Long-term impact of LLM-based social decision-making on human cooperation

Ascertain the long-term effects of using large language models to support social decision-making on human cooperation dynamics, particularly in settings where reputation, observability of interactions, and indirect reciprocity guide cooperative behavior.

Background

The paper motivates this problem by noting that people increasingly rely on LLMs for judgments in social settings, and prior work indicates LLMs can shape moral and political judgments. Because large-scale cooperation often depends on indirect reciprocity—where reputations and observers’ judgments influence future helping—understanding how LLM-based advice affects these mechanisms is crucial.

To probe this question, the authors extract social norms from 21 LLMs using a large prompt dataset that encodes donation-game interactions with reputational context, and then evaluate the inferred norms in an evolutionary game-theoretic model. The stated open issue is the broader, long-term societal implication: whether and how these LLM-shaped judgments alter cooperation trajectories over time.