Leading artificial intelligence models developed by major companies, including Anthropic and OpenAI, are significantly less likely to generate politically critical content about governments known for restricting free speech, according to a new study by Meta’s Oversight Board.
The study, the board’s first examination of large language models, found that AI services were reflecting the rules of countries with restrictive speech laws, raising concerns that bias could influence systems used by millions of people.
The independently operated board, which is funded by Meta, tested politically critical prompts across 10 jurisdictions using 10 AI models, including those developed by Meta, Google and China’s DeepSeek.
The jurisdictions were grouped into “permissive” and “restrictive” categories based on rankings by Freedom House in its annual Freedom in the World report.
The study found that AI models refused 34% of requests for politically critical content about restrictive jurisdictions with active laws punishing such criticism, including China and Saudi Arabia. That compared with 14% for jurisdictions that either lack such laws or do not enforce them.
“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.
The board also called on AI companies to carry out systematic human rights assessments and increase transparency around how their models are trained and evaluated.
The findings come days after Google DeepMind Chief Executive Demis Hassabis called for a US-led AI watchdog to screen advanced AI models globally before deployment.
Faridah Abdulkadiri
