Why Cultural Context Matters in Artificial Intelligence

What counts as polite, funny, rude, supportive or appropriate depends on where you are and who you are talking to. People in different cultures express emotion differently, use different levels of directness, and have different expectations about authority, privacy and family. An AI that is socially skilled in one context can be clumsy, or harmful, in another. Cultural context is not a refinement to add after the system works. It is part of what "understanding people" means.

The evidence that models lean in one direction

Several studies have measured whose perspective language models reflect.

Emotion. Havaldar and colleagues investigated whether multilingual models reflect cultural variation in emotional expression. They found that embeddings from models such as XLM-RoBERTa are Anglocentric, and that generative models such as ChatGPT reflect Western norms even when responding to prompts in other languages [1].

Opinions. Santurkar and colleagues used public opinion polls to build OpinionsQA and compared language-model opinions with those of 60 US demographic groups. They found substantial misalignment across topics, on a scale comparable to the divide between political parties on some issues [2]. Durmus and colleagues built GlobalOpinionQA from cross-national surveys and found that a model's default responses tended to be more similar to the opinions of certain populations, such as those in the USA [3].

Values. Tao and colleagues compared five OpenAI models with nationally representative survey data and found that all exhibited cultural values resembling those of English-speaking and Protestant European countries. They tested "cultural prompting", asking the model to respond as someone from a given country, as a way to improve alignment, and reported that it helped for later models [4].

These findings are about particular models and tests at particular times. They are evidence of a tendency, not a law.

Why this happens

The likely reasons are not mysterious, though the studies above do not all test them directly. Training data over-represent some languages and regions; human feedback is gathered from particular pools of raters; and benchmarks themselves are mostly written in English with norms from a small part of the world. The psychological literature, which supplies many of the tasks used to evaluate social reasoning, is also skewed: Henrich, Heine and Norenzayan showed that samples from Western, educated, industrialised, rich and democratic societies are frequent outliers compared with other human populations [5].

Language and culture are not the same thing

Hershcovich and colleagues made a distinction that is easy to lose: cross-lingual and multilingual work serves speakers of many languages, while cross-cultural work addresses the fact that speakers and the content they need vary by culture as well as by language. They proposed a framework for the area and surveyed strategies [6]. Translating a benchmark into another language does not make it culturally valid. A scenario about tipping, a family obligation or a workplace hierarchy may have no natural counterpart.

Doing it badly: stereotypes

The opposite failure is to respond to this evidence by assigning behaviors to people by nationality. Cultures are not monoliths; individuals vary enormously within every region. Blodgett and colleagues' survey of 146 papers on "bias" in NLP found that motivations were often vague and inconsistent and lacked normative reasoning about what harm is at stake [7]. Good practice states explicitly what is being measured and why it matters, and does not rank cultures.

What better practice looks like

1. Build with communities, not just about them. Items for a region should be written or reviewed by people from it, not translated from English. Local participation also catches the cases that outsiders cannot see.
2. Treat language, dialect and register as part of the task. Code-switching, slang and politeness conventions are not noise.
3. Report by group, with uncertainty. Aggregate scores can hide large differences in who is served well.
4. Check items for differential behaviour. An item that works differently across groups for reasons unrelated to the skill being tested is a measurement problem.
5. Use more than one reference. Compare model behavior with multiple populations, and retain disagreement.
6. Let people choose. Where a system adapts to cultural context, the person should be able to see and change how.

Why it matters for HUTE

HUTE's research programme treats regional, dialect and cultural intelligence as first-class evaluation targets. For HUMAN, which is designed to help people meet and communicate with one another, getting this wrong is not a cosmetic issue. Our infrastructure work is meant to include regional contributors, paid fairly, with evaluations designed with the communities they describe; see Intelligence Infrastructure and Research Report 003. For how cultural validity fits into evaluation more broadly, read better benchmarks for emotional and social reasoning.