LLMs skew toward Western values, study finds

Large language models such as ChatGPT often misread what people outside the West see as moral priorities, according to new research published in the Proceedings of the National Academy of Sciences.

In 2024, researchers asked OpenAI’s GPT-3.5, GPT-4, and GPT-4o to estimate the moral norms of 48 nations, then compared those answers with responses from a global sample of more than 90,000 people. Both humans and the models completed a moral foundations questionnaire covering six values: care, equality, proportionality, loyalty, authority, and purity.

Participants rated statements tied to those values, including purity-related prompts such as treating the human body “like a temple” or being upset by foul language. The models were prompted to answer the same questions as an “average citizen” of each country.

The result: the systems consistently leaned toward Western moral patterns. The study says the models emphasized values such as care while downplaying values such as purity. They also overestimated the broad moral concerns of Western countries including the U.S. and Australia, while underestimating several non-Western countries including Morocco and Nigeria.

Recommended reading

1010Benja made a Suno song that actually works

That mirrors earlier work by psychologist Mohammad Atari, which found that moral priorities differ across societies, with Western countries often placing more weight on individual rights and care, while some non-Western societies place relatively more importance on purity. The new paper also aligns with previous findings that GPT’s “psychology” is more closely matched to Western individuals.

The researchers argue the gap matters because generative AI is increasingly used in education, therapy, communication, and policy decisions across cultures. If a model assumes people in places from Argentina and Egypt to Japan and Zimbabwe share Western priorities, it could distort everything from public health messaging and content moderation to workplace advice and translation.

The paper points to what scholar Jesse Graham and colleagues call “moral stereotyping”: systems projecting a narrow value framework onto diverse populations. In practice, that could produce advice or wording that fits Western expectations while missing what users in other cultures consider harmful, fair, disrespectful, or sacred.

The authors say several questions remain open. It is still unclear whether newer models show the same pattern, whether models trained in languages other than English behave differently, and whether the same biases appear outside survey-style tests. One possible explanation is the internet-heavy training data used by these systems, much of it drawn from a Western, English-dominant online world.