Abstract
<p>The success of human-AI teams depends on humans' willingness to cooperate with artificial intelligence (AI), yet little is known about whether the social group membership people associate with AI shapes this cooperation. Across three preregistered experiments using a two-player public goods game, we tested whether ingroup favoritism, operationalized through political affiliation, influences how people train and cooperate with AI. When training AI, participants taught it to be more cooperative when it would interact with members of their own political party than with members of the opposing party (Experiment 1). When playing alongside trained AIs, participants were more cooperative when paired with AI trained by their political ingroup (Experiment 2), even when they could observe that both AIs were equally cooperative (Experiment 3). These findings reveal that people shape and evaluate AI based on social group membership, extending ingroup favoritism to artificial agents that have no group membership of their own.</p>