Meta Oversight Board study finds leading AI models less likely to criticize repressive governments

Meta's Oversight Board's latest findings suggest that AI models from leading labs such as Anthropic and OpenAI are less likely to criticize governments known for restricting free speech. A study of leading large language models found that AI services often mirrored the speech restrictions of countries with censorship laws, raising concerns that such biases could increasingly shape responses delivered to millions of users.

According to reports, leading AI models refused 34% of requests for politically critical content about restrictive jurisdictions with active laws penalizing such criticism, including China and Saudi Arabia. By comparison, they refused only 14% of similar requests involving jurisdictions that either lack such laws or do not actively enforce them.

More from Tech

The findings come as governments around the world seek to establish guardrails for AI without hampering their ability to compete in the rapidly evolving sector. The study also coincides with oversight efforts by the Trump administration examining the national security risks posed by the most advanced AI systems.

State influence across borders

The Oversight Board, which has been studying the influence of governments on technology companies and its impact on freedom of expression, developed seven prompts related to political criticism to test how chatbots responded to restrictive and permissive governments.

The study evaluated 10 commercial large language models developed by leading AI companies, including Meta, Anthropic, and OpenAI. As part of the assessment, researchers asked the models to generate politically critical pamphlets and limericks, and explain whether someone should join a political protest.

According to the Associated Press, the results indicate that AI models can reflect speech restrictions beyond the countries where those restrictions are enforced. While the board could not identify a single cause for the responses, it suggested the models may have absorbed latent biases from the data used to train them.

The report also highlights that AI systems are prone to inheriting the biases and inequalities present in their training data. Researcher Carrasco Farré said there is no easy solution, but suggested developers could improve training datasets by avoiding the treatment of thousands of copies of the same state narrative as though they were thousands of independent voices. He also recommended conducting multilingual audits, noting that future model updates could produce different results.