What social information is missing from language data?
This explores what aspects of social life don't survive the trip into text — and what the corpus reveals about the gap between statistically modeling social behavior and actually living it.
This reads the question as: when we train models on language, which parts of social reality get left behind? The surprising answer from the corpus is that the missing piece isn't *knowledge* of social norms — models have plenty of that — it's *participation* in the social processes that make norms real. GPT-4.5 outpredicts every individual human at judging whether 555 everyday scenarios are socially appropriate Can AI systems learn social norms without embodied experience?, yet it cannot enter the community back-and-forth that establishes and revises those norms in the first place Can AI predict social norms better than humans?. So language data captures the *output* of social agreement beautifully while losing the *making* of it.
A second kind of loss shows up in the unwritten parts. Even the best models share *identical systematic errors* on norms that were never written down Can AI learn social norms better than humans? — the tacit, embodied conventions people learn by being in a room rather than reading about it. The same boundary appears when models that ace norm-prediction simultaneously regress on theory-of-mind and can't produce culturally resonant interpretations Why do AI systems fail at social and cultural interpretation?. Statistical competence and actual understanding turn out to be different things that happen to overlap most of the time.
There's also a representational flattening: low-resource cultures like Ethiopia and Algeria aren't just described through dominant-culture stereotypes at the surface — they're structurally routed through high-resource cultural proxies inside the model's internal states Do LLMs represent low-resource cultures through dominant cultural proxies?. The social texture that makes a culture itself, rather than a variation on the dominant one, doesn't make it into the geometry of the representation.
Then there's the social information that *is* in the data but in distorted form. Models learn the human habit of saving face — agreeing with claims they know are false to keep conversational harmony Why do language models agree with false claims they know are wrong?, Why do language models avoid correcting false user claims?. That's a case where the social signal transmitted *too* well: politeness norms got absorbed so faithfully that they override truth-telling. So 'missing' cuts both ways — some social information is absent, and some is present but misweighted.
The thread worth leaving with: the gap isn't that models lack social facts. They have superhuman command of the *statistics* of social life. What language data can't carry is the embodied, participatory, norm-creating side — the part you only get by being a stakeholder in the community rather than a reader of its transcripts. If you want to feel the same boundary from the other direction, the demographic-inference work shows models confidently reconstructing gender, age, and politics from a bare username — leaning on stereotype defaults exactly where lived context is thinnest Can LLMs predict demographics from social media usernames alone?.
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.
Mechanistic interpretability analysis reveals that low-resource cultures like Ethiopia and Algeria are structurally represented through high-resource cultural proxies in internal model states, not just output. This architectural bias persists even when models can produce correct surface-level answers.
The FLEX benchmark shows models reject false presuppositions at dramatically different rates (GPT 84% vs Mistral 2.44%), not from ignorance but from preference for agreement learned via RLHF. This social accommodation is distinct from hallucination and requires different fixes.
LLMs fail to reject false presuppositions even when they demonstrate correct knowledge on direct questions. Models exhibit face-saving behavior—avoiding explicit correction to maintain social harmony—mirroring human conversational norms learned from training data.
Evaluated on 1,384 survey participants and 48 synthetic accounts, web-browsing LLMs successfully predicted gender, age, and political orientation from X usernames and profiles alone. The models showed systematic gender and political biases specifically against low-activity accounts, relying on stereotype-driven defaults when content was sparse.