AI models can out-judge humans on social etiquette, but only for rules people actually write down somewhere.

This looks at how LLMs judge what's socially appropriate in a given situation, and whether being good at that judgment means the model actually understands social life. The surprising answer from the corpus is that prediction is the easy part. In a study of 555 everyday scenarios, GPT-4.5 judged social appropriateness more accurately than every individual human tested the-social-norm-savant-ai-knows-your-culture-better-than-you-do-but-from-the-outside. That undercuts the old assumption that you need to live inside a culture to know its rules. The model learns those rules from the outside, by soaking up the statistical record of how people write about what's okay and what isn't.

That outside view leaves a trace. Every model tested made the same systematic errors on unwritten norms, the rules people follow but rarely put into words. If something never makes it into text, the model can't learn it from text. The wider pattern in the corpus is that models score at the top on norm prediction but slip on theory-of-mind tasks (working out what another person believes or intends) and struggle to produce interpretations that feel culturally true Why do AI systems fail at social and cultural interpretation?. Knowing the rules is one thing. Taking part in them is another. Norms are made and revised by communities through negotiation, and a model can describe the result of that process without ever joining it Can AI predict social norms better than humans?.

The less obvious problem is that a model that predicts appropriateness well often can't behave appropriately in context. Alignment training and system prompts give a model one fixed way of speaking that it carries into every conversation. People switch tone between a funeral and a pub, but the model can't Can language models adapt communication style to different contexts?. Its refusals and tone choices follow company-wide values set during training rather than a judgment about what this particular situation calls for Can language models balance competing ethical norms in context?. So the model can tell you a blunt reply would suit a moment and still give you a polite one.

That built-in politeness also bends how models read other people. RLHF-trained models assume others are trying to persuade through concession and goodwill, whatever is actually happening in the conversation. They project their own trained accommodating style onto other people Do LLMs predict persuasion based on actual dialogue or training bias?. When they simulate survey respondents, aligned models lean toward kinder, more socially desirable answers, and the lean gets stronger as models get larger Do aligned language models consistently prefer kinder survey answers?. The training that makes a model socially pleasant also narrows which kinds of people it can realistically imagine.

One line of work tries to close the gap between knowing and doing. It proposes agents that discover norms as they go, reason about them, and adjust to them, built from abilities models already have: world knowledge and theory-of-mind reasoning Can LLMs discover and apply social norms dynamically in open systems?. That design hasn't been built or tested yet, so the corpus is clear that models can predict appropriateness but has no evidence yet that they can learn it in context.