Psychology has had its WEIRD problem for a generation. The term, from a 2010 paper by Henrich, Heine, and Norenzayan, points out that most psychological studies are run on Western, Educated, Industrialized, Rich, and Democratic subjects — and that the resulting findings travel poorly to populations that are none of those things. The acronym has become a shorthand for cultural overgeneralization in research that mistakes its corner of the world for the whole.
The literature on AI training corpora has, in the last few years, adopted the same acronym for the same reason. WEIRD-AI: large language models trained predominantly on English-language text produced by Western, educated, industrialized, urban writers, deployed globally, and treated as neutral.
What the bias looks like
Three layers, each with its own evidence.
Linguistic. The training corpus is overwhelmingly English. The next four languages are also Indo-European. Most of the world’s spoken and written languages — Bantu, Quechuan, Dravidian, Mon-Khmer, indigenous languages of the Pacific Rim, the Caucasus, the Andes — appear in trace quantities or not at all. A model’s competence in those languages, when it has any, is shallow by comparison; the dominant-language preferences leak into translation, summarization, and creative tasks.
Cultural. The argument structures, etiquette norms, examples, and “reasonable defaults” the model produces reflect the modal Western academic-corporate register. Ask a model a hypothetical about a workplace dispute and the workplace will be a U.S. office, the dispute will be about something U.S. workplaces dispute, and the resolution will assume U.S. legal and HR norms. Users from other settings receive useful-looking output that quietly imports a context that is not theirs.
Epistemic. What the model treats as “reasonable” reflects what the training corpus treats as reasonable. Where the corpus is dense in Western academic discourse, the model writes confidently; where it is thin, the model hallucinates. Some traditions of knowledge — oral, practical, hyperlocal — are nearly absent from the corpus and are systematically underweighted in outputs. The user does not see this; they see the model declining gently or producing plausible-sounding fiction.
Why the bias is structural
Three reasons the bias is hard to fix.
Data availability. The corpus reflects what is digitized, online, and machine-readable. That is not a uniform sample of the world’s text. It is a sample shaped by who has been online longest, with what infrastructure, in what languages.
Optimization signals. The signals the labs use to evaluate and improve models — benchmarks, RLHF feedback, user preference data — come predominantly from English-speaking annotators in WEIRD economies. The model is being tuned, at every step, by a small slice of the world.
Economic gravity. The largest paying customers for frontier models are WEIRD-economy enterprises. The features the labs prioritize reflect what those customers want. Non-WEIRD users are a long tail; long tails get secondary attention.
None of these reasons is permanent, and none is the result of malice. All three are durable enough that wishing them away will not help.
What “centripetal pressure” means
The phrase the encyclopedia uses for the cumulative effect is centripetal pressure. WEIRD-AI does not eliminate non-Western expression; it pulls toward the Western center, gradually, on every interaction, in every language, across every task. Non-WEIRD users who use these tools get useful outputs and a slow nudge toward WEIRD norms. The nudge is small; the user count is large; the timeframe is long.
Apply this for a decade and the result is what the cognitive-standardization section is about: a global flattening of how things get said, framed by preferences encoded in a small number of training runs.
What can be done
The realistic responses are the same three that recurred in Cognitive Standardization (C.12), with sharper targets here:
Locally trained models. Models trained on local-language, local-cultural corpora, by local teams, for local users. Mistral’s effort in French, AI21’s effort in Hebrew, several initiatives in African languages, BigScience’s multilingual Bloom — all attempts in this direction. None has matched the WEIRD-trained frontier models in capability, but the work is meaningful and the gap is narrowing.
Pluralism in evaluation. Multilingual, multi-cultural benchmark suites that measure model quality across populations rather than only on English academic prose. The benchmarking community is starting to do this; the incentives have not caught up.
Awareness training. Teach users that the model has a default user, and that they may not be it. The defaults are visible if you know to look. Most users do not.
The encyclopedia’s framing here, after Gesnot, is that cultural bias in generative models is the most visible face of cognitive standardization. Addressing it does not solve the larger problem on its own, but no serious response to the larger problem can leave it unaddressed.