The Chameleon Machine: Mimicry, Entrainment and Robots That Meet You Halfway

People unconsciously copy each other's postures, words and voices, and a growing body of research asks what happens when a machine joins in.

We are all copycats

Watch two friends talking in a café and you will often see a quiet choreography. One leans back, the other follows. One rubs their chin, and a minute later so does the other. Nobody planned it. In 1999 Tanya Chartrand and John Bargh gave this a name, the chameleon effect, in a paper in the Journal of Personality and Social Psychology.

Across three experiments they showed that people tended to copy the mannerisms of a stranger they worked with, such as face rubbing or foot shaking, without being aware of it. When a confederate deliberately mirrored participants' postures and movements, the interaction felt smoother and participants liked the confederate more. People who scored higher on a measure of empathy showed more mimicry. The proposed mechanism was a direct link between perceiving a behaviour and the tendency to perform it.

Later research has refined and sometimes complicated these findings, and the social psychology of mimicry has had its share of debate over how robust specific effects are. But the broad idea that people adapt their behaviour to their partners has strong support, especially in language.

Talking alike

Conversation is full of convergence. In 1996 Susan Brennan and Herbert Clark studied how people refer to objects over repeated exchanges. Partners quickly settled on shared labels, what they called conceptual pacts, and kept using them even when a simpler term would do. If you and I have agreed to call something 'the pinstriped shoe', I will keep saying it, but I may switch if I start talking to someone new. Word choice is negotiated, and it is tied to particular partners.

Martin Pickering and Simon Garrod pushed this into a general theory of dialogue in a 2004 article in Behavioral and Brain Sciences. They argued that alignment, the tendency of conversational partners to converge on the same words, grammatical structures and situation models, is not a side effect of conversation but central to how it works. Aligning at one level makes alignment at others easier, and it reduces the effort each person must spend modelling the other's mind.

The same holds for the sound of speech. Researchers in spoken dialogue, notably Julia Hirschberg's group at Columbia University, have measured how speakers come to resemble each other in features such as loudness, pitch and speaking rate over the course of a conversation. This acoustic-prosodic entrainment has been linked in various studies to perceptions of engagement and rapport, though the size and direction of the effect vary between speakers and settings.

People align with machines too

Do people do this with computers? Holly Branigan, Pickering and colleagues reviewed the evidence in a 2010 paper in the Journal of Pragmatics and argued that they do, in some respects more strongly than with humans. Users repeat the words and structures a system uses. Part of this is automatic, but part appears strategic: in related experiments from the same group, people aligned more with partners they believed to be computers, and more with computers presented as less capable, as if trying to stay within the system's vocabulary to make sure they are understood. The reasons may differ from human conversation, but the behaviour looks similar.

The reverse direction is what interests robot designers. If people like partners who mirror them, perhaps machines should mirror people. Jeremy Bailenson and Nick Yee tested this in immersive virtual reality in 2005. Participants listened to a virtual agent deliver a persuasive message. For half of them, the agent copied their head movements with a delay of a few seconds. Agents that mimicked were rated as more persuasive and more likeable, and few participants consciously spotted the copying.

Robots that adapt their voices

Social robots offer a natural test bed because they can adapt speech in real time. In a 2016 study at the ACM/IEEE Human-Robot Interaction conference, Nichola Lubold, Erin Walker and Heather Pon-Barry had students teach maths problems to a small teachable robot called Quinn. In one version the robot added social chat to the task. In another it also adapted its voice, shifting the average pitch of its synthetic speech to match the student's while keeping its intonation pattern. A control version did neither. Social presence was rated higher when social dialogue was combined with vocal adaptation than with social dialogue alone.

Findings like these suggest that prosodic entrainment, long studied as something humans do, can be a design feature. A robot that subtly meets its user's speaking style may feel more attentive than one that speaks identically to everyone.

Mirrors with limits

There are reasons for caution. Mimicry works in part because it goes unnoticed. Copying that is too fast, too exact or too obvious risks feeling mocking or manipulative rather than warm. Alignment also has ethical edges: a machine designed to win liking and persuasion through unconscious channels is exerting influence that users may not detect. And convergence is not always desirable. A tutoring robot should not drift towards a student's mistakes, and a robot used by many people should not absorb the quirks of whoever spoke to it last.

The more defensible goal is the one human conversation already models. Good conversational partners adapt to each other to understand and be understood. They settle on shared words, match pace and tone, and keep track of what they have agreed. A robot that does those things is not tricking anyone. It is simply learning to talk the way people talk to each other.

References

1. Chartrand, T. L., & Bargh, J. A. (1999). The chameleon effect: The perception-behavior link and social interaction. Journal of Personality and Social Psychology, 76(6), 893-910. Link
2. Brennan, S. E., & Clark, H. H. (1996). Conceptual pacts and lexical choice in conversation. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22(6), 1482-1493. Link
3. Pickering, M. J., & Garrod, S. (2004). Toward a mechanistic psychology of dialogue. Behavioral and Brain Sciences, 27(2), 169-190. Link
4. Branigan, H. P., Pickering, M. J., Pearson, J., & McLean, J. F. (2010). Linguistic alignment between people and computers. Journal of Pragmatics, 42(9), 2355-2368. Link
5. Bailenson, J. N., & Yee, N. (2005). Digital chameleons: Automatic assimilation of nonverbal gestures in immersive virtual environments. Psychological Science, 16(10), 814-819. Link
6. Lubold, N., Walker, E., & Pon-Barry, H. (2016). Effects of voice-adaptation and social dialogue on perceptions of a robotic learning companion. Proceedings of the 11th ACM/IEEE International Conference on Human-Robot Interaction, 255-262. Link

Feature drafted with AI from the published research listed above, for social-robotics.com.