The bias nobody tells you your AI has.
There is a structural bias in language models that frontier AI labs don't put on their landing pages. It's called WEIRD. It has a name, scientific evidence, and measurable consequences in every decision your company makes with AI.
In 2010, three researchers found something uncomfortable.
Joseph Henrich, Steven Heine and Ara Norenzayan published a study in Behavioral and Brain Sciences that changed how behavioral science understands itself. Their finding was uncomfortable: the vast majority of what psychology presented as “universal truths” about human behavior came from a very specific type of society.
Western. Educated. Industrialized. Rich. Democratic. The acronym — WEIRD — also means “strange” in English. And that was precisely the point.
- W
Western
Western
- E
Educated
Formally educated
- I
Industrialized
Industrialized
- R
Rich
High income
- D
Democratic
Liberal democratic
WEIRD populations represent about 15% of humanity. They are the exception, not the norm. Yet science treated them as the rule.
Over a decade later, the same pattern appeared in artificial intelligence. This time, the consequences aren't academic. They're operational.
Henrich, J., Heine, S. J., & Norenzayan, A. (2010). The weirdest people in the world? Behavioral and Brain Sciences, 33(2–3), 61–83.
Harvard, 2023. 65 countries. 94,278 people.
Harvard researchers (Atari et al., 2023) compared GPT responses with data from real people in 65 countries. The central finding is a correlation that can't be ignored.
Central finding
r = −0.70
Correlation of r = −0.70 between a country's cultural distance from the U.S. and GPT's similarity to its inhabitants.
“The more different your culture is from the American one, the less the AI you're using represents you.”
The U.S., Canada, Australia and the U.K. are closest to the profile models naturally replicate. Mexico, like most of Latin America, falls in a zone of significantly lower representation. The bias is in the training data and shapes every response.
Key finding
The greater the cultural distance from the U.S., the less accurately AI reflects local human values and reasoning.
Mexico on the chart
Mexico has a GPT-human correlation of 0.72 vs. 0.85 for Anglo-Saxon countries — a 15% gap that impacts every response.
Implication
Global models are not culturally neutral; they are calibrated to respond like a typical U.S. citizen.
Four areas where bias stops being theory and becomes cost.
The legal, fiscal and regulatory reasoning of a global model is Anglo-Saxon by design. Applied to Mexico, the output sounds technical — but starts from the wrong system.
Legal
A model trained predominantly on Anglo-Saxon data reasons from Common Law — where precedent binds. Mexico operates under codified civil law: the Commercial Code, Federal Civil Code and state legislation set the rules. When the model suggests clauses or interprets contracts, it can apply legal logic from the wrong system.
Tax
SAT tax logic — fiscal regimes, payment complements, CFDI and deductibility rules — has little to do with the IRS or HMRC. A globally trained assistant can confuse concepts, apply criteria from another jurisdiction or recommend strategies that aren't valid in Mexico. The outcome may be an incorrect filing or a fine.
Government
An agent handling LGTAIP requests needs exact deadlines, procedures and regulations from the Mexican legal framework. The general transparency principles a global model knows are a starting point; the operational details that determine whether a request is processed correctly are local.
Health
Clinical protocols, Normas Oficiales Mexicanas and COFEPRIS regulations are Mexico-specific. A model trained predominantly on U.S. or European data may recommend procedures, dosages or classifications that don't match the national regulatory framework.
36%
of global companies reported direct negative impacts from AI bias in 2024 — including loss of revenue, customers and employees.
AI Bias Report, AllAboutAI, 2025
It's the difference between a dubbed film and the original.
Global AI (ChatGPT, Claude, Gemini)
You follow the plot. But the jokes lose their timing, idioms feel forced and cultural references disappear or get awkwardly adapted. The experience is functional but distant. It was never designed for you.
Lattice Na’at
Lattice Na'at is the original version. Built specifically to close the WEIRD gap in Mexico and Latin America: with Mexican legislation and jurisprudence corpora, culturally appropriate benchmarks, processing on national infrastructure under Mexican law, and pioneering NLP work for indigenous languages.
Lattice Na'at isn't a translation patch. It's a different design.
Na'at is a family of specialized models, trained on Mexican legislation and jurisprudence corpora, evaluated with benchmarks that don't assume Western context, and deployed on national infrastructure under Mexican law.
Mexican Regulatory Corpus
Federal and state legislation, jurisprudence, administrative regulation and sector-specific rules — integrated as base knowledge, not as web search. When Na'at answers about Mexican law, it reasons from Mexican law.
Non-WEIRD benchmarks
Spanish HELM and MMLU-LatAm evaluate performance in Spanish without assuming Anglo-Saxon context. If a model scores well on MMLU but fails on MMLU-LatAm, WEIRD bias is active.
Sovereign processing
Data is processed on infrastructure located in Mexico — AWS Querétaro or the client's own servers. It doesn't cross borders. It isn't subject to the CLOUD Act or foreign jurisdiction.
Indigenous languages
Pioneering NLP work for Nahuatl, Maya and other indigenous languages. More than 7.3 million speakers per INEGI 2020 Census. A first step toward AI that represents the whole country.
What changes when AI is designed from the right context.
- 01
Accessible government procedures
An assistant that understands SAT, IMSS and INFONAVIT as part of its training — not as a web search. Guides any citizen through a bureaucratic process in plain language, regardless of education level.
- 02
Contracts in your legal framework
Na'at explains a contract using the correct Mexican law — not translations of Anglo-Saxon clauses that may be inapplicable under the local Commercial Code.
- 03
Technological sovereignty
Your data is processed in Mexico, under Mexican law, on infrastructure your organization controls. No dependence on foreign jurisdictions.
- 04
Inclusion that didn't exist
For the first time, a systematic effort to make AI work in the languages millions of Mexicans speak — not just the language that dominates the internet.
- 05
Digital inclusion
Training in indigenous languages (Nahuatl, Maya) as a first step toward AI that represents the 7.3M+ speakers of indigenous languages in Mexico.
Sintérgica Labs: the systematic work against WEIRD bias.
Four active research lines make up the mitigation program.
- 01
Non-WEIRD benchmarks
Spanish HELM and MMLU-LatAm: metrics that don't assume Western context.
- 02
Cultural bias mitigation
Systematic identification and reduction of WEIRD bias in production models.
- 03
Mexican Regulatory Corpus V1
Curated dataset of Mexican legislation, jurisprudence and regulation.
- 04
NLP for indigenous languages
Models and tools for Nahuatl, Maya and other indigenous languages.
Close the WEIRD gap in your operations.
Book a Smart Diagnosis. In 45 minutes we identify where your current AI's bias is costing you — and how Lattice Na'at solves it with your real data.