look the least human theleast human-looking head movements were seen asmost appropriate This finding that avoiding mis-matches is the most effective manner of combining anembodied agentrsquos appearance and behavior is consistentwith previous research on visually embodied agents(Slater amp Steed 2001 Garau et al 2003) as well asvoice agents (Lee amp Nass 2003)

Furthermore the convergence of memory andperceived-copresence ratings provides some evidencethat better memory is indicative of higher copresenceIn other words the more copresence an embodiedagent elicits the more a person interacting with that

embodied agent remembers about it It is interesting tonote however that despite the similar patterns of find-ings between the perceived-copresence data and thememory data results of a Pearson correlation revealedthat there was no significant linear relationship betweenperceived copresence and peoplersquos ability to rememberinformation about an embodied agent

There was however a significant simple main effectof behavioral realism within the human-representationcondition F (3 133) 549 p 001 2 11Specifically participants in the mimic and random-movement conditions remembered significantly fewercharacters than participants in either the static or re-corded conditions ps 05 Participants in the teddy-bear condition displayed the same pattern but in thatcondition none of the differences reached statisticalsignificance

These results suggest that participants may have beendistracted by a human-looking embodied agent thatmimicked their own movements which in turn af-fected their ability to notice and recall information inthe virtual environment Moreover participants mayalso have been distracted in the random-movement con-dition since this condition represented the greatest mis-match between the appearance (human looking) andbehavior (not human looking) of the embodied agentFurther research should examine these potential expla-nations as they suggest that different processes maybe involved for different algorithms for generatingembodied-agent head movements

37 Interpersonal Distance

There were no significant effects of behavioral re-alism or representation type on minimum distance (ourfirst measure of interpersonal distance) Thus partici-pants in all conditions maintained equivalent levels ofinterpersonal distance between themselves and the em-bodied agent There was also no significant correlationbetween perceived copresence and either of our behav-ioral dependent measures (ie minimum distance andnumber of returns)

However as hypothesized results of a 3 (representa-tion type human vs teddy bear vs blockhead) 4 (be-

Figure 5 Gender-covariate adjusted mean number of correctly

recalled characters out of eight on a surprise memory test as a

function of representation type and behavioral realism Conditions that

share the same letter do not differ within that level of representation

type at p 05

388 PRESENCE VOLUME 14 NUMBER 4

havioral realism static vs random movements vs mimicvs recorded) analysis of covariance controlling for par-ticipant gender revealed a significant main effect of be-havioral realism on reversal count or the number oftimes a participant changed direction along the z-axisduring his or her examination of the embodied agentF(3 133) 306 p 05 2 07 As can be seen inFigure 6 participants in the random-movement condi-tion exhibited significantly fewer reversals than partici-pants in any of the other three behavioral-realism condi-tions Pairwise comparisons indicated that there weresignificantly fewer reversals in the random-movementcondition compared to the other three conditions(ps 05) which are not significantly different fromeach other

This finding in part confirms our hypothesis con-cerning our behavioral dependent variable For an em-bodied agent whose head moves with curves randomlygenerated by the computer participants show little tono hesitation in approaching In contrast when thehead movements were controlled by a real person (ei-ther a mimic of their own movements or the move-ments from another participant) participants showedgreater hesitation in approaching the embodied agentThus computer-generated movements generated lesshesitancy than human-generated movements

Given this explanation however one might expecteven less hesitation that is fewer reversals in approach-ing the static representation This finding did not occurand one potential explanation is that overall a ran-domly moving representation may elicit less social influ-ence and copresence than a completely static representa-tion In other words the static representation has atleast the potential to exhibit realistic movements whilethe unrealistically moving embodied agent immediatelylowers the sense of copresence In sum participantsrsquohesitancy in approaching an embodied agent was influ-enced by the behavioral realism of the embodied agentbut not its appearance

4 Conclusions

41 Summary of Findings

To summarize the appearance and behavior ofembodied agents had varying effects on the dependentvariables In many instances there was convergence inthat self-report cognitive and behavioral measureschanged in similar ways as a function of our indepen-dent variable manipulations For example the perceived-copresence likability and embarrassment measures cor-related in the predicted directions and the blockheadrepresentation elicited similar reactions across behaviortypes for both the memory measure and the perceived-copresence measure In other instances however therewas a lack of convergence where varying realism pro-duced different outcomes for the various responsemodes One inevitable challenge facing researchers whoexamine a large number of dependent variables is tounderstand the instances in which those dependentvariables do and do not converge

In the current study for some of our dependent mea-sures of copresence (willingness to commit embarrassingacts and likability) manipulation of representation typeaffected copresence responses On other dependentmeasures (self-reported copresence ratings interper-sonal distance approach reversals and memory) ma-nipulation of behavioral realism changed participantsrsquoresponses One way to interpret these results is thatself-report measures were adequate for measuring

Figure 6 Mean number of direction reversals while approaching the

embodied agent as a function of behavioral realism

Bailenson et al 389

copresence levels as affected by variation in appear-ance whereas self-report measures were not sensitiveenough to detect the subtle differences elicited by thebehavioral-realism manipulation Indeed the only reliablemain effects of the behavioral-realism manipulation oc-curred with cognitive and behavioral dependent measures

In other words self-report questionnaires are effectivewhen measuring how people perceive an embodiedagent but not necessarily how they will respond to thatembodied agent Consequently not only does it matterhow one defines copresence it also matters how oneoperationalizes copresence If the goal of a researcher isto understand the way in which people fathom embod-ied agents on a conceptual level then self-report ques-tionnaires may be the best tool However if the goal ofthe researcher is to understand the way in which peoplewill interact with learn from or behave toward anagent then he or she should augment questionnaireswith other types of measures In this paper we are notlobbying for the abandonment of copresence self-reportquestionnaires however we do urge that researchersuse these questionnaires as a complement to other typesof measures whenever possible

While this explanation that self-report measures lacksensitivity receives support from the current data itmust be qualified On most measures there was an in-teraction between behavioral realism and representationtype At times these interactions did not conform toour specific predictions and were difficult to interpretPart of the reason for this difficulty is most likely due tothe fact that we assumed a priori that certain types ofhead movements would be viewed as more realistic thanothers Specifically we assumed that the measures basedon actual human behaviors (mimic and recorded) wouldappear more realistic than either those generated by thecomputer (random) or than no movements at all Re-corded human movements applied to 3D models how-ever may not always be the most effective method ofrendering realistic behaviors and the differences be-tween our various levels of behavioral realism may havebeen too subtle in the current study As previous workhas demonstrated (Garau et al 2003) the interactionbetween appearance and behavioral realism is compli-cated and the finding that seems to be emerging is that

large disparities between the two types of realism cancause diminished levels of copresence Only when theblockhead representation the least humanlike of thethree demonstrated unrealistic random head move-ments did it elicit high levels of self-report and cognitivecopresence

Our findings with respect to the main effects of ap-pearance on self-report measures suggest that how hu-man an embodied agent looks may also be an importantprecondition for experiencing copresence This findingis unanticipated given predictions made by the Blasco-vich model and given findings from previous work(Bailenson et al 2001 Garau et al 2003) It seemsthat virtual representations that look like a realistic hu-man as opposed to either a familiar or unfamiliar non-human elicit embarrassment reactions that would beappropriate only in the presence of a real human

42 Limitations and Future Research

Though this effect surfaces for representations thatare humanlike to varying degrees how this effect of ap-pearance would apply to interactions with multiple hu-man representations of varying photographic realismremains an empirical question One key direction forfuture research is to examine how variation in photo-graphic realism (ie visual fidelity) of humanlike em-bodied agents affects copresence and how it interactswith behavioral realism Furthermore another impor-tant direction is to begin future studies with pretestedlevels of all realism variables (behavioral photographicand others) to assist in a more thorough interpretationof results Moreover the current study is limited in thatwe implemented behavioral realism solely by varyinghead movements A more thorough investigation of thisphenomenon should investigate other types of behav-iors such as facial gestures hand movements and gaitas well as more complex behaviors such as planning andcoordinated actions In addition the interaction task wechose (a purely nonverbal task) is limited and researchshould definitely examine copresence in richer social-interaction settings Finally our method of translationusing the game pad as opposed to naturally trackedwalking was not a perfect behavioral measure since

390 PRESENCE VOLUME 14 NUMBER 4

people do not translate by pushing buttons in the realworld Future work should provide a further examinationof copresence using more natural behavioral measures

Importantly had we measured only self-report re-sponses we would have demonstrated a different patternof results in the current study Attempting to augment anystudy geared toward understanding human interactionwith embodied agents or avatars with dependent measuresfrom multiple response channels certainly can make datainterpretation more complicated Patterns from self-reportdata that seem clean and consistent often are not replicatedwhen a less direct measurement paradigm is implementedFurthermore when asking peripheral implicitly relatedquestions about presence such as likability or embarrass-ment one risks measuring constructs inside the humanmind that are not ldquocentralrdquo to the definition of copres-ence Adding more implicit and disparate measures maynot paint a cleaner picture but it does paint a more com-plete one The challenge to the researcher is interpretingthese complex patterns in the data in order to assess amore thorough understanding of an extremely novel andabstract phenomenon of interacting with a virtual digitalentity

It may be the case that copresence is a purely affectiveconstruct inside the human mind alternatively copres-ence may be more of a cognitive construct In this earlystage of the research field examining copresence in vir-tual environments the theoretical and empirical work isextremely limited and does not allow one to rule outeither of these two (or many other) possibilities Deter-mining the exact conceptual nature of an abstract con-struct in the human mind is a challenge that faces manysocial scientists The measurement framework we offerallows the researcher to get a more complete picture ofthe copresence construct and allows for future researchto utilize more specific types of construct-type hypothe-ses as well as measurement strategies

In addition to the basic research implications with re-spect to the use of various modes of dependent measuresthe current study also has some important applied implica-tions By understanding the components of a virtual repre-sentation that lead to elicitations of heightened copresencewith that representation we can better develop our virtualagents and virtual worlds to immerse and engage the user

As we as well as others have demonstrated creating highlevels of copresence is a complicated process involving anappropriate fit between levels of each type of realism Aswe grow to better understand these interactions we canbetter match agent design to context and perhaps reach alevel where it becomes easy to use agents to develop thedegree of copresence usually found with virtual avatarsand perhaps in the long term the degree of copresencefound with flesh-and-blood people during face-to-faceinteraction

References

Bailenson J N Aharoni E Beall A C Guadagno R Di-mov A amp Blascovich J (2004) Comparing behavioraland self-report measures of embodied agents social presence inimmersive virtual environments Proceedings of the 7th An-nual International Workshop on Presence Valencia Spain

Bailenson J N Beall A C amp Blascovich J (2002) Mutualgaze and task performance in shared virtual environmentsJournal of Visualization and Computer Animation 13 1ndash8

Bailenson J N Beall A C Loomis J Blascovich J ampTurk M (2004) Transformed social interaction Decou-pling representation from behavior and form in collabora-tive virtual environments Presence Teleoperators and Vir-tual Environments 13(4) 428ndash441

Bailenson J N Blascovich J Beall A C amp Loomis J(2001) Equilibrium revisited Mutual gaze and personalspace in virtual environments Presence Teleoperators andVirtual Environments 10 583ndash598

Bailenson J N Blascovich J Beall A C amp Loomis J M(2003) Interpersonal distance in immersive virtual environ-ments Personality and Social Psychology Bulletin 29 1ndash 15

Bailenson J N amp Yee N (in press) Digital chameleonsAutomatic assimilation of nonverbal gestures in immersivevirtual environments Psychological Science

Beall A C Bailenson J N Loomis J Blascovich J ampRex C (2003) Non-zero-sum mutual gaze in collaborativevirtual environments Proceedings of HCI InternationalCrete

Bente G Ruggenberg S Tietz B amp Wortberg S (2004)Measuring behavioral correlates of social presence in virtualencounters Paper presented at the International Communi-cation Association Conference May 27ndash31

Biocca F (1997) Cyborgrsquos dilemma Progressive embodi-

Bailenson et al 391

effect of behavioral realism such that behavioral real-ism would matter more for less human-looking em-bodied agents than for more human-looking embod-ied agents Finally we expected to demonstrateconvergence in our dependent measures such thatthey would correlate with one another and change inthe same way as a function of our independent vari-able manipulations

32 Perceived Copresence Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movement vs mimic vs recorded) analysisof covariance controlling for participant gender4 re-vealed a significant main effect of behavioral realism onself-reported perceived copresence F (3 133) 272p 05 2 06 As Figure 4 demonstrates perceivedcopresence was higher in the random-movements con-dition compared to the static condition p 01 How-ever this main effect of behavioral realism was qualifiedby a significant representation type by behavioral-realisminteraction F (6 133) 218 p 05 2 09

Simple main effects tests revealed that within theblockhead condition perceived copresence varied signif-icantly as a function of behavioral realism F (3 133)

339 p 05 2 07 Specifically perceived copres-ence was significantly higher when the blockhead repre-sentation moved according to our random-movementalgorithm compared with either the static or mimic con-ditions ps 05 Simple main effects tests also revealedthat perceived copresence varied as a function of behav-

ioral realism for participants in the teddy-bear condi-tion F (3 133) 330 p 05 2 07 Specificallyperceived copresence was significantly higher in mimiccondition than in either the static or recorded condi-tions ps 05 and higher in the random-movementscondition relative to the recorded and static conditions(ps 05 and 10 respectively)

In contrast there were no significant differences inperceived copresence between any of the four behavioral-realism conditions for participants in the humanldquootherrdquo condition This pattern of findings is consis-tent with our prediction that behavioral realismwould have less of an influence on perceived copres-ence for embodied agents that appear human ascompared to embodied agents with a nonhuman ap-pearance In addition it is interesting to note that thetype of head movement that elicited the greatest levelof perceived copresence varied as a function of the em-bodied agentrsquos appearance with participants attributinghigh copresence ratings to both mimicking and ran-domly moving teddy-bear representations but only ran-domly moving blockhead representations

4 Although both male and female participants were used in thisstudy gender was not included as a separate factor in analyses involv-ing both behavioral realism and representation type When randomlyassigning participants to conditions we did not block participants bygender and as a result when behavioral realism and representationtype were fully crossed there were not enough participants of bothgenders in all 12 experimental conditions to reliably compute a three-way analysis of variance Thus our data analytic technique involvedincluding participant gender as a covariate in our analyses involvingboth behavioral realism and representation type in order to control forthe potential effect of participant gender The results do not changeappreciably if we remove gender as a covariate in our analyses but weinclude gender as a covariate to provide the most thorough analysispossible

Figure 4 Mean perceived-copresence ratings (using gender-

covariate adjusted means) as a function of representation type and

behavioral realism Conditions that share the same letter do not differ

within that level of representation type at p 05

386 PRESENCE VOLUME 14 NUMBER 4

33 Embarrassment Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender re-vealed a significant main effect of representation type onparticipantsrsquo self-reported willingness to perform embar-rassing acts in front of the embodied agent (ie embar-rassment) F (2 133) 421 p 05 2 06 Spe-cifically participants in the human embodied-agentcondition (M 260 SD 155) were significantlyless willing to perform embarrassing acts in front ofthe ldquootherrdquo than were participants in the teddy-bear(M 351 SD 154) embodied-agent conditionp 05 Participants in the human-representationcondition were also marginally less willing to performembarrassing acts than participants in the blockhead(M 313 SD 156) condition p 10 In sumthe presence of a human-looking representationseemed to reduce participantsrsquo willingness to performin embarrassing acts Regardless of the behavioral re-alism of the representation embodied agents wholooked like humans elicited greater social influencethan did non-human-looking representations

34 Likability Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender alsorevealed a significant main effect of representation typeon self-reported likability ratings F (2 133) 543p 01 2 08 Specifically both the human-lookingrepresentation (M 250 SD 109) and the teddy-bear representation (M 228 SD 109) were ratedas significantly more likable than the blockhead repre-sentation (M 179 SD 110) ps 05 Partici-pants found the blockhead representation (which wasthe least human looking of the three representations) tobe less attractive than either of the other two represen-tation types and they liked the blockhead representationthe least

35 Correlations between PerceivedCopresence Embarrassment andLikability Ratings

In order to examine the relationships among per-ceived copresence embarrassment and likability a se-ries of Pearson correlations was computed Consistentwith out predictions results of these analyses revealed asignificant negative correlation between perceived co-presence and a participantrsquos willingness to perform em-barrassing acts in front of the embodied agent r(144) 27 p 001 as well as a significant positivecorrelation between perceived copresence and likabilityr (144) 31 p 001 Specifically we found that thehigher the perceived copresence experienced the lesswilling participants were to perform embarrassing acts infront of the embodied agent (r2 07) and the moreparticipants liked the embodied agent (r2 10) Therewas no significant correlation between likability and em-barrassment ratings

36 Memory

We computed a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender to ex-amine the effects of our independent variables on partic-ipantsrsquo memory Results of these analyses revealed a sig-nificant main effect of behavioral realism on memory F(3 133) 414 p 01 2 09 Specifically mem-ory was significantly worse in the mimic condition thanin either the static or random-movement conditionsps 05 suggesting that overall participants had theworst memory for characters worn by a representationthat mimicked their own movements with a 4-s delayAgain however this main effect was qualified by a sig-nificant representation type by behavioral-realism inter-action (see Figure 5) F (6 133) 270 p 02 2

11Simple main effects tests revealed that within the

blockhead condition memory varied significantly as afunction of behavioral realism F (3 133) 271p 05 2 06 Specifically individuals rememberedsignificantly more characters when they were in the

Bailenson et al 387

random-movement condition than in either the mimicor the recorded conditions (the two behavioral-realismconditions that appeared most humanlike) ps 05and participants in the random-movement conditionremembered marginally more than participants in thestatic condition p 10 These results converge quitewell with the perceived copresence data in which partic-ipants indicated higher copresence ratings for blockheadrepresentations only when the blockhead demonstratedrandom head movements Interestingly this suggeststhat for representations who look the least human theleast human-looking head movements were seen asmost appropriate This finding that avoiding mis-matches is the most effective manner of combining anembodied agentrsquos appearance and behavior is consistentwith previous research on visually embodied agents(Slater amp Steed 2001 Garau et al 2003) as well asvoice agents (Lee amp Nass 2003)

Furthermore the convergence of memory andperceived-copresence ratings provides some evidencethat better memory is indicative of higher copresenceIn other words the more copresence an embodiedagent elicits the more a person interacting with that

embodied agent remembers about it It is interesting tonote however that despite the similar patterns of find-ings between the perceived-copresence data and thememory data results of a Pearson correlation revealedthat there was no significant linear relationship betweenperceived copresence and peoplersquos ability to rememberinformation about an embodied agent

There was however a significant simple main effectof behavioral realism within the human-representationcondition F (3 133) 549 p 001 2 11Specifically participants in the mimic and random-movement conditions remembered significantly fewercharacters than participants in either the static or re-corded conditions ps 05 Participants in the teddy-bear condition displayed the same pattern but in thatcondition none of the differences reached statisticalsignificance

These results suggest that participants may have beendistracted by a human-looking embodied agent thatmimicked their own movements which in turn af-fected their ability to notice and recall information inthe virtual environment Moreover participants mayalso have been distracted in the random-movement con-dition since this condition represented the greatest mis-match between the appearance (human looking) andbehavior (not human looking) of the embodied agentFurther research should examine these potential expla-nations as they suggest that different processes maybe involved for different algorithms for generatingembodied-agent head movements

37 Interpersonal Distance

There were no significant effects of behavioral re-alism or representation type on minimum distance (ourfirst measure of interpersonal distance) Thus partici-pants in all conditions maintained equivalent levels ofinterpersonal distance between themselves and the em-bodied agent There was also no significant correlationbetween perceived copresence and either of our behav-ioral dependent measures (ie minimum distance andnumber of returns)

However as hypothesized results of a 3 (representa-tion type human vs teddy bear vs blockhead) 4 (be-

Figure 5 Gender-covariate adjusted mean number of correctly

recalled characters out of eight on a surprise memory test as a

function of representation type and behavioral realism Conditions that

share the same letter do not differ within that level of representation

type at p 05

388 PRESENCE VOLUME 14 NUMBER 4

havioral realism static vs random movements vs mimicvs recorded) analysis of covariance controlling for par-ticipant gender revealed a significant main effect of be-havioral realism on reversal count or the number oftimes a participant changed direction along the z-axisduring his or her examination of the embodied agentF(3 133) 306 p 05 2 07 As can be seen inFigure 6 participants in the random-movement condi-tion exhibited significantly fewer reversals than partici-pants in any of the other three behavioral-realism condi-tions Pairwise comparisons indicated that there weresignificantly fewer reversals in the random-movementcondition compared to the other three conditions(ps 05) which are not significantly different fromeach other

This finding in part confirms our hypothesis con-cerning our behavioral dependent variable For an em-bodied agent whose head moves with curves randomlygenerated by the computer participants show little tono hesitation in approaching In contrast when thehead movements were controlled by a real person (ei-ther a mimic of their own movements or the move-ments from another participant) participants showedgreater hesitation in approaching the embodied agentThus computer-generated movements generated lesshesitancy than human-generated movements

Given this explanation however one might expecteven less hesitation that is fewer reversals in approach-ing the static representation This finding did not occurand one potential explanation is that overall a ran-domly moving representation may elicit less social influ-ence and copresence than a completely static representa-tion In other words the static representation has atleast the potential to exhibit realistic movements whilethe unrealistically moving embodied agent immediatelylowers the sense of copresence In sum participantsrsquohesitancy in approaching an embodied agent was influ-enced by the behavioral realism of the embodied agentbut not its appearance

4 Conclusions

41 Summary of Findings

To summarize the appearance and behavior ofembodied agents had varying effects on the dependentvariables In many instances there was convergence inthat self-report cognitive and behavioral measureschanged in similar ways as a function of our indepen-dent variable manipulations For example the perceived-copresence likability and embarrassment measures cor-related in the predicted directions and the blockheadrepresentation elicited similar reactions across behaviortypes for both the memory measure and the perceived-copresence measure In other instances however therewas a lack of convergence where varying realism pro-duced different outcomes for the various responsemodes One inevitable challenge facing researchers whoexamine a large number of dependent variables is tounderstand the instances in which those dependentvariables do and do not converge

In the current study for some of our dependent mea-sures of copresence (willingness to commit embarrassingacts and likability) manipulation of representation typeaffected copresence responses On other dependentmeasures (self-reported copresence ratings interper-sonal distance approach reversals and memory) ma-nipulation of behavioral realism changed participantsrsquoresponses One way to interpret these results is thatself-report measures were adequate for measuring

Figure 6 Mean number of direction reversals while approaching the

embodied agent as a function of behavioral realism

Bailenson et al 389

copresence levels as affected by variation in appear-ance whereas self-report measures were not sensitiveenough to detect the subtle differences elicited by thebehavioral-realism manipulation Indeed the only reliablemain effects of the behavioral-realism manipulation oc-curred with cognitive and behavioral dependent measures

In other words self-report questionnaires are effectivewhen measuring how people perceive an embodiedagent but not necessarily how they will respond to thatembodied agent Consequently not only does it matterhow one defines copresence it also matters how oneoperationalizes copresence If the goal of a researcher isto understand the way in which people fathom embod-ied agents on a conceptual level then self-report ques-tionnaires may be the best tool However if the goal ofthe researcher is to understand the way in which peoplewill interact with learn from or behave toward anagent then he or she should augment questionnaireswith other types of measures In this paper we are notlobbying for the abandonment of copresence self-reportquestionnaires however we do urge that researchersuse these questionnaires as a complement to other typesof measures whenever possible

While this explanation that self-report measures lacksensitivity receives support from the current data itmust be qualified On most measures there was an in-teraction between behavioral realism and representationtype At times these interactions did not conform toour specific predictions and were difficult to interpretPart of the reason for this difficulty is most likely due tothe fact that we assumed a priori that certain types ofhead movements would be viewed as more realistic thanothers Specifically we assumed that the measures basedon actual human behaviors (mimic and recorded) wouldappear more realistic than either those generated by thecomputer (random) or than no movements at all Re-corded human movements applied to 3D models how-ever may not always be the most effective method ofrendering realistic behaviors and the differences be-tween our various levels of behavioral realism may havebeen too subtle in the current study As previous workhas demonstrated (Garau et al 2003) the interactionbetween appearance and behavioral realism is compli-cated and the finding that seems to be emerging is that

large disparities between the two types of realism cancause diminished levels of copresence Only when theblockhead representation the least humanlike of thethree demonstrated unrealistic random head move-ments did it elicit high levels of self-report and cognitivecopresence

Our findings with respect to the main effects of ap-pearance on self-report measures suggest that how hu-man an embodied agent looks may also be an importantprecondition for experiencing copresence This findingis unanticipated given predictions made by the Blasco-vich model and given findings from previous work(Bailenson et al 2001 Garau et al 2003) It seemsthat virtual representations that look like a realistic hu-man as opposed to either a familiar or unfamiliar non-human elicit embarrassment reactions that would beappropriate only in the presence of a real human

42 Limitations and Future Research

Though this effect surfaces for representations thatare humanlike to varying degrees how this effect of ap-pearance would apply to interactions with multiple hu-man representations of varying photographic realismremains an empirical question One key direction forfuture research is to examine how variation in photo-graphic realism (ie visual fidelity) of humanlike em-bodied agents affects copresence and how it interactswith behavioral realism Furthermore another impor-tant direction is to begin future studies with pretestedlevels of all realism variables (behavioral photographicand others) to assist in a more thorough interpretationof results Moreover the current study is limited in thatwe implemented behavioral realism solely by varyinghead movements A more thorough investigation of thisphenomenon should investigate other types of behav-iors such as facial gestures hand movements and gaitas well as more complex behaviors such as planning andcoordinated actions In addition the interaction task wechose (a purely nonverbal task) is limited and researchshould definitely examine copresence in richer social-interaction settings Finally our method of translationusing the game pad as opposed to naturally trackedwalking was not a perfect behavioral measure since

390 PRESENCE VOLUME 14 NUMBER 4

people do not translate by pushing buttons in the realworld Future work should provide a further examinationof copresence using more natural behavioral measures

Importantly had we measured only self-report re-sponses we would have demonstrated a different patternof results in the current study Attempting to augment anystudy geared toward understanding human interactionwith embodied agents or avatars with dependent measuresfrom multiple response channels certainly can make datainterpretation more complicated Patterns from self-reportdata that seem clean and consistent often are not replicatedwhen a less direct measurement paradigm is implementedFurthermore when asking peripheral implicitly relatedquestions about presence such as likability or embarrass-ment one risks measuring constructs inside the humanmind that are not ldquocentralrdquo to the definition of copres-ence Adding more implicit and disparate measures maynot paint a cleaner picture but it does paint a more com-plete one The challenge to the researcher is interpretingthese complex patterns in the data in order to assess amore thorough understanding of an extremely novel andabstract phenomenon of interacting with a virtual digitalentity

It may be the case that copresence is a purely affectiveconstruct inside the human mind alternatively copres-ence may be more of a cognitive construct In this earlystage of the research field examining copresence in vir-tual environments the theoretical and empirical work isextremely limited and does not allow one to rule outeither of these two (or many other) possibilities Deter-mining the exact conceptual nature of an abstract con-struct in the human mind is a challenge that faces manysocial scientists The measurement framework we offerallows the researcher to get a more complete picture ofthe copresence construct and allows for future researchto utilize more specific types of construct-type hypothe-ses as well as measurement strategies

In addition to the basic research implications with re-spect to the use of various modes of dependent measuresthe current study also has some important applied implica-tions By understanding the components of a virtual repre-sentation that lead to elicitations of heightened copresencewith that representation we can better develop our virtualagents and virtual worlds to immerse and engage the user

As we as well as others have demonstrated creating highlevels of copresence is a complicated process involving anappropriate fit between levels of each type of realism Aswe grow to better understand these interactions we canbetter match agent design to context and perhaps reach alevel where it becomes easy to use agents to develop thedegree of copresence usually found with virtual avatarsand perhaps in the long term the degree of copresencefound with flesh-and-blood people during face-to-faceinteraction

References

Bailenson J N Aharoni E Beall A C Guadagno R Di-mov A amp Blascovich J (2004) Comparing behavioraland self-report measures of embodied agents social presence inimmersive virtual environments Proceedings of the 7th An-nual International Workshop on Presence Valencia Spain

Bailenson J N Beall A C amp Blascovich J (2002) Mutualgaze and task performance in shared virtual environmentsJournal of Visualization and Computer Animation 13 1ndash8

Bailenson J N Beall A C Loomis J Blascovich J ampTurk M (2004) Transformed social interaction Decou-pling representation from behavior and form in collabora-tive virtual environments Presence Teleoperators and Vir-tual Environments 13(4) 428ndash441

Bailenson J N Blascovich J Beall A C amp Loomis J(2001) Equilibrium revisited Mutual gaze and personalspace in virtual environments Presence Teleoperators andVirtual Environments 10 583ndash598

Bailenson J N Blascovich J Beall A C amp Loomis J M(2003) Interpersonal distance in immersive virtual environ-ments Personality and Social Psychology Bulletin 29 1ndash 15

Bailenson J N amp Yee N (in press) Digital chameleonsAutomatic assimilation of nonverbal gestures in immersivevirtual environments Psychological Science

Beall A C Bailenson J N Loomis J Blascovich J ampRex C (2003) Non-zero-sum mutual gaze in collaborativevirtual environments Proceedings of HCI InternationalCrete

Bente G Ruggenberg S Tietz B amp Wortberg S (2004)Measuring behavioral correlates of social presence in virtualencounters Paper presented at the International Communi-cation Association Conference May 27ndash31

Biocca F (1997) Cyborgrsquos dilemma Progressive embodi-

Bailenson et al 391

33 Embarrassment Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender re-vealed a significant main effect of representation type onparticipantsrsquo self-reported willingness to perform embar-rassing acts in front of the embodied agent (ie embar-rassment) F (2 133) 421 p 05 2 06 Spe-cifically participants in the human embodied-agentcondition (M 260 SD 155) were significantlyless willing to perform embarrassing acts in front ofthe ldquootherrdquo than were participants in the teddy-bear(M 351 SD 154) embodied-agent conditionp 05 Participants in the human-representationcondition were also marginally less willing to performembarrassing acts than participants in the blockhead(M 313 SD 156) condition p 10 In sumthe presence of a human-looking representationseemed to reduce participantsrsquo willingness to performin embarrassing acts Regardless of the behavioral re-alism of the representation embodied agents wholooked like humans elicited greater social influencethan did non-human-looking representations

34 Likability Ratings

Results of a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender alsorevealed a significant main effect of representation typeon self-reported likability ratings F (2 133) 543p 01 2 08 Specifically both the human-lookingrepresentation (M 250 SD 109) and the teddy-bear representation (M 228 SD 109) were ratedas significantly more likable than the blockhead repre-sentation (M 179 SD 110) ps 05 Partici-pants found the blockhead representation (which wasthe least human looking of the three representations) tobe less attractive than either of the other two represen-tation types and they liked the blockhead representationthe least

35 Correlations between PerceivedCopresence Embarrassment andLikability Ratings

In order to examine the relationships among per-ceived copresence embarrassment and likability a se-ries of Pearson correlations was computed Consistentwith out predictions results of these analyses revealed asignificant negative correlation between perceived co-presence and a participantrsquos willingness to perform em-barrassing acts in front of the embodied agent r(144) 27 p 001 as well as a significant positivecorrelation between perceived copresence and likabilityr (144) 31 p 001 Specifically we found that thehigher the perceived copresence experienced the lesswilling participants were to perform embarrassing acts infront of the embodied agent (r2 07) and the moreparticipants liked the embodied agent (r2 10) Therewas no significant correlation between likability and em-barrassment ratings

36 Memory

We computed a 3 (representation type human vsteddy bear vs blockhead) 4 (behavioral realism staticvs random movements vs mimic vs recorded) analysisof covariance controlling for participant gender to ex-amine the effects of our independent variables on partic-ipantsrsquo memory Results of these analyses revealed a sig-nificant main effect of behavioral realism on memory F(3 133) 414 p 01 2 09 Specifically mem-ory was significantly worse in the mimic condition thanin either the static or random-movement conditionsps 05 suggesting that overall participants had theworst memory for characters worn by a representationthat mimicked their own movements with a 4-s delayAgain however this main effect was qualified by a sig-nificant representation type by behavioral-realism inter-action (see Figure 5) F (6 133) 270 p 02 2

11Simple main effects tests revealed that within the

blockhead condition memory varied significantly as afunction of behavioral realism F (3 133) 271p 05 2 06 Specifically individuals rememberedsignificantly more characters when they were in the

Bailenson et al 387

random-movement condition than in either the mimicor the recorded conditions (the two behavioral-realismconditions that appeared most humanlike) ps 05and participants in the random-movement conditionremembered marginally more than participants in thestatic condition p 10 These results converge quitewell with the perceived copresence data in which partic-ipants indicated higher copresence ratings for blockheadrepresentations only when the blockhead demonstratedrandom head movements Interestingly this suggeststhat for representations who look the least human theleast human-looking head movements were seen asmost appropriate This finding that avoiding mis-matches is the most effective manner of combining anembodied agentrsquos appearance and behavior is consistentwith previous research on visually embodied agents(Slater amp Steed 2001 Garau et al 2003) as well asvoice agents (Lee amp Nass 2003)

Furthermore the convergence of memory andperceived-copresence ratings provides some evidencethat better memory is indicative of higher copresenceIn other words the more copresence an embodiedagent elicits the more a person interacting with that

embodied agent remembers about it It is interesting tonote however that despite the similar patterns of find-ings between the perceived-copresence data and thememory data results of a Pearson correlation revealedthat there was no significant linear relationship betweenperceived copresence and peoplersquos ability to rememberinformation about an embodied agent

There was however a significant simple main effectof behavioral realism within the human-representationcondition F (3 133) 549 p 001 2 11Specifically participants in the mimic and random-movement conditions remembered significantly fewercharacters than participants in either the static or re-corded conditions ps 05 Participants in the teddy-bear condition displayed the same pattern but in thatcondition none of the differences reached statisticalsignificance

These results suggest that participants may have beendistracted by a human-looking embodied agent thatmimicked their own movements which in turn af-fected their ability to notice and recall information inthe virtual environment Moreover participants mayalso have been distracted in the random-movement con-dition since this condition represented the greatest mis-match between the appearance (human looking) andbehavior (not human looking) of the embodied agentFurther research should examine these potential expla-nations as they suggest that different processes maybe involved for different algorithms for generatingembodied-agent head movements

37 Interpersonal Distance

There were no significant effects of behavioral re-alism or representation type on minimum distance (ourfirst measure of interpersonal distance) Thus partici-pants in all conditions maintained equivalent levels ofinterpersonal distance between themselves and the em-bodied agent There was also no significant correlationbetween perceived copresence and either of our behav-ioral dependent measures (ie minimum distance andnumber of returns)

However as hypothesized results of a 3 (representa-tion type human vs teddy bear vs blockhead) 4 (be-

Figure 5 Gender-covariate adjusted mean number of correctly

recalled characters out of eight on a surprise memory test as a

function of representation type and behavioral realism Conditions that

share the same letter do not differ within that level of representation

type at p 05

388 PRESENCE VOLUME 14 NUMBER 4

havioral realism static vs random movements vs mimicvs recorded) analysis of covariance controlling for par-ticipant gender revealed a significant main effect of be-havioral realism on reversal count or the number oftimes a participant changed direction along the z-axisduring his or her examination of the embodied agentF(3 133) 306 p 05 2 07 As can be seen inFigure 6 participants in the random-movement condi-tion exhibited significantly fewer reversals than partici-pants in any of the other three behavioral-realism condi-tions Pairwise comparisons indicated that there weresignificantly fewer reversals in the random-movementcondition compared to the other three conditions(ps 05) which are not significantly different fromeach other

This finding in part confirms our hypothesis con-cerning our behavioral dependent variable For an em-bodied agent whose head moves with curves randomlygenerated by the computer participants show little tono hesitation in approaching In contrast when thehead movements were controlled by a real person (ei-ther a mimic of their own movements or the move-ments from another participant) participants showedgreater hesitation in approaching the embodied agentThus computer-generated movements generated lesshesitancy than human-generated movements

Given this explanation however one might expecteven less hesitation that is fewer reversals in approach-ing the static representation This finding did not occurand one potential explanation is that overall a ran-domly moving representation may elicit less social influ-ence and copresence than a completely static representa-tion In other words the static representation has atleast the potential to exhibit realistic movements whilethe unrealistically moving embodied agent immediatelylowers the sense of copresence In sum participantsrsquohesitancy in approaching an embodied agent was influ-enced by the behavioral realism of the embodied agentbut not its appearance

4 Conclusions

41 Summary of Findings

To summarize the appearance and behavior ofembodied agents had varying effects on the dependentvariables In many instances there was convergence inthat self-report cognitive and behavioral measureschanged in similar ways as a function of our indepen-dent variable manipulations For example the perceived-copresence likability and embarrassment measures cor-related in the predicted directions and the blockheadrepresentation elicited similar reactions across behaviortypes for both the memory measure and the perceived-copresence measure In other instances however therewas a lack of convergence where varying realism pro-duced different outcomes for the various responsemodes One inevitable challenge facing researchers whoexamine a large number of dependent variables is tounderstand the instances in which those dependentvariables do and do not converge

In the current study for some of our dependent mea-sures of copresence (willingness to commit embarrassingacts and likability) manipulation of representation typeaffected copresence responses On other dependentmeasures (self-reported copresence ratings interper-sonal distance approach reversals and memory) ma-nipulation of behavioral realism changed participantsrsquoresponses One way to interpret these results is thatself-report measures were adequate for measuring

Figure 6 Mean number of direction reversals while approaching the

embodied agent as a function of behavioral realism

Bailenson et al 389

copresence levels as affected by variation in appear-ance whereas self-report measures were not sensitiveenough to detect the subtle differences elicited by thebehavioral-realism manipulation Indeed the only reliablemain effects of the behavioral-realism manipulation oc-curred with cognitive and behavioral dependent measures

In other words self-report questionnaires are effectivewhen measuring how people perceive an embodiedagent but not necessarily how they will respond to thatembodied agent Consequently not only does it matterhow one defines copresence it also matters how oneoperationalizes copresence If the goal of a researcher isto understand the way in which people fathom embod-ied agents on a conceptual level then self-report ques-tionnaires may be the best tool However if the goal ofthe researcher is to understand the way in which peoplewill interact with learn from or behave toward anagent then he or she should augment questionnaireswith other types of measures In this paper we are notlobbying for the abandonment of copresence self-reportquestionnaires however we do urge that researchersuse these questionnaires as a complement to other typesof measures whenever possible

While this explanation that self-report measures lacksensitivity receives support from the current data itmust be qualified On most measures there was an in-teraction between behavioral realism and representationtype At times these interactions did not conform toour specific predictions and were difficult to interpretPart of the reason for this difficulty is most likely due tothe fact that we assumed a priori that certain types ofhead movements would be viewed as more realistic thanothers Specifically we assumed that the measures basedon actual human behaviors (mimic and recorded) wouldappear more realistic than either those generated by thecomputer (random) or than no movements at all Re-corded human movements applied to 3D models how-ever may not always be the most effective method ofrendering realistic behaviors and the differences be-tween our various levels of behavioral realism may havebeen too subtle in the current study As previous workhas demonstrated (Garau et al 2003) the interactionbetween appearance and behavioral realism is compli-cated and the finding that seems to be emerging is that

large disparities between the two types of realism cancause diminished levels of copresence Only when theblockhead representation the least humanlike of thethree demonstrated unrealistic random head move-ments did it elicit high levels of self-report and cognitivecopresence

Our findings with respect to the main effects of ap-pearance on self-report measures suggest that how hu-man an embodied agent looks may also be an importantprecondition for experiencing copresence This findingis unanticipated given predictions made by the Blasco-vich model and given findings from previous work(Bailenson et al 2001 Garau et al 2003) It seemsthat virtual representations that look like a realistic hu-man as opposed to either a familiar or unfamiliar non-human elicit embarrassment reactions that would beappropriate only in the presence of a real human

42 Limitations and Future Research

Though this effect surfaces for representations thatare humanlike to varying degrees how this effect of ap-pearance would apply to interactions with multiple hu-man representations of varying photographic realismremains an empirical question One key direction forfuture research is to examine how variation in photo-graphic realism (ie visual fidelity) of humanlike em-bodied agents affects copresence and how it interactswith behavioral realism Furthermore another impor-tant direction is to begin future studies with pretestedlevels of all realism variables (behavioral photographicand others) to assist in a more thorough interpretationof results Moreover the current study is limited in thatwe implemented behavioral realism solely by varyinghead movements A more thorough investigation of thisphenomenon should investigate other types of behav-iors such as facial gestures hand movements and gaitas well as more complex behaviors such as planning andcoordinated actions In addition the interaction task wechose (a purely nonverbal task) is limited and researchshould definitely examine copresence in richer social-interaction settings Finally our method of translationusing the game pad as opposed to naturally trackedwalking was not a perfect behavioral measure since

390 PRESENCE VOLUME 14 NUMBER 4

people do not translate by pushing buttons in the realworld Future work should provide a further examinationof copresence using more natural behavioral measures

Importantly had we measured only self-report re-sponses we would have demonstrated a different patternof results in the current study Attempting to augment anystudy geared toward understanding human interactionwith embodied agents or avatars with dependent measuresfrom multiple response channels certainly can make datainterpretation more complicated Patterns from self-reportdata that seem clean and consistent often are not replicatedwhen a less direct measurement paradigm is implementedFurthermore when asking peripheral implicitly relatedquestions about presence such as likability or embarrass-ment one risks measuring constructs inside the humanmind that are not ldquocentralrdquo to the definition of copres-ence Adding more implicit and disparate measures maynot paint a cleaner picture but it does paint a more com-plete one The challenge to the researcher is interpretingthese complex patterns in the data in order to assess amore thorough understanding of an extremely novel andabstract phenomenon of interacting with a virtual digitalentity

It may be the case that copresence is a purely affectiveconstruct inside the human mind alternatively copres-ence may be more of a cognitive construct In this earlystage of the research field examining copresence in vir-tual environments the theoretical and empirical work isextremely limited and does not allow one to rule outeither of these two (or many other) possibilities Deter-mining the exact conceptual nature of an abstract con-struct in the human mind is a challenge that faces manysocial scientists The measurement framework we offerallows the researcher to get a more complete picture ofthe copresence construct and allows for future researchto utilize more specific types of construct-type hypothe-ses as well as measurement strategies

In addition to the basic research implications with re-spect to the use of various modes of dependent measuresthe current study also has some important applied implica-tions By understanding the components of a virtual repre-sentation that lead to elicitations of heightened copresencewith that representation we can better develop our virtualagents and virtual worlds to immerse and engage the user

As we as well as others have demonstrated creating highlevels of copresence is a complicated process involving anappropriate fit between levels of each type of realism Aswe grow to better understand these interactions we canbetter match agent design to context and perhaps reach alevel where it becomes easy to use agents to develop thedegree of copresence usually found with virtual avatarsand perhaps in the long term the degree of copresencefound with flesh-and-blood people during face-to-faceinteraction

References

Bailenson J N Aharoni E Beall A C Guadagno R Di-mov A amp Blascovich J (2004) Comparing behavioraland self-report measures of embodied agents social presence inimmersive virtual environments Proceedings of the 7th An-nual International Workshop on Presence Valencia Spain

Bailenson J N Beall A C amp Blascovich J (2002) Mutualgaze and task performance in shared virtual environmentsJournal of Visualization and Computer Animation 13 1ndash8

Bailenson J N Beall A C Loomis J Blascovich J ampTurk M (2004) Transformed social interaction Decou-pling representation from behavior and form in collabora-tive virtual environments Presence Teleoperators and Vir-tual Environments 13(4) 428ndash441

Bailenson J N Blascovich J Beall A C amp Loomis J(2001) Equilibrium revisited Mutual gaze and personalspace in virtual environments Presence Teleoperators andVirtual Environments 10 583ndash598

Bailenson J N Blascovich J Beall A C amp Loomis J M(2003) Interpersonal distance in immersive virtual environ-ments Personality and Social Psychology Bulletin 29 1ndash 15

Bailenson J N amp Yee N (in press) Digital chameleonsAutomatic assimilation of nonverbal gestures in immersivevirtual environments Psychological Science

Beall A C Bailenson J N Loomis J Blascovich J ampRex C (2003) Non-zero-sum mutual gaze in collaborativevirtual environments Proceedings of HCI InternationalCrete

Bente G Ruggenberg S Tietz B amp Wortberg S (2004)Measuring behavioral correlates of social presence in virtualencounters Paper presented at the International Communi-cation Association Conference May 27ndash31

Biocca F (1997) Cyborgrsquos dilemma Progressive embodi-

Bailenson et al 391

random-movement condition than in either the mimicor the recorded conditions (the two behavioral-realismconditions that appeared most humanlike) ps 05and participants in the random-movement conditionremembered marginally more than participants in thestatic condition p 10 These results converge quitewell with the perceived copresence data in which partic-ipants indicated higher copresence ratings for blockheadrepresentations only when the blockhead demonstratedrandom head movements Interestingly this suggeststhat for representations who look the least human theleast human-looking head movements were seen asmost appropriate This finding that avoiding mis-matches is the most effective manner of combining anembodied agentrsquos appearance and behavior is consistentwith previous research on visually embodied agents(Slater amp Steed 2001 Garau et al 2003) as well asvoice agents (Lee amp Nass 2003)

Furthermore the convergence of memory andperceived-copresence ratings provides some evidencethat better memory is indicative of higher copresenceIn other words the more copresence an embodiedagent elicits the more a person interacting with that

embodied agent remembers about it It is interesting tonote however that despite the similar patterns of find-ings between the perceived-copresence data and thememory data results of a Pearson correlation revealedthat there was no significant linear relationship betweenperceived copresence and peoplersquos ability to rememberinformation about an embodied agent

There was however a significant simple main effectof behavioral realism within the human-representationcondition F (3 133) 549 p 001 2 11Specifically participants in the mimic and random-movement conditions remembered significantly fewercharacters than participants in either the static or re-corded conditions ps 05 Participants in the teddy-bear condition displayed the same pattern but in thatcondition none of the differences reached statisticalsignificance

These results suggest that participants may have beendistracted by a human-looking embodied agent thatmimicked their own movements which in turn af-fected their ability to notice and recall information inthe virtual environment Moreover participants mayalso have been distracted in the random-movement con-dition since this condition represented the greatest mis-match between the appearance (human looking) andbehavior (not human looking) of the embodied agentFurther research should examine these potential expla-nations as they suggest that different processes maybe involved for different algorithms for generatingembodied-agent head movements

37 Interpersonal Distance

There were no significant effects of behavioral re-alism or representation type on minimum distance (ourfirst measure of interpersonal distance) Thus partici-pants in all conditions maintained equivalent levels ofinterpersonal distance between themselves and the em-bodied agent There was also no significant correlationbetween perceived copresence and either of our behav-ioral dependent measures (ie minimum distance andnumber of returns)

However as hypothesized results of a 3 (representa-tion type human vs teddy bear vs blockhead) 4 (be-

Figure 5 Gender-covariate adjusted mean number of correctly

recalled characters out of eight on a surprise memory test as a

function of representation type and behavioral realism Conditions that

share the same letter do not differ within that level of representation

type at p 05

388 PRESENCE VOLUME 14 NUMBER 4

havioral realism static vs random movements vs mimicvs recorded) analysis of covariance controlling for par-ticipant gender revealed a significant main effect of be-havioral realism on reversal count or the number oftimes a participant changed direction along the z-axisduring his or her examination of the embodied agentF(3 133) 306 p 05 2 07 As can be seen inFigure 6 participants in the random-movement condi-tion exhibited significantly fewer reversals than partici-pants in any of the other three behavioral-realism condi-tions Pairwise comparisons indicated that there weresignificantly fewer reversals in the random-movementcondition compared to the other three conditions(ps 05) which are not significantly different fromeach other

This finding in part confirms our hypothesis con-cerning our behavioral dependent variable For an em-bodied agent whose head moves with curves randomlygenerated by the computer participants show little tono hesitation in approaching In contrast when thehead movements were controlled by a real person (ei-ther a mimic of their own movements or the move-ments from another participant) participants showedgreater hesitation in approaching the embodied agentThus computer-generated movements generated lesshesitancy than human-generated movements

Given this explanation however one might expecteven less hesitation that is fewer reversals in approach-ing the static representation This finding did not occurand one potential explanation is that overall a ran-domly moving representation may elicit less social influ-ence and copresence than a completely static representa-tion In other words the static representation has atleast the potential to exhibit realistic movements whilethe unrealistically moving embodied agent immediatelylowers the sense of copresence In sum participantsrsquohesitancy in approaching an embodied agent was influ-enced by the behavioral realism of the embodied agentbut not its appearance

4 Conclusions

41 Summary of Findings

To summarize the appearance and behavior ofembodied agents had varying effects on the dependentvariables In many instances there was convergence inthat self-report cognitive and behavioral measureschanged in similar ways as a function of our indepen-dent variable manipulations For example the perceived-copresence likability and embarrassment measures cor-related in the predicted directions and the blockheadrepresentation elicited similar reactions across behaviortypes for both the memory measure and the perceived-copresence measure In other instances however therewas a lack of convergence where varying realism pro-duced different outcomes for the various responsemodes One inevitable challenge facing researchers whoexamine a large number of dependent variables is tounderstand the instances in which those dependentvariables do and do not converge

In the current study for some of our dependent mea-sures of copresence (willingness to commit embarrassingacts and likability) manipulation of representation typeaffected copresence responses On other dependentmeasures (self-reported copresence ratings interper-sonal distance approach reversals and memory) ma-nipulation of behavioral realism changed participantsrsquoresponses One way to interpret these results is thatself-report measures were adequate for measuring

Figure 6 Mean number of direction reversals while approaching the

embodied agent as a function of behavioral realism

Bailenson et al 389

copresence levels as affected by variation in appear-ance whereas self-report measures were not sensitiveenough to detect the subtle differences elicited by thebehavioral-realism manipulation Indeed the only reliablemain effects of the behavioral-realism manipulation oc-curred with cognitive and behavioral dependent measures

In other words self-report questionnaires are effectivewhen measuring how people perceive an embodiedagent but not necessarily how they will respond to thatembodied agent Consequently not only does it matterhow one defines copresence it also matters how oneoperationalizes copresence If the goal of a researcher isto understand the way in which people fathom embod-ied agents on a conceptual level then self-report ques-tionnaires may be the best tool However if the goal ofthe researcher is to understand the way in which peoplewill interact with learn from or behave toward anagent then he or she should augment questionnaireswith other types of measures In this paper we are notlobbying for the abandonment of copresence self-reportquestionnaires however we do urge that researchersuse these questionnaires as a complement to other typesof measures whenever possible

While this explanation that self-report measures lacksensitivity receives support from the current data itmust be qualified On most measures there was an in-teraction between behavioral realism and representationtype At times these interactions did not conform toour specific predictions and were difficult to interpretPart of the reason for this difficulty is most likely due tothe fact that we assumed a priori that certain types ofhead movements would be viewed as more realistic thanothers Specifically we assumed that the measures basedon actual human behaviors (mimic and recorded) wouldappear more realistic than either those generated by thecomputer (random) or than no movements at all Re-corded human movements applied to 3D models how-ever may not always be the most effective method ofrendering realistic behaviors and the differences be-tween our various levels of behavioral realism may havebeen too subtle in the current study As previous workhas demonstrated (Garau et al 2003) the interactionbetween appearance and behavioral realism is compli-cated and the finding that seems to be emerging is that

large disparities between the two types of realism cancause diminished levels of copresence Only when theblockhead representation the least humanlike of thethree demonstrated unrealistic random head move-ments did it elicit high levels of self-report and cognitivecopresence

Our findings with respect to the main effects of ap-pearance on self-report measures suggest that how hu-man an embodied agent looks may also be an importantprecondition for experiencing copresence This findingis unanticipated given predictions made by the Blasco-vich model and given findings from previous work(Bailenson et al 2001 Garau et al 2003) It seemsthat virtual representations that look like a realistic hu-man as opposed to either a familiar or unfamiliar non-human elicit embarrassment reactions that would beappropriate only in the presence of a real human

42 Limitations and Future Research

Though this effect surfaces for representations thatare humanlike to varying degrees how this effect of ap-pearance would apply to interactions with multiple hu-man representations of varying photographic realismremains an empirical question One key direction forfuture research is to examine how variation in photo-graphic realism (ie visual fidelity) of humanlike em-bodied agents affects copresence and how it interactswith behavioral realism Furthermore another impor-tant direction is to begin future studies with pretestedlevels of all realism variables (behavioral photographicand others) to assist in a more thorough interpretationof results Moreover the current study is limited in thatwe implemented behavioral realism solely by varyinghead movements A more thorough investigation of thisphenomenon should investigate other types of behav-iors such as facial gestures hand movements and gaitas well as more complex behaviors such as planning andcoordinated actions In addition the interaction task wechose (a purely nonverbal task) is limited and researchshould definitely examine copresence in richer social-interaction settings Finally our method of translationusing the game pad as opposed to naturally trackedwalking was not a perfect behavioral measure since

390 PRESENCE VOLUME 14 NUMBER 4

people do not translate by pushing buttons in the realworld Future work should provide a further examinationof copresence using more natural behavioral measures

Importantly had we measured only self-report re-sponses we would have demonstrated a different patternof results in the current study Attempting to augment anystudy geared toward understanding human interactionwith embodied agents or avatars with dependent measuresfrom multiple response channels certainly can make datainterpretation more complicated Patterns from self-reportdata that seem clean and consistent often are not replicatedwhen a less direct measurement paradigm is implementedFurthermore when asking peripheral implicitly relatedquestions about presence such as likability or embarrass-ment one risks measuring constructs inside the humanmind that are not ldquocentralrdquo to the definition of copres-ence Adding more implicit and disparate measures maynot paint a cleaner picture but it does paint a more com-plete one The challenge to the researcher is interpretingthese complex patterns in the data in order to assess amore thorough understanding of an extremely novel andabstract phenomenon of interacting with a virtual digitalentity

It may be the case that copresence is a purely affectiveconstruct inside the human mind alternatively copres-ence may be more of a cognitive construct In this earlystage of the research field examining copresence in vir-tual environments the theoretical and empirical work isextremely limited and does not allow one to rule outeither of these two (or many other) possibilities Deter-mining the exact conceptual nature of an abstract con-struct in the human mind is a challenge that faces manysocial scientists The measurement framework we offerallows the researcher to get a more complete picture ofthe copresence construct and allows for future researchto utilize more specific types of construct-type hypothe-ses as well as measurement strategies

In addition to the basic research implications with re-spect to the use of various modes of dependent measuresthe current study also has some important applied implica-tions By understanding the components of a virtual repre-sentation that lead to elicitations of heightened copresencewith that representation we can better develop our virtualagents and virtual worlds to immerse and engage the user

As we as well as others have demonstrated creating highlevels of copresence is a complicated process involving anappropriate fit between levels of each type of realism Aswe grow to better understand these interactions we canbetter match agent design to context and perhaps reach alevel where it becomes easy to use agents to develop thedegree of copresence usually found with virtual avatarsand perhaps in the long term the degree of copresencefound with flesh-and-blood people during face-to-faceinteraction

References

Bailenson J N Aharoni E Beall A C Guadagno R Di-mov A amp Blascovich J (2004) Comparing behavioraland self-report measures of embodied agents social presence inimmersive virtual environments Proceedings of the 7th An-nual International Workshop on Presence Valencia Spain

Bailenson J N Beall A C amp Blascovich J (2002) Mutualgaze and task performance in shared virtual environmentsJournal of Visualization and Computer Animation 13 1ndash8

Bailenson J N Beall A C Loomis J Blascovich J ampTurk M (2004) Transformed social interaction Decou-pling representation from behavior and form in collabora-tive virtual environments Presence Teleoperators and Vir-tual Environments 13(4) 428ndash441

Bailenson J N Blascovich J Beall A C amp Loomis J(2001) Equilibrium revisited Mutual gaze and personalspace in virtual environments Presence Teleoperators andVirtual Environments 10 583ndash598

Bailenson J N Blascovich J Beall A C amp Loomis J M(2003) Interpersonal distance in immersive virtual environ-ments Personality and Social Psychology Bulletin 29 1ndash 15

Bailenson J N amp Yee N (in press) Digital chameleonsAutomatic assimilation of nonverbal gestures in immersivevirtual environments Psychological Science

Beall A C Bailenson J N Loomis J Blascovich J ampRex C (2003) Non-zero-sum mutual gaze in collaborativevirtual environments Proceedings of HCI InternationalCrete

Bente G Ruggenberg S Tietz B amp Wortberg S (2004)Measuring behavioral correlates of social presence in virtualencounters Paper presented at the International Communi-cation Association Conference May 27ndash31

Biocca F (1997) Cyborgrsquos dilemma Progressive embodi-

Bailenson et al 391

havioral realism static vs random movements vs mimicvs recorded) analysis of covariance controlling for par-ticipant gender revealed a significant main effect of be-havioral realism on reversal count or the number oftimes a participant changed direction along the z-axisduring his or her examination of the embodied agentF(3 133) 306 p 05 2 07 As can be seen inFigure 6 participants in the random-movement condi-tion exhibited significantly fewer reversals than partici-pants in any of the other three behavioral-realism condi-tions Pairwise comparisons indicated that there weresignificantly fewer reversals in the random-movementcondition compared to the other three conditions(ps 05) which are not significantly different fromeach other

This finding in part confirms our hypothesis con-cerning our behavioral dependent variable For an em-bodied agent whose head moves with curves randomlygenerated by the computer participants show little tono hesitation in approaching In contrast when thehead movements were controlled by a real person (ei-ther a mimic of their own movements or the move-ments from another participant) participants showedgreater hesitation in approaching the embodied agentThus computer-generated movements generated lesshesitancy than human-generated movements

Given this explanation however one might expecteven less hesitation that is fewer reversals in approach-ing the static representation This finding did not occurand one potential explanation is that overall a ran-domly moving representation may elicit less social influ-ence and copresence than a completely static representa-tion In other words the static representation has atleast the potential to exhibit realistic movements whilethe unrealistically moving embodied agent immediatelylowers the sense of copresence In sum participantsrsquohesitancy in approaching an embodied agent was influ-enced by the behavioral realism of the embodied agentbut not its appearance

4 Conclusions

41 Summary of Findings

To summarize the appearance and behavior ofembodied agents had varying effects on the dependentvariables In many instances there was convergence inthat self-report cognitive and behavioral measureschanged in similar ways as a function of our indepen-dent variable manipulations For example the perceived-copresence likability and embarrassment measures cor-related in the predicted directions and the blockheadrepresentation elicited similar reactions across behaviortypes for both the memory measure and the perceived-copresence measure In other instances however therewas a lack of convergence where varying realism pro-duced different outcomes for the various responsemodes One inevitable challenge facing researchers whoexamine a large number of dependent variables is tounderstand the instances in which those dependentvariables do and do not converge

In the current study for some of our dependent mea-sures of copresence (willingness to commit embarrassingacts and likability) manipulation of representation typeaffected copresence responses On other dependentmeasures (self-reported copresence ratings interper-sonal distance approach reversals and memory) ma-nipulation of behavioral realism changed participantsrsquoresponses One way to interpret these results is thatself-report measures were adequate for measuring

Figure 6 Mean number of direction reversals while approaching the

embodied agent as a function of behavioral realism

Bailenson et al 389

copresence levels as affected by variation in appear-ance whereas self-report measures were not sensitiveenough to detect the subtle differences elicited by thebehavioral-realism manipulation Indeed the only reliablemain effects of the behavioral-realism manipulation oc-curred with cognitive and behavioral dependent measures

In other words self-report questionnaires are effectivewhen measuring how people perceive an embodiedagent but not necessarily how they will respond to thatembodied agent Consequently not only does it matterhow one defines copresence it also matters how oneoperationalizes copresence If the goal of a researcher isto understand the way in which people fathom embod-ied agents on a conceptual level then self-report ques-tionnaires may be the best tool However if the goal ofthe researcher is to understand the way in which peoplewill interact with learn from or behave toward anagent then he or she should augment questionnaireswith other types of measures In this paper we are notlobbying for the abandonment of copresence self-reportquestionnaires however we do urge that researchersuse these questionnaires as a complement to other typesof measures whenever possible

While this explanation that self-report measures lacksensitivity receives support from the current data itmust be qualified On most measures there was an in-teraction between behavioral realism and representationtype At times these interactions did not conform toour specific predictions and were difficult to interpretPart of the reason for this difficulty is most likely due tothe fact that we assumed a priori that certain types ofhead movements would be viewed as more realistic thanothers Specifically we assumed that the measures basedon actual human behaviors (mimic and recorded) wouldappear more realistic than either those generated by thecomputer (random) or than no movements at all Re-corded human movements applied to 3D models how-ever may not always be the most effective method ofrendering realistic behaviors and the differences be-tween our various levels of behavioral realism may havebeen too subtle in the current study As previous workhas demonstrated (Garau et al 2003) the interactionbetween appearance and behavioral realism is compli-cated and the finding that seems to be emerging is that

large disparities between the two types of realism cancause diminished levels of copresence Only when theblockhead representation the least humanlike of thethree demonstrated unrealistic random head move-ments did it elicit high levels of self-report and cognitivecopresence

Our findings with respect to the main effects of ap-pearance on self-report measures suggest that how hu-man an embodied agent looks may also be an importantprecondition for experiencing copresence This findingis unanticipated given predictions made by the Blasco-vich model and given findings from previous work(Bailenson et al 2001 Garau et al 2003) It seemsthat virtual representations that look like a realistic hu-man as opposed to either a familiar or unfamiliar non-human elicit embarrassment reactions that would beappropriate only in the presence of a real human

42 Limitations and Future Research

Though this effect surfaces for representations thatare humanlike to varying degrees how this effect of ap-pearance would apply to interactions with multiple hu-man representations of varying photographic realismremains an empirical question One key direction forfuture research is to examine how variation in photo-graphic realism (ie visual fidelity) of humanlike em-bodied agents affects copresence and how it interactswith behavioral realism Furthermore another impor-tant direction is to begin future studies with pretestedlevels of all realism variables (behavioral photographicand others) to assist in a more thorough interpretationof results Moreover the current study is limited in thatwe implemented behavioral realism solely by varyinghead movements A more thorough investigation of thisphenomenon should investigate other types of behav-iors such as facial gestures hand movements and gaitas well as more complex behaviors such as planning andcoordinated actions In addition the interaction task wechose (a purely nonverbal task) is limited and researchshould definitely examine copresence in richer social-interaction settings Finally our method of translationusing the game pad as opposed to naturally trackedwalking was not a perfect behavioral measure since

390 PRESENCE VOLUME 14 NUMBER 4

people do not translate by pushing buttons in the realworld Future work should provide a further examinationof copresence using more natural behavioral measures

Importantly had we measured only self-report re-sponses we would have demonstrated a different patternof results in the current study Attempting to augment anystudy geared toward understanding human interactionwith embodied agents or avatars with dependent measuresfrom multiple response channels certainly can make datainterpretation more complicated Patterns from self-reportdata that seem clean and consistent often are not replicatedwhen a less direct measurement paradigm is implementedFurthermore when asking peripheral implicitly relatedquestions about presence such as likability or embarrass-ment one risks measuring constructs inside the humanmind that are not ldquocentralrdquo to the definition of copres-ence Adding more implicit and disparate measures maynot paint a cleaner picture but it does paint a more com-plete one The challenge to the researcher is interpretingthese complex patterns in the data in order to assess amore thorough understanding of an extremely novel andabstract phenomenon of interacting with a virtual digitalentity

It may be the case that copresence is a purely affectiveconstruct inside the human mind alternatively copres-ence may be more of a cognitive construct In this earlystage of the research field examining copresence in vir-tual environments the theoretical and empirical work isextremely limited and does not allow one to rule outeither of these two (or many other) possibilities Deter-mining the exact conceptual nature of an abstract con-struct in the human mind is a challenge that faces manysocial scientists The measurement framework we offerallows the researcher to get a more complete picture ofthe copresence construct and allows for future researchto utilize more specific types of construct-type hypothe-ses as well as measurement strategies

In addition to the basic research implications with re-spect to the use of various modes of dependent measuresthe current study also has some important applied implica-tions By understanding the components of a virtual repre-sentation that lead to elicitations of heightened copresencewith that representation we can better develop our virtualagents and virtual worlds to immerse and engage the user

As we as well as others have demonstrated creating highlevels of copresence is a complicated process involving anappropriate fit between levels of each type of realism Aswe grow to better understand these interactions we canbetter match agent design to context and perhaps reach alevel where it becomes easy to use agents to develop thedegree of copresence usually found with virtual avatarsand perhaps in the long term the degree of copresencefound with flesh-and-blood people during face-to-faceinteraction

References

Bailenson J N Aharoni E Beall A C Guadagno R Di-mov A amp Blascovich J (2004) Comparing behavioraland self-report measures of embodied agents social presence inimmersive virtual environments Proceedings of the 7th An-nual International Workshop on Presence Valencia Spain

Bailenson J N Beall A C amp Blascovich J (2002) Mutualgaze and task performance in shared virtual environmentsJournal of Visualization and Computer Animation 13 1ndash8

Bailenson J N Beall A C Loomis J Blascovich J ampTurk M (2004) Transformed social interaction Decou-pling representation from behavior and form in collabora-tive virtual environments Presence Teleoperators and Vir-tual Environments 13(4) 428ndash441

Bailenson J N Blascovich J Beall A C amp Loomis J(2001) Equilibrium revisited Mutual gaze and personalspace in virtual environments Presence Teleoperators andVirtual Environments 10 583ndash598

Bailenson J N Blascovich J Beall A C amp Loomis J M(2003) Interpersonal distance in immersive virtual environ-ments Personality and Social Psychology Bulletin 29 1ndash 15

Bailenson J N amp Yee N (in press) Digital chameleonsAutomatic assimilation of nonverbal gestures in immersivevirtual environments Psychological Science

Beall A C Bailenson J N Loomis J Blascovich J ampRex C (2003) Non-zero-sum mutual gaze in collaborativevirtual environments Proceedings of HCI InternationalCrete

Bente G Ruggenberg S Tietz B amp Wortberg S (2004)Measuring behavioral correlates of social presence in virtualencounters Paper presented at the International Communi-cation Association Conference May 27ndash31

Biocca F (1997) Cyborgrsquos dilemma Progressive embodi-

Bailenson et al 391

copresence levels as affected by variation in appear-ance whereas self-report measures were not sensitiveenough to detect the subtle differences elicited by thebehavioral-realism manipulation Indeed the only reliablemain effects of the behavioral-realism manipulation oc-curred with cognitive and behavioral dependent measures

In other words self-report questionnaires are effectivewhen measuring how people perceive an embodiedagent but not necessarily how they will respond to thatembodied agent Consequently not only does it matterhow one defines copresence it also matters how oneoperationalizes copresence If the goal of a researcher isto understand the way in which people fathom embod-ied agents on a conceptual level then self-report ques-tionnaires may be the best tool However if the goal ofthe researcher is to understand the way in which peoplewill interact with learn from or behave toward anagent then he or she should augment questionnaireswith other types of measures In this paper we are notlobbying for the abandonment of copresence self-reportquestionnaires however we do urge that researchersuse these questionnaires as a complement to other typesof measures whenever possible

While this explanation that self-report measures lacksensitivity receives support from the current data itmust be qualified On most measures there was an in-teraction between behavioral realism and representationtype At times these interactions did not conform toour specific predictions and were difficult to interpretPart of the reason for this difficulty is most likely due tothe fact that we assumed a priori that certain types ofhead movements would be viewed as more realistic thanothers Specifically we assumed that the measures basedon actual human behaviors (mimic and recorded) wouldappear more realistic than either those generated by thecomputer (random) or than no movements at all Re-corded human movements applied to 3D models how-ever may not always be the most effective method ofrendering realistic behaviors and the differences be-tween our various levels of behavioral realism may havebeen too subtle in the current study As previous workhas demonstrated (Garau et al 2003) the interactionbetween appearance and behavioral realism is compli-cated and the finding that seems to be emerging is that

large disparities between the two types of realism cancause diminished levels of copresence Only when theblockhead representation the least humanlike of thethree demonstrated unrealistic random head move-ments did it elicit high levels of self-report and cognitivecopresence

Our findings with respect to the main effects of ap-pearance on self-report measures suggest that how hu-man an embodied agent looks may also be an importantprecondition for experiencing copresence This findingis unanticipated given predictions made by the Blasco-vich model and given findings from previous work(Bailenson et al 2001 Garau et al 2003) It seemsthat virtual representations that look like a realistic hu-man as opposed to either a familiar or unfamiliar non-human elicit embarrassment reactions that would beappropriate only in the presence of a real human

42 Limitations and Future Research

Though this effect surfaces for representations thatare humanlike to varying degrees how this effect of ap-pearance would apply to interactions with multiple hu-man representations of varying photographic realismremains an empirical question One key direction forfuture research is to examine how variation in photo-graphic realism (ie visual fidelity) of humanlike em-bodied agents affects copresence and how it interactswith behavioral realism Furthermore another impor-tant direction is to begin future studies with pretestedlevels of all realism variables (behavioral photographicand others) to assist in a more thorough interpretationof results Moreover the current study is limited in thatwe implemented behavioral realism solely by varyinghead movements A more thorough investigation of thisphenomenon should investigate other types of behav-iors such as facial gestures hand movements and gaitas well as more complex behaviors such as planning andcoordinated actions In addition the interaction task wechose (a purely nonverbal task) is limited and researchshould definitely examine copresence in richer social-interaction settings Finally our method of translationusing the game pad as opposed to naturally trackedwalking was not a perfect behavioral measure since

390 PRESENCE VOLUME 14 NUMBER 4

people do not translate by pushing buttons in the realworld Future work should provide a further examinationof copresence using more natural behavioral measures

Importantly had we measured only self-report re-sponses we would have demonstrated a different patternof results in the current study Attempting to augment anystudy geared toward understanding human interactionwith embodied agents or avatars with dependent measuresfrom multiple response channels certainly can make datainterpretation more complicated Patterns from self-reportdata that seem clean and consistent often are not replicatedwhen a less direct measurement paradigm is implementedFurthermore when asking peripheral implicitly relatedquestions about presence such as likability or embarrass-ment one risks measuring constructs inside the humanmind that are not ldquocentralrdquo to the definition of copres-ence Adding more implicit and disparate measures maynot paint a cleaner picture but it does paint a more com-plete one The challenge to the researcher is interpretingthese complex patterns in the data in order to assess amore thorough understanding of an extremely novel andabstract phenomenon of interacting with a virtual digitalentity

It may be the case that copresence is a purely affectiveconstruct inside the human mind alternatively copres-ence may be more of a cognitive construct In this earlystage of the research field examining copresence in vir-tual environments the theoretical and empirical work isextremely limited and does not allow one to rule outeither of these two (or many other) possibilities Deter-mining the exact conceptual nature of an abstract con-struct in the human mind is a challenge that faces manysocial scientists The measurement framework we offerallows the researcher to get a more complete picture ofthe copresence construct and allows for future researchto utilize more specific types of construct-type hypothe-ses as well as measurement strategies

In addition to the basic research implications with re-spect to the use of various modes of dependent measuresthe current study also has some important applied implica-tions By understanding the components of a virtual repre-sentation that lead to elicitations of heightened copresencewith that representation we can better develop our virtualagents and virtual worlds to immerse and engage the user

As we as well as others have demonstrated creating highlevels of copresence is a complicated process involving anappropriate fit between levels of each type of realism Aswe grow to better understand these interactions we canbetter match agent design to context and perhaps reach alevel where it becomes easy to use agents to develop thedegree of copresence usually found with virtual avatarsand perhaps in the long term the degree of copresencefound with flesh-and-blood people during face-to-faceinteraction

References

Bailenson J N Aharoni E Beall A C Guadagno R Di-mov A amp Blascovich J (2004) Comparing behavioraland self-report measures of embodied agents social presence inimmersive virtual environments Proceedings of the 7th An-nual International Workshop on Presence Valencia Spain

Bailenson J N Beall A C amp Blascovich J (2002) Mutualgaze and task performance in shared virtual environmentsJournal of Visualization and Computer Animation 13 1ndash8

Bailenson J N Beall A C Loomis J Blascovich J ampTurk M (2004) Transformed social interaction Decou-pling representation from behavior and form in collabora-tive virtual environments Presence Teleoperators and Vir-tual Environments 13(4) 428ndash441

Bailenson J N Blascovich J Beall A C amp Loomis J(2001) Equilibrium revisited Mutual gaze and personalspace in virtual environments Presence Teleoperators andVirtual Environments 10 583ndash598

Bailenson J N Blascovich J Beall A C amp Loomis J M(2003) Interpersonal distance in immersive virtual environ-ments Personality and Social Psychology Bulletin 29 1ndash 15

Bailenson J N amp Yee N (in press) Digital chameleonsAutomatic assimilation of nonverbal gestures in immersivevirtual environments Psychological Science

Beall A C Bailenson J N Loomis J Blascovich J ampRex C (2003) Non-zero-sum mutual gaze in collaborativevirtual environments Proceedings of HCI InternationalCrete

Bente G Ruggenberg S Tietz B amp Wortberg S (2004)Measuring behavioral correlates of social presence in virtualencounters Paper presented at the International Communi-cation Association Conference May 27ndash31

Biocca F (1997) Cyborgrsquos dilemma Progressive embodi-

Bailenson et al 391

people do not translate by pushing buttons in the realworld Future work should provide a further examinationof copresence using more natural behavioral measures

Importantly had we measured only self-report re-sponses we would have demonstrated a different patternof results in the current study Attempting to augment anystudy geared toward understanding human interactionwith embodied agents or avatars with dependent measuresfrom multiple response channels certainly can make datainterpretation more complicated Patterns from self-reportdata that seem clean and consistent often are not replicatedwhen a less direct measurement paradigm is implementedFurthermore when asking peripheral implicitly relatedquestions about presence such as likability or embarrass-ment one risks measuring constructs inside the humanmind that are not ldquocentralrdquo to the definition of copres-ence Adding more implicit and disparate measures maynot paint a cleaner picture but it does paint a more com-plete one The challenge to the researcher is interpretingthese complex patterns in the data in order to assess amore thorough understanding of an extremely novel andabstract phenomenon of interacting with a virtual digitalentity

It may be the case that copresence is a purely affectiveconstruct inside the human mind alternatively copres-ence may be more of a cognitive construct In this earlystage of the research field examining copresence in vir-tual environments the theoretical and empirical work isextremely limited and does not allow one to rule outeither of these two (or many other) possibilities Deter-mining the exact conceptual nature of an abstract con-struct in the human mind is a challenge that faces manysocial scientists The measurement framework we offerallows the researcher to get a more complete picture ofthe copresence construct and allows for future researchto utilize more specific types of construct-type hypothe-ses as well as measurement strategies

In addition to the basic research implications with re-spect to the use of various modes of dependent measuresthe current study also has some important applied implica-tions By understanding the components of a virtual repre-sentation that lead to elicitations of heightened copresencewith that representation we can better develop our virtualagents and virtual worlds to immerse and engage the user

As we as well as others have demonstrated creating highlevels of copresence is a complicated process involving anappropriate fit between levels of each type of realism Aswe grow to better understand these interactions we canbetter match agent design to context and perhaps reach alevel where it becomes easy to use agents to develop thedegree of copresence usually found with virtual avatarsand perhaps in the long term the degree of copresencefound with flesh-and-blood people during face-to-faceinteraction

References

Bailenson J N Aharoni E Beall A C Guadagno R Di-mov A amp Blascovich J (2004) Comparing behavioraland self-report measures of embodied agents social presence inimmersive virtual environments Proceedings of the 7th An-nual International Workshop on Presence Valencia Spain

Bailenson J N Beall A C amp Blascovich J (2002) Mutualgaze and task performance in shared virtual environmentsJournal of Visualization and Computer Animation 13 1ndash8

Bailenson J N Beall A C Loomis J Blascovich J ampTurk M (2004) Transformed social interaction Decou-pling representation from behavior and form in collabora-tive virtual environments Presence Teleoperators and Vir-tual Environments 13(4) 428ndash441

Bailenson J N Blascovich J Beall A C amp Loomis J(2001) Equilibrium revisited Mutual gaze and personalspace in virtual environments Presence Teleoperators andVirtual Environments 10 583ndash598

Bailenson J N Blascovich J Beall A C amp Loomis J M(2003) Interpersonal distance in immersive virtual environ-ments Personality and Social Psychology Bulletin 29 1ndash 15

Bailenson J N amp Yee N (in press) Digital chameleonsAutomatic assimilation of nonverbal gestures in immersivevirtual environments Psychological Science

Beall A C Bailenson J N Loomis J Blascovich J ampRex C (2003) Non-zero-sum mutual gaze in collaborativevirtual environments Proceedings of HCI InternationalCrete

Bente G Ruggenberg S Tietz B amp Wortberg S (2004)Measuring behavioral correlates of social presence in virtualencounters Paper presented at the International Communi-cation Association Conference May 27ndash31

Biocca F (1997) Cyborgrsquos dilemma Progressive embodi-

Bailenson et al 391

ment in virtual environments Journal of Computer-Mediated Communication 3(2) Available from httpwwwascuscorgjcmcvol3issue2biocca2html

Blascovich J (2002) Social influences within immersive vir-tual environments In R Schroeder (Ed) The social life ofavatars London Springer-Verlag

Blascovich J Loomis J Beall A Swinth K Hoyt C ampBailenson J N (2002) Immersive virtual environmenttechnology as a methodological tool for social psychologyPsychological Inquiry 13 103ndash124

Garau M (2003) The impact of avatar fidelity on social inter-action in virtual environments Doctoral thesis UniversityCollege London London England

Garau M Slater M Bee S amp Sasse M A (2001) Theimpact of eye gaze on communication using humanoid ava-tars Proceedings of the SIG-CHI Conference on Human Fac-tors in Computing Systems Seattle WA

Garau M Slater M Vinayagamoorthy V Brogni ASteed A Sasse M A (2003) The impact of avatar realismand eye gaze control on perceived quality of communica-tion in a shared immersive virtual environment Proceedingsof the SIG-CHI Conference on Human Factors in ComputingSystems Fort Lauderdale FL

Heeter C (1992) Being there The subjective experience ofpresence Presence Teleoperators and Virtual Environments1(2) 262ndash271

Hoyt C L Blascovich J amp Swinth K R (2003) Socialinhibition in immersive virtual environments Presence Tele-operators and Virtual Environments 12(2) 183ndash195

Krikorian D H Lee J Chock T M amp Harms C (2000)Isnrsquot that spatial Distance and communication in a 2-Dvirtual environment Journal of Computer-Mediated Com-munication 5(4) Available from httpwwwascuscorgjcmcvol5issue4krikorianhtml

Lee K M amp Nass C (2003) Designing social presence ofsocial actors in human computer interaction Proceedings ofthe CHI 2003 Conference on Human Factors in ComputingSystems 289ndash296

Lee K M amp Nass C (2004) The multiple source effect andsynthesized speech Doubly disembodied language as a con-ceptual framework Human Communication Research 30182ndash207

Lessiter J Freeman J Keogh E amp Davidoff J D (2001)A cross-media presence questionnaire The ITC sense ofpresence inventory Presence Teleoperators and Virtual En-vironments 10(3) 282ndash297

Lombard M amp Ditton T (1997) At the heart of it all The

concept of presence Journal of Computer-Mediated Com-munication 3(2) Available from httpwwwascuscorgjcmcvol3issue2

Meehan M Insko B Whitton M Brooks F P (2002)Physiological measures of presence in stressful virtual envi-ronments ACM Transactions on Graphics 21(3) 645ndash653

New J J (2003) A content-specific attenuation of changeblindness Preferential attention to animate beings in naturalscenes Journal of Vision 3(9) 643 Available from httpjournalofvisionorg39643

Nowak K (2001 May) The influence of anthropomorphism onsocial judgment in social virtual environments Paper pre-sented at the International Communication AssociationWashington DC

Nowak K amp Biocca F (2001 May) The influence of agencyand the virtual body on presence social presence and copresencein computer mediated interaction Paper presented at the 4thInternational Workshop on Presence Philadelphia PA

Parise S Kiesler S B Sproull S amp Waters K (1996) Mypartner is a real dog Cooperation with social agents Proceed-ings of Computer Supported Cooperative Work 399ndash 408

Reeves B amp Nass C (1996) The media equation How peo-ple treat computers television and new media like real peopleand places New York Cambridge University Press

Russo T C (2001 May) Operationalizing mediated pres-ence Initial steps toward a measure of the construct Paperpresented at the 4th International Workshop on PresencePhiladelphia PA

Schroeder R (2002) Social interaction in virtual environ-ments Key issues common themes and a framework forresearch In R Schroeder (Ed) The social life of avatarsPresence and interaction in shared virtual environments (pp1ndash18) London Springer-Verlag

Slater M (2004) How colorful was your day Why question-naires cannot assess presence in virtual environments PresenceTeleoperators and Virtual Environments 13(4) 484ndash493

Slater M Sadagic A Usoh M Schroeder R (2000)Small group behavior in a virtual and real environment Acomparative study Presence Teleoperators and Virtual Envi-ronments 9 37ndash51

Slater M amp Steed A (2001) Meeting people virtually Experi-ments in virtual environments In R Schroeder (Ed) The so-cial life of avatars Presence and interaction in shared virtualenvironments (pp 146ndash171) London Springer-Verlag

Swinth K R amp Blascovich J (2002) Perceiving and re-sponding to others Human-human and human-computersocial interaction in collaborative virtual environments Pro-

392 PRESENCE VOLUME 14 NUMBER 4

ceedings of the 5th Annual International Workshop on PRES-ENCE Porto Portugal

Appendix Questionnaires

Note (r) denotes reverse-coded itemsCopresence ( 71)

1 Even when the ldquootherrdquo was present I still feltalone in the virtual room (r)

2 I felt like there was someone else in the room with me3 I felt like the ldquootherrdquo was aware of my presence in

the room

Embarrassment ( 72)

1 I would be willing to change clothes in front ofthe ldquootherrdquo

2 I would be willing to pick my nose in front of theldquootherrdquo

3 I would be willing to act out a scene from themovie ldquoTitanicrdquo in front of the ldquootherrdquo

Likability ( 71)

1 I like the ldquootherrdquo2 I would like to meet this ldquootherrdquo again3 The ldquootherrdquo is attractive4 Spending time with the ldquootherrdquo was NOT satisfy-

ing (r)

Bailenson et al 393