AI Agents in Human Social Decision-Making: A Framework from Social Inference to Social Action

This is the authors' abstract. We don't add key points for this paper.

Artificial intelligence (AI) agents are no longer encountered only as tools for search, prediction, or automation. AI agents increasingly generate information, participate in interactions, and influence how people perceive and respond to their social environment. This shift poses a challenge for cognitive science. Human social cognition evolved to infer others’ mental states from socially generated cues and to predict the actions of goal-directed agents to guide decision-making in social contexts. AI agents, however, now produce synthetic social cues that often resemble human-generated signals while arising from fundamentally different generative processes, potentially violating the inferential and predictive assumptions embedded in human social cognition. Emerging evidence suggests that this mismatch alters how people perceive, interpret, and act upon social information. This review organizes the literature around two established routes linking social signals to social decision-making. The retrospective inferential route examines how AI-generated social cues may reshape the mapping from observed signals to the underlying mental states and incentives that generated them, influencing source attribution, cue diagnosticity, social comparison, and self-evaluation. The prospective predictive route examines how AI agents may reshape expectations about outcomes, norms, authorship, and accountability, thereby influencing trust, punishment, delegation, coordination, and disclosure. Together, these analytically distinct but dynamically interconnected routes provide an integrative organizing framework for understanding how AI changes both the interpretation of social evidence and the anticipation of social consequences, while highlighting individual-, system-, task-, and culture-level boundary conditions.