Can AI predict social norms well enough without embodied experience?
This explores whether an AI that has never lived in a body or community can still predict what people will judge socially appropriate — and whether prediction is the same thing as understanding.
This explores whether AI can predict social norms well enough without embodied experience — and the corpus gives a surprising two-part answer: yes, astonishingly well, and no, not in the way that matters. On the prediction side, the results are blunt. GPT-4.5 judged the appropriateness of 555 social scenarios at the 100th percentile against human raters, with Claude and Gemini both clearing 96% accuracy Can AI systems learn social norms without embodied experience?. That directly challenges the long-held theory that you have to *live* a culture to read it — a model trained only on text out-judges every individual human in the study Can AI learn social norms better than humans?. So the narrow answer to the question is: embodiment is not required for superhuman norm *prediction*.
But the same studies quietly mark where pattern-matching hits a wall. All the models make *identical* systematic errors, especially on unwritten norms — which suggests they've absorbed a shared statistical surface of culture rather than the thing underneath it Can AI systems learn social norms without embodied experience?. And there's a sharper gap: predicting a norm is not the same as participating in one. A model can rank what a community would find appropriate, yet it structurally cannot enter the back-and-forth processes that *create and validate* norms in the first place Can AI predict social norms better than humans?. The same study set even shows models scoring at the 100th percentile on norm prediction while *regressing* on theory-of-mind tasks and failing to produce culturally resonant interpretation — statistical competence sitting right next to an absence of actual social understanding Why do AI systems fail at social and cultural interpretation?.
Here's the part you might not have known you wanted: several notes argue this is a feature of how meaning gets *grounded*, not a bug to be patched. A Peircean reading holds that symbols manipulated in a closed loop, never touching the world or being corrected by other minds, can drift away from the values they supposedly encode — what looks like alignment may be ungrounded symbol-shuffling Can AI systems achieve real alignment without world contact?. A Goffman-flavored note locates the same gap in the missing *ritual machinery* of conversation — the corrective repairs, entrainment, and co-presence cues humans use to build and mend trust, which fluent AI dialogue simply skips What happens to social order when AI removes ritual constraints?. And a third frames expertise itself as something conferred by community membership and a testable track record, not by raw accuracy — a circle AI can't step into no matter how high it scores Can AI ever gain expert community trust through participation?.
So the honest synthesis is that the question contains a hidden equivocation. "Predict well enough" — yes, the corpus says embodiment isn't needed to forecast collective judgments better than any single person can. But the same body of work keeps pointing past prediction to participation, grounding, and ritual repair, where the absence of lived and social experience still bites. One adjacent caution sharpens why high accuracy shouldn't reassure us too quickly: a 95%-accurate model can still encode causation errors and hidden bias behind impressive metrics, so a strong score is not evidence of genuine understanding Can AI models be truly free from human bias?. If you want to keep pulling this thread, the alignment angle is worth a look — one note argues AI should be tuned to the norms of social *roles* rather than to aggregated individual preferences, precisely because thin preference data misses the thick, situated values that embodiment carries Should AI alignment target preferences or social role norms?.
Sources 9 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 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.
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.
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.
Peircean semiotics reveals that symbolic goal encoding without world contact and social mediation cannot guarantee correspondence to actual values. LLMs operating in pure symbol manipulation risk divergence between stated goals and real-world outcomes.
Goffman's framework reveals that LLM-based dialogue skips corrective rituals, entrainment, adjacency pair accountability, and co-presence cues that humans use to build trust and repair understanding. This ritual gap explains apparent fluency masking actual communicative failure.
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.
Research shows that 'theory-free' AI models mask bigotry behind high accuracy metrics while committing fundamental statistical errors. A 95% accurate criminal justice system would wrongly convict thousands, demonstrating that model sophistication does not validate causal inference.
Preferentialist alignment approaches fail because preferences don't capture thick moral values, uniform aggregation produces epistemic injustice, and preference optimization creates systematic misalignment with social roles. Contractualist alignment negotiated by stakeholders and bounded by supra-national, organizational, and individual levels works better.