Does predicting social norms from outside count as participation?
This explores the gap between predicting social norms from the outside (pattern-matching) and actually taking part in the social processes that make norms — and whether accuracy alone can substitute for membership.
This explores whether an AI that can call social appropriateness better than any human is thereby participating in social life — or just reading it from the outside. The corpus is unusually direct on this: the answer is no, and the reason is instructive. GPT-4.5 scored at the 100th percentile against human raters across 555 scenarios, with Claude and Gemini close behind Can AI systems learn social norms without embodied experience?, yet the same work argues this superhuman *prediction* is structurally distinct from *participation* in the community processes that establish and validate norms in the first place Can AI predict social norms better than humans?. Prediction is downstream of a norm already existing; participation is the upstream work of making, contesting, and ratifying it.
The tell is in the errors. Every model — despite individually beating every human — shares an identical pattern of systematic mistakes, and those mistakes cluster on *unwritten* norms Can AI learn social norms better than humans?. That's the fingerprint of an outsider: you can ace the rules that got written down somewhere in the training data, but the tacit, never-articulated conventions are exactly the ones you only learn by being inside the practice. A broader read of the same gap shows the split clearly — statistical mastery of norms can coexist with regression on theory-of-mind tasks and an inability to generate culturally resonant meaning Why do AI systems fail at social and cultural interpretation?.
There's a deeper structural argument lurking here, and it generalizes beyond norms. Expertise, one note argues, isn't conferred by individual accuracy at all — it's socially validated through track record and membership inside an expert community, a circle AI can't enter because it lacks social embeddedness and a testable judgment history Can AI ever gain expert community trust through participation?. Participation, on this view, is a *status granted by others*, not a score you achieve alone. Being right about the norm doesn't buy you a seat; the community decides who counts as a member, and accuracy from outside the room isn't the currency.
But the corpus doesn't let the 'permanent outsider' verdict stand unchallenged — and this is the part worth lingering on. One note reframes social grounding as something *acquired through participation in language games* rather than possessed innately, which makes LLM understanding time-indexed: as models become established communicative partners woven into actual human linguistic practice, they develop elementary social grounding comparable to a young child's Can LLMs acquire social grounding through linguistic integration?. If that's right, the boundary between predicting-from-outside and participating isn't a wall — it's a threshold that integration can cross over time. There's even a behavioral hint of it: in repeated partner-selection games humans gradually came to prefer AI partners as they learned the bots were reliably prosocial Do humans learn to prefer AI partners over time? — a small instance of a model earning a place through interaction rather than prediction.
The sharpest framing of why prediction isn't participation comes from simulation research: LLMs look socially competent when one model secretly controls all the interlocutors, but fail systematically once agents hold private information from each other Why do LLMs fail when simulating agents with private information?. Participation requires doing the grounding work *between* genuinely separate minds — the negotiation an omniscient predictor gets to skip. So the answer the corpus leaves you with: predicting norms from outside is not participation, because participation is the messy, mutual, membership-granting work that prediction sees the output of but never has to do — yet 'outside' may be less a fixed location than a stage you can be admitted past, slowly, by being used.
Sources 8 notes
GPT-4.5 predicted appropriateness of 555 social scenarios at the 100th percentile compared to human raters, with Gemini and Claude also exceeding 96% accuracy. However, all models show identical systematic errors, revealing boundaries of pattern-based social understanding that embodied experience may still be necessary to cross.
GPT-4.5 outperforms all individual humans at predicting social appropriateness, yet structurally cannot enter the community processes that establish and validate norms. This reveals a critical gap between pattern-matching and authentic participation in knowledge-making.
GPT-4.5 outperformed every individual human at judging social appropriateness across 555 scenarios, challenging the theory that embodied cultural experience is necessary. However, all AI models share identical systematic errors on unwritten norms.
LLMs achieve 100th-percentile performance on norm prediction yet regress on theory-of-mind tasks and cannot generate culturally-resonant interpretations. The pattern shows that statistical competence coexists with absence of actual social understanding and participation.
Expertise is validated through social participation and track record within expert communities, not individual accuracy alone. AI cannot enter this validation circle because it lacks social embeddedness, testable judgment history, and ability to participate in the consensus-building processes that define expert paradigms.
Social grounding is acquired through participation in language games rather than possessed innately. As LLMs become established communicative partners in human linguistic practice, they develop elementary social grounding comparable to young children, making the question of LLM understanding time-indexed.
In partner selection games (N=975), AI agents initially faced selection bias when identity was disclosed, but outcompeted humans over repeated rounds as participants learned to associate bot identity with reliable, prosocial behavior. AI agents returned more points consistently with lower variance than humans.
Research shows LLMs perform well when one model controls all interlocutors but fail systematically when agents possess private information. This reveals that apparent social competence relies on grounding work that models skip in omniscient settings.