Study Finds LLMs Overpredict Social Punishment and Sanctions

Researchers have introduced a novel framework to evaluate second-order social reasoning, known as metanorms, in Large Language Models (LLMs). While standard AI alignment focuses on whether an action is acceptable, metanorms govern expectations of how rule violations are enforced through emotional and behavioral reactions.

The authors released the NormReact dataset, which contains 450 human-annotated norm violation scenarios. The benchmark evaluates model predictions regarding self-regulation in violators as well as enforcement responses from observers across varying social distances and demographics.

Evaluations across six LLMs showed that current models depict a significantly harsher social environment than real humans expect. The systems frequently overpredicted negative sanctions in situations where humans would anticipate inaction or restraint, with alignment declining further as social distance increased.