ChatGPT leans toward Western moral values, new study finds

A 2024 study published in Proceedings of the National Academy of Sciences (PNAS) found that large language models often make moral judgments through a WEIRD lens: Western, Educated, Industrialized, Rich, and Democratic.

ChatGPT Download

The researchers compared GPT-3.5, GPT-4, and GPT-4o with responses from more than 90,000 people across 48 nations. What they found was a pretty consistent tilt toward WEIRD norms, with the models often putting more emphasis on care and equality while giving less weight to values like purity, which carry more force in some non-Western contexts.

If you’re using chatbots for education, health information, moderation, or emotional support, that has real consequences. Training data drawn heavily from the English-language internet can wash out perspectives from the Global South, and separate studies show that ChatGPT and Llama2Chat-70B can shift moral judgments across languages.

So the risk isn’t abstract. You could end up with skewed public-health messaging, shaky cross-cultural guidance, or mental-health advice that lands badly, especially in communities where family loyalty and collective duties matter more than individual boundaries.

Researchers are now trying cultural prompting as one way to reduce the problem. A Cornell-led study reported lower bias across most of the 100-plus countries it examined, while other scholars are pushing for a more pluralistic approach to alignment, one that makes room for traditions like Ubuntu and language-centered efforts such as Masakhane.

The full study is available in PNAS.