A new study examines the ability of Large Language Models to understand and predict second-order social norms, often referred to as metanorms. Previous alignment efforts have focused on direct social rules, but social intelligence requires anticipating enforcement mechanisms and expected reactions. The research introduces a framework for evaluating LLMs across emotional appraisal and behavioral response dimensions. Specifically, tasks were created to predict self-regulation in violators and other-regulation in observers.
The study released a dataset, NormReact, containing 450 norm violation scenarios. The data was annotated for emotions and behavioral responses, considering the gender of the violator and the social closeness of the observer. Current LLMs demonstrate a tendency to predict harsher social consequences than humans would expect. Across six models, predictions of negative sanctions were overemphasized when human judgment would have anticipated inaction.
Furthermore, alignment with human judgments weakens as the social distance between the model and the observer increases. This indicates a potential distortion in AI systems’ understanding of social regulation, potentially over-representing punishment and under-representing tolerance and restraint. This has implications for AI systems in norm-sensitive domains.
This research highlights the importance of incorporating metanorm reasoning into LLM development. It suggests a need for improved evaluation metrics and datasets to accurately represent real-world social dynamics. Source: https://arxiv.org/abs/2609.05437v1