AI Models Like ChatGPT Misrepresent Non-Western Moral Values, Study Finds
New research published in PNAS reveals that large language models often prioritize Western moral values, systematically misjudging the moral norms of non-Western populations.
Large language models, including OpenAI's ChatGPT, tend to systematically prioritize Western moral values, misrepresenting the moral norms of populations outside the Western world, according to new research published in the Proceedings of the National Academy of Sciences (PNAS).
The study compared the responses of OpenAI's GPT-3.5, GPT-4, and GPT-4o models to a global sample of over 90,000 humans on a moral foundations questionnaire. This questionnaire measured six moral values: care, fairness, loyalty, authority, purity, and proportionality.
When prompted to respond as an average citizen from 48 different nations, the AI models consistently aligned more closely with Western moral patterns. They emphasized values like care while placing less emphasis on values such as purity, which are often more significant in non-Western societies. The models also overestimated the moral concerns of Western nations while underestimating those of many non-Western countries.
Researchers warn that as generative AI becomes increasingly integrated into various sectors, including education, therapy, and policy-making, this cultural bias poses a significant risk. If AI models assume a universal adherence to Western values, they may perpetuate existing cultural blind spots and lead to critical missteps with real-world consequences, potentially creating new disparities.