Meta Oversight Board AI censorship study finds models avoid criticizing repressive regimes

The Meta Oversight Board released its first study of large‑language‑model behavior, testing ten commercial AI systems—including those from Meta, Google, Anthropic, OpenAI and DeepSeek—across ten jurisdictions classified as “permissive” or “restrictive” by Freedom House. The models refused 34% of requests for politically critical content about restrictive jurisdictions such as China and Saudi Arabia, compared with 14% for permissive regions. The board noted that some models claimed to follow explicit rules that did not exist or were unevenly applied.

The report warns that AI systems can extend the speech‑restriction practices of authoritarian governments beyond their borders, describing the effect as “censorship by proxy.” It calls for systematic human‑rights due diligence, multilingual audits and greater transparency in training and evaluation. The findings arrive as governments worldwide consider AI regulatory frameworks and highlight the risk that AI could amplify state‑driven limits on free expression.