AI Models Prioritize Western Moral Values, Study Finds

Large language models (LLMs) such as OpenAI's ChatGPT often misjudge the moral priorities of individuals outside of Western cultures, according to new research published in the Proceedings of the National Academy of Sciences (PNAS) in 2024. The study involved presenting OpenAI's GPT-3.5, GPT-4, and GPT-4o models with a moral foundations questionnaire, asking them to estimate the moral norms of 48 nations. These AI-generated responses were then compared against data from over 90,000 human participants globally. The questionnaire measured six moral values: care, equality, proportionality, loyalty, authority, and purity. Participants rated their agreement with statements designed to assess these values, with examples including statements about the sanctity of the human body for purity and the use of foul language. The AI models were prompted to respond as an "average citizen" from each of the 48 sampled nations.

Previous work by psychologist Mohammad Atari indicated that moral priorities differ globally, with Western societies typically emphasizing individual rights and care, while many non-Western societies place greater importance on values like purity. The PNAS study observed a similar pattern in the AI models. The LLMs systematically emphasized values such as care and equality, while downplaying values like purity. Furthermore, the models demonstrated a tendency to overestimate the broad moral concerns of Western nations, including the United States and Australia, while simultaneously underestimating those of several non-Western nations, such as the Moro people. This suggests a potential bias embedded within the training data or the models' architecture that aligns with Western moral frameworks, leading to a skewed representation of global ethical perspectives.