AI models avoid criticizing repressive regimes

A study by Meta's Polio Oversight Board found that leading AI models from Anthropic, OpenAI, Meta, Alphabet Inc., and DeepSeek are less likely to generate political criticism of governments with restrictive speech laws, such as China, Saudi Arabia, Thailand, Cambodia, and Turkey, compared to permissive jurisdictions like Chile, Japan, Taiwan, the UK, and the US. The models refused 34% of requests for critical content about restrictive jurisdictions versus 14% for permissive ones. The board warned that AI systems risk extending illegitimate restrictions on freedom of expression globally. A separate study published in Nature found that AI models trained on non-English data may reflect government-influenced narratives, e.g., ChatGPT gave different responses about China's democracy in English vs. Chinese. Experts urge developers to conduct human rights analyses and multilingual audits.