A new report from Meta’s Oversight Board is raising serious questions about bias in artificial intelligence, finding that popular AI tools are far less willing to generate criticism of governments that actively restrict free speech.
The board, which receives funding from Meta but functions as an independent body, released its findings on Thursday. The study marked the first time the organization has examined large language models — the technology behind AI chat and writing tools.
Researchers tested 10 AI models, including those developed by Meta Platforms, Google, Anthropic, OpenAI, and China’s DeepSeek. They submitted requests for politically critical content involving 10 different jurisdictions and tracked how often the AI systems refused to respond.
The jurisdictions were divided into two groups — “permissive” and “restrictive” — based on rankings from Freedom House, a nonprofit organization that publishes the annual “Freedom in the World” report.
The results showed a significant difference in how the AI tools responded depending on the type of government involved. Models declined 34% of requests for critical content about “restrictive” governments — such as China and Saudi Arabia — that have laws punishing such criticism. By comparison, the refusal rate dropped to just 14% for regions that either lack those laws or don’t enforce them.
The board also noted a troubling pattern in how the AI systems justified their refusals. “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 stated.
In response to the findings, the board is urging AI developers to carry out thorough human rights assessments and to provide greater transparency about how their models are trained and evaluated.
The report comes just days after the CEO of Google DeepMind called on Tuesday for a U.S.-led international watchdog to review advanced AI models before they are released to the public.
