COMPUTER SCIENCE:
Virtually Trustworthy

Our impression of others in face-to-face interactions derives from multi-modal observations: We hear voices, see faces and gestures, and listen to each other's words. Our online interactions are less sensory, dominated instead by text: We send e-mail, read discussion boards, and participate in text chats. Yet that may be changing. Online 3D spaces such as There and Second Life feature detailed virtual environments in which users represent themselves with graphical avatars, often highly customizable.

Avatars may closely resemble a person, or they may be fantastical creations such as aliens or talking lizards. Their online use dates back to the mid-1980s, when they were used in games such as Ultima and in the online social space Habitat. As technology has advanced, they have gone from being static 2D images to complex 3D animations, complete with realistic gait, fashionable clothing, and dynamic facial expressions.

Held in trust. Self presentation in real and virtual life. CREDIT: DIETMAR OFFENHUBER

We carry out social interactions with a large number of communicative behaviors that indicate our intention, state of mind, communicative competency, and so on. For instance, you may see an acquaintance across the room at a cocktail party and decide to go speak to him. You carry out this goal not only by walking across the room but also by making eye contact, smiling, raising your brows, adjusting your clothes--a complex set of communicative behaviors that indicate your intention to start a conversation, allow you to gauge his willingness to do so, and show your level of determination. Cassell and Vilhjálmsson (1) argued that avatars without these social behaviors seem stilted and awkward. They can be moved to stand next to each other to talk, but stare blankly into space, inert and unengaged.

Just giving the user finer control over the avatar is not a satisfactory option, however, for if users must specify their avatar's every eye movement and gesture, they would be too distracted to engage in the conversation itself. Cassell and Vilhjálmsson's solution was to program such behaviors into the avatars, to be set off when the user indicated some desired action. For instance, if the user indicated (with a simple typed command) that she wanted to end a conversation, the avatar would break away by averting its gaze, and upon leaving would look at its recent partner, nod its head, and wave. Users of this system found it more natural and engaging than one with static avatars and felt that their conversational partners were more expressive.

Researchers have found that users infer a number of character traits from avatar behavior and appearance. They judge avatars that are humanlike and clearly gendered (as opposed to androgynous) to be the most attractive and the most credible (2). In an audio-only conversation, simply adding an avatar whose head and eye movements match the conversation flow increases users' perception of their partners' trustworthiness and friendliness.

Today's online graphical interactions are still rather awkward. Behavioral sophistication lags behind rendering skill, so we have avatars whose appearance raises high expectations of humanlike behaviors but whose gaze and gestures are relatively primitive. However, it is quite conceivable that in a few years avatars whose behavior is nearly imperceptible from humans' will be available.

Yet this raises important questions about the reliability of the impressions we form in avatar-mediated interactions. In our face-to-face interactions, many of the cues we read to assess traits such as trustworthiness have real links to the trait. Gaze direction, for example, links directly to what one is seeing. When long averted, it is thus a sign of inattention. During times of high cognitive load, such as when inventing a story, people may make less eye contact, possibly because gazing at another face is itself a cognitively intensive process. Perhaps because of this, we have the popular (though unsubstantiated) belief that someone who makes steady and direct eye contact is being honest (3).

Online, however, behaviors generated by a software program can create the same impression of trustworthiness or friendliness, but without a grounding connection to an underlying cognitive process or other causative element. As behavioral software becomes more sophisticated, are we creating avatars that will be increasingly attractive and seemingly friendly but are in fact the ideal mask behind which a dishonest or manipulative person can operate? Once an interface includes humanlike avatars, the issue of user interpretation of character traits from ungrounded avatar behaviors is inevitable, for even nonaction is an interpretable behavior; it conveys an impression of social ineptness and distance.

How we assess these issues depends on the context in which they are used and how we view them in relation to real-world practice. In the real world, we use many strategies to enhance the impression we make on others. We employ resume consultants and speech coaches, wear makeup, and undergo plastic surgery. In the virtual realm, idealized bodies and perfect skin are the norm (2), but there are also a whole new range of possible enhancements. Bailenson and colleagues have conducted several experiments on digital mimicry, such as morphing a person's own face into the avatar of their conversational partner or having that avatar closely mimic their gestures (4). They found that people in a group paid closer attention to messages delivered by a "team face" avatar, made by combining the features of the people in the group. This could be a useful technique for enhancing the cohesiveness of geographically dispersed groups. Yet they also found that politicians' arguments were more persuasive when their faces were made to subtly resemble the listener's, raising the specter of a world in which you are bombarded with oddly compelling ad campaigns presented by people just like you.

Reliability involves tradeoffs. Less reliable communication is often cheaper or easier; when deception becomes too prevalent, more costly signals or social sanctions may be needed (5). Suspicious citizens of the future may demand to interact with candidates only through trusted "manipulation-free" sites. Today, most 3D graphical sites are fantasy games, where role-playing and artifice are not only accepted but also required. As social sites such as Second Life gain popularity, other uses are emerging, including academic lectures, retail stores, and business meetings. These will require a range of avatar designs, not only in terms of technical sophistication but also across a continuum from the most attractive and impressively persuasive to the most rigorously and reliably grounded.

References

1. J. Cassell, H. Vilhjálmsson, Autonomous Agents and Multi-Agent Systems 2, 45 (1999).
2. K. L. Nowak, C. Rauh, J. Computer-Mediated Comm. 11, 153 (2005).
3. B. M. DePaulo et al., Psychol. Bull. 129, 74 (2003).
4. J. N. Bailenson, A. C. Beall, in Avatars at Work and Play: Collaboration and Interaction in Shared Virtual Environments, R. Schroeder, A. Axelsson, Eds. (Springer-Verlag, London, 2006), pp. 1-16.
5. J. Donath, Signals, Truth, and Design (MIT Press, Cambridge, MA, in press).

10.1126/science.1142770