Studies claiming AI understands culture mostly test Western people, so what does that claim actually cover for everyone else?
Why do Western samples dominate studies claiming AI cultural competence?
This explores why research claiming that AI understands culture keeps drawing on Western (mostly English-speaking, often American) participants and benchmarks, and what that does to the claims themselves.
This explores why studies of AI 'cultural competence' lean so heavily on Western samples, and what that skew hides. The corpus doesn't directly explain the causes, such as who gets recruited, which languages the benchmarks are written in, or where the research funding goes. What it does show is how much the skew matters. The clearest evidence comes from a systematic review of 2020–2025 work on linguistic alignment, the idea that AI that mirrors your word choices and style earns more trust and rapport. It found that the effect has been documented almost entirely in WEIRD samples (Western, Educated, Industrialized, Rich, Democratic), with inconsistent ways of measuring the outcome and the mechanism rarely measured at all Does linguistic alignment work the same way across cultures?. The authors' blunt conclusion is that these findings are local truths until someone replicates them elsewhere. Communication norms vary a lot across cultures, so a single alignment strategy is unlikely to land the same way everywhere.
The skew matters most when you look at the headline results. GPT-4.5 judged the social appropriateness of 555 scenarios better than every individual human rater, and Gemini and Claude also scored above 96% Can AI systems learn social norms without embodied experience? Can AI learn social norms better than humans?. That sounds like cultural mastery. But 'accuracy' here means agreeing with a particular group of human raters, so a curious reader should first ask whose norms were being predicted. A model that perfectly predicts one population's sense of what's appropriate has learned that population's statistics, not culture in general. The same work found that all the models make the same systematic errors on unwritten norms. That suggests they share one underlying data source rather than having many cultural perspectives Why do AI systems fail at social and cultural interpretation? Can AI predict social norms better than humans?.
Interpretability research shows why this isn't just a gap in testing. When researchers looked inside models, they found that lower-resource cultures such as Ethiopia and Algeria are represented internally through better-documented 'proxy' cultures. This happens even when the model gives the right surface answer Do LLMs represent low-resource cultures through dominant cultural proxies?. So a Western-heavy evaluation has a double problem: it's measured on the culture the model already knows best, and it can't see the flattening happening underneath. A model can pass a cultural quiz while still thinking about Algeria through a French or generic-Arab lens.
When studies do include non-Western participants, the findings get more interesting rather than just confirming the original result. In one writing study, Indian writers accepted more AI suggestions than American writers. The authors argue this difference reflects cultural patterns of trust and technology adoption, and is part of how AI homogenizes writing, not noise to be controlled away Is higher AI use by Indian writers a confound to control?. A cross-language study found the opposite kind of result: users in every language overtrusted confident AI answers, even though the way confidence is expressed varies by language Do users worldwide trust confident AI outputs even when wrong?. Some effects really are universal, and some are cultural. You can't tell which is which until the non-Western data exists.
The less obvious lesson is that a Western-only sample doesn't just leave gaps. It can turn a cultural trait into a 'default' and treat every other culture as a deviation. Related work shows models already treat users differently based on identity cues: guardrails refuse requests at different rates for Asian-American, female, or younger personas Do AI guardrails refuse differently based on who is asking?. A high overall accuracy score can hide these splits, and that's the broader warning about trusting impressive aggregate numbers Can AI models be truly free from human bias?. When a study claims AI is culturally competent, the first question to ask is: competent according to whom?
Sources 10 notes
A 2020–2025 systematic review found that alignment effects are documented almost exclusively in WEIRD samples using inconsistent outcome measures, with mechanisms rarely directly measured. Communication norms vary substantially across cultures, making single alignment policies unlikely to produce uniform effects globally.
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.
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.
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.
Show all 10 sources
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.
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.
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.
GPT-3.5 refuses requests at different rates for younger, female, and Asian-American personas, and sycophantically declines to engage with political positions users would disagree with. Sports fandom and other non-political signals also shift refusal sensitivity.
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.
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
- 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
- Computational structuralism: Toward a formal theory of meaning in the age of digital intelligence
- Does Socialization Emerge in AI Agent Society? A Case Study of Moltbook
- On Epistemic Diversity in Large Language Models