Moral stereotyping in large language models.

Area of Science:

• Computational Social Science
• Cross-Cultural Psychology
• Artificial Intelligence Ethics

Background:

• Large Language Models (LLMs) are increasingly used for data analysis, but their accuracy in representing diverse human values is unknown.
• Understanding cross-cultural moral values is crucial for global cooperation and ethical AI development.

Purpose of the Study:

• To evaluate the accuracy of LLMs in estimating the moral values of populations across 48 nations.
• To compare LLM-generated moral perceptions with large-scale survey data on six key moral values.

Main Methods:

• LLMs were prompted to estimate the moral norms of the 'average' person in 48 countries.
• LLM outputs were compared against empirical survey data on six moral values: Care, Equality, Proportionality, Loyalty, Authority, and Purity.

Main Results:

• LLMs demonstrated poor accuracy in capturing global moral diversity, systematically misrepresenting moral values.
• Models overestimated values like Care and underestimated values like Purity.
• LLMs exhibited biases, overestimating Western moral concerns and underestimating non-Western ones, indicating stereotyping.

Conclusions:

• LLMs are unreliable for generating cross-cultural moral value estimations.
• The reliance on LLMs for such tasks poses significant ethical and epistemic risks due to inherent biases and inaccuracies.
• Findings underscore the need for caution when using LLMs in sensitive cross-cultural research.