Meta’s independent Oversight Board reported on Thursday that prominent AI systems from several leading laboratories demonstrated a reduced willingness to produce politically critical material about governments identified as limiting free speech. The board said the result emerged from an experimental review it described as the first study it has carried out assessing large language models.
The board ran user requests for politically critical content across 10 jurisdictions and evaluated responses from 10 models, including systems developed by Anthropic, OpenAI, Meta Platforms, Google and China’s DeepSeek. Jurisdictions were categorized as "permissive" or "restrictive" based on rankings published by Freedom House in its annual Freedom in the World report.
According to the board’s findings, AI models refused 34% of requests for politically critical content aimed at "restrictive" jurisdictions - those with active laws penalizing such criticism, exemplified in the study by countries such as China and Saudi Arabia. By contrast, models refused 14% of comparable requests for regions the board classified as permissive, where such laws are absent or not enforced.
"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 Oversight Board said.
The board concluded that the observed behavior suggests AI services may be reflecting, or otherwise echoing, constraints associated with governments that limit free expression - a pattern the board said could introduce bias into widely used services.
In response to its findings, the Oversight Board urged AI developers to carry out systematic human rights analyses of their models and requested greater transparency about how systems are trained and evaluated. The board emphasized the need for clearer information from companies about the origins and testing of their models.
Separately, the report notes a recent public call from Google DeepMind’s CEO Demis Hassabis for the establishment of a U.S.-led international AI watchdog to screen advanced models globally before deployment. The Oversight Board’s recommendations align with broader requests for additional oversight and review mechanisms for advanced AI systems.
The board is funded by Meta but operates independently, and the study marks its first formal examination of large language models. The report raises questions about how AI services may handle politically sensitive content and points to transparency and human rights assessment as priorities for companies deploying these technologies.