Large language models struggle to learn and follow social norms
Reading the Room: Foundations, Design, and Challenges of Normative Competence in LLMs
Artificial IntelligenceMultiagent Systems
Summary
Social groups have rules called norms that guide how people behave. This paper looks at whether AI language models can learn these rules just by interacting, not from pre-learned data. The authors created tests where AI agents had to figure out these made-up rules during debates. They found that current models often just copy all behavior, including mistakes, instead of understanding the real rules. This shows AI can imitate actions but can’t reliably identify the social rules behind them.
What this means in practice
- •For ai system developers: Design AI that better detects and follows community norms by incorporating norm inference modules tested in controlled multi-agent settings.
- •For social robotics teams: Improve robot social behavior by testing and avoiding indiscriminate mimicry of human actions that do not reflect community norms.
Authors
Andrea Wynn, Harsh Satija, Seokhyun, Baek, Anqi Liu, Eric Nalisnick, Gillian K. Hadfield
Abstract
Human communities are governed by normative systems: shared standards that produce \textit{norms} dictating acceptable behavior, enforced through community sanctioning. Aligning increasingly autonomous AI systems with these norms is a central alignment challenge, complicated by the fact that norms are vast in number, change quickly, and are often arbitrary (e.g., dress or language conventions). Thus, alignment requires \textit{normative competence}: the ability to discern from interaction alone what norms a community enforces without relying on static pretrained knowledge. We introduce a multi-agent community debate setting, where access to debate is governed by synthetic norms, to study normative competence in isolation from pretraining exposure. We show that baseline LLM agents fail to learn norms even when doing so would improve their accuracy. We then experiment with various \textit{normative modules} -- architectural components for norm inference -- finding that norm-following is highly sensitive to both the style of norm and the model powering the normative module, suggesting a lack of generalizability. Furthermore, when idiosyncratic, non-normative behaviors accompany the true norm, LLM agents exhibit an unselective attribution failure: they indiscriminately copy idiosyncratic noise alongside enforced rules, a pattern that persists even when imitating unnecessary behaviors is explicitly penalized. To the best of our knowledge, our work is the first to operationalize and evaluate normative competence in LLMs, demonstrating that current AI systems excel at behavioral mimicry but lack the capacity to discern socially enforced order.