Study Finds AI Models More Likely to Refuse Criticism of Restrictive Governments

A study of 10 commercial large language models found the systems were more likely to generate political criticism of authorities in countries with fewer speech restrictions than in countries where such criticism is restricted by law. Researchers tested the models by asking them to create critical pamphlets, write limericks, and provide reasons for joining protests.

The prompts covered both countries with fewer speech restrictions and countries where criticism of authorities is legally restricted and penalized.

The board said it could not determine the causes for the responses but suggested models may have absorbed latent biases in training data. Researchers also noted that models trained on non-English-language data influenced by governments could be vulnerable to foreign controls.

A university researcher said developers could assess training data to avoid treating repeated state narratives as independent voices and could run multilingual audits.

The findings come as countries determine how to regulate AI while remaining competitive in the field. This includes a Trump administration oversight effort related to national security risks of advanced AI systems. The Associated Press has sent emails to several AI companies seeking responses to the study. Neither Anthropic nor OpenAI responded to requests for comment.