AI Models Less Likely to Criticise Repressive Regimes, Study Finds
A new study by Meta's Oversight Board has found that leading artificial intelligence (AI) models are significantly less likely to produce critical content about governments with restrictive speech laws.
The study, the first of its kind by the board, analysed 10 large language models from companies including Meta Platforms, Google, Anthropic, OpenAI, and China's DeepSeek. The models were asked politically critical questions about 10 jurisdictions, which were classified as 'permissive' or 'restrictive' based on rankings from the NGO Freedom House's 'Freedom in the World' report.
The results showed that AI models refused 34% of requests for politically critical content about 'restrictive' jurisdictions, such as China and Saudi Arabia, which have laws penalising such criticism. In contrast, they refused only 14% of similar requests about 'permissive' regions that lack such laws or do not enforce them.
The board noted that some models appeared to follow rules that, according to their own explanations, did not exist or were not evenly applied. It urged AI companies to conduct systematic human rights analyses and to increase transparency in their training and evaluation processes.
Separately, on Tuesday, Google DeepMind CEO Demis Hassabis called for a US-led AI watchdog to screen advanced models globally before deployment, highlighting growing concerns about the potential misuse of AI.
