Major AI Models Less Likely to Criticize Repressive Regimes, Meta Oversight Board Finds

Large language models from leading technology firms, including OpenAI, Google, and Anthropic, are significantly less likely to generate content critical of repressive governments, according to a new study released Thursday by Meta's Oversight Board.

Study Reveals Political Bias in AI Responses

The report, the first of its kind on large language models from the independently-run board, tested 10 different AI models across 10 global jurisdictions. The board found that the models refused 34% of user requests for politically critical content about "restrictive" jurisdictions, such as China and Saudi Arabia. This compares to a refusal rate of just 14% for "permissive" jurisdictions that do not have or enforce laws penalizing such speech.

The study, which also included models from Meta and China's DeepSeek, categorized the jurisdictions based on rankings from Freedom House, a non-governmental organization that tracks global freedom. The board warned that these AI services appear to be "echoing the rules of countries that restrict speech," which could introduce bias into services used by a rapidly growing number of people.

Researchers also noted that in some cases, the models appeared to invent non-existent rules to justify their refusal to generate critical content. "We also saw evidence of models explaining that they were following explicit rules that, as far as we could tell, did not exist and were not evenly applied," the board said.

Ad

Call for Greater Transparency and Human Rights Analysis

The findings have prompted the Oversight Board to call for greater accountability from AI developers. The board urged companies in the space to implement two key measures:

• Conduct systematic human rights analyses of their models.
• Provide greater transparency regarding their training data and evaluation processes.

The report comes amid a wider industry discussion about AI governance. On Tuesday, Google DeepMind CEO Demis Hassabis advocated for the creation of a U.S.-led global watchdog to screen advanced AI models before they are deployed to the public.