Simulated emotional reactions affect (dis)honesty in speech-free human–agent interactions

Can simulated emotional reactions displayed by embodied virtual agents influence human moral behaviour? Specifically, can they affect honesty, a core moral norm, even if the reaction is neither sentient nor genuine? In a preregistered and incentivised experiment, 372 participants played a repeated Prisoner’s Dilemma with a cooperative agent whose simulated facial reactions were designed to promote cooperation, defection, or neither. We then measured honesty in two unrelated tasks without emotional feedback, where dishonesty had either only self-regarding consequences (Mind Game) or also other-regarding consequences (Deception Game). When dishonesty affected the agent’s outcomes, honesty remained relatively high given the stakes, as if the agent’s outcomes were treated as morally relevant. In contrast, when dishonesty had no consequences for the agent, honesty depended on prior exposure: participants were more honest after interacting with a defection-promoting agent than with a cooperation-promoting one, consistent with moral compensation. These results suggest that even brief, speech-free interactions with emotionally expressive yet non-sentient agents can influence moral behaviour, suggesting that people respond to such agents as they would to real social partners. This raises important questions about the ethical impact of simulated emotion in human-AI interactions.