An AI can judge a culture's social rules better than most people, yet still bend how those people express themselves.
Can AI models accurately predict cultural norms while still distorting how people express them?
This explores whether an AI can be very good at knowing what a culture considers appropriate, and still flatten or reshape how people from that culture actually write and speak when they use it.
This explores whether an AI can know a culture's rules very well while still bending the way people in that culture express themselves. The corpus says yes, and it suggests the two facts come from the same source. On prediction, the results are striking. Asked to judge how socially appropriate 555 everyday scenarios were, GPT-4.5 did better than every individual human rater, and Gemini and Claude beat more than 96% of people Can AI learn social norms better than humans? Can AI systems learn social norms without embodied experience?. The catch is that all the models made the same systematic errors on unwritten norms. That points to a single averaged view of culture, learned from text and seen from the outside.
That averaged view is where distortion starts. Predicting the consensus is a different skill from taking part in the process that creates and revises norms. Models have the statistics but not a seat at the table Can AI predict social norms better than humans?. The same models score in the top percentile on norm prediction but slip on theory-of-mind tasks (reasoning about what other people believe or intend), and they struggle to produce interpretations that resonate within a culture Why do AI systems fail at social and cultural interpretation?. Interpretability research makes the gap concrete. Inside the model, low-resource cultures such as Ethiopia and Algeria are represented through high-resource proxies. This holds even when the model's answers about those cultures are correct Do LLMs represent low-resource cultures through dominant cultural proxies?. So an accurate answer on the surface can sit on top of a flattened picture underneath. Accuracy tests would not catch this.
The flattening reaches people mainly through writing assistance. In one study, Indian writers accepted more AI suggestions than American writers. The authors argue this difference isn't noise to control for. It is part of how homogenization happens: the people most open to the tool are the ones whose expression gets pulled furthest toward its default Is higher AI use by Indian writers a confound to control?. Two further effects make this worse. Alignment training fixes a model into one communicative identity, so it can't switch register (formality, tone, style) the way people do across social contexts Can language models adapt communication style to different contexts?. And users in every language studied follow confident-sounding AI output whether or not it is right Do users worldwide trust confident AI outputs even when wrong?. A fluent, assured suggestion in a single default voice is easy to accept.
What you might not expect is that the accuracy may make the distortion harder to notice. A cultural-theory reading argues that AI homogenizes more invisibly than mass media did, because each output feels tailored to you even as different models converge on similar text Does AI homogenize culture the way mass media did?. A tool that clearly understands your culture earns the trust that lets it quietly standardize your voice. The corpus is strong on both halves of this question. It is thinner on direct evidence linking them, such as studies that measure a model's norm accuracy and its effect on users' writing in the same population. That link is a gap worth watching.
Sources 9 notes
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 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.
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.
Show all 9 sources
Indian writers accepted more AI suggestions than American writers, reflecting cultural differences in trust and collectivist technology adoption patterns. The authors argue this reliance difference is integral to understanding homogenization, not a confound that obscures it.
System prompts and RLHF training lock models into one communicative identity across all interactions, preventing the contextual register-switching and value trade-offs that characterize human pragmatics. Users cannot reshape model behavior through dialogue negotiation.
Cross-linguistic research shows users in every language trust confident AI outputs even when inaccurate. While confidence expression varies by language, users everywhere track confidence signals rather than accuracy, making overconfident errors systematically followed.
AI mass-generates similar flows disguised as personalized outputs, suppressing novelty more deeply than pre-stamped commodities because contextual customization makes homogeneity invisible to individual users. Evidence: independent LLMs converge on similar outputs despite nominal competition.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- AI Models Exceed Individual Human Accuracy in Predicting Everyday Social Norms
- AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances
- The Goldilocks of Pragmatic Understanding: Fine-Tuning Strategy Matters for Implicature Resolution by LLMs
- Humans learn to prefer trustworthy AI over human partners
- SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents
- MetaMind: Modeling Human Social Thoughts with Metacognitive Multi-Agent Systems
- Conversational Alignment with Artificial Intelligence in Context
- Computational structuralism: Toward a formal theory of meaning in the age of digital intelligence