The promise and peril of interactive embodied agents for studying non-verbal communication: a machine learning perspective.

In face-to-face interactions, parties rapidly react and adapt to each other's words, movements and expressions. Any science of face-to-face interaction must develop approaches to hypothesize and rigorously test mechanisms that explain such interdependent behaviour. Yet conventional experimental designs often sacrifice interactivity to establish experimental control. Interactive virtual and robotic agents have been offered as a way to study true interactivity while enforcing a measure of experimental control by allowing participants to interact with realistic but carefully controlled partners. But as researchers increasingly turn to machine learning to add realism to such agents, they may unintentionally distort the very interactivity they seek to illuminate, particularly when investigating the role of non-verbal signals such as emotion or active-listening behaviours. Here I discuss some of the methodological challenges that may arise when machine learning is used to model the behaviour of interaction partners. By articulating and explicitly considering these commitments, researchers can transform 'unintentional distortions' into valuable methodological tools that yield new insights and better contextualize existing experimental findings that rely on learning technology. This article is part of a discussion meeting issue 'Face2face: advancing the science of social interaction'.