Meta’s Oversight Board reports that leading AI models are twice as likely to refuse content critical of restrictive governments. The study tested 10 models from developers including Meta, Google, OpenAI, Anthropic, and xAI.
Models refused 34% of requests for politically critical content regarding countries like China and Saudi Arabia. This compares to a 14% refusal rate for more permissive nations.
The Board warns these findings suggest a trend of censorship by proxy. It urges all AI developers to increase transparency and conduct human rights analyses to prevent technology from extending the reach of authoritarian censorship.