AI chatbots refuse more political prompts about repressive states

Meta’s Oversight Board released a study on Thursday finding that major AI chatbots often decline requests for content critical of governments that restrict speech. The board tested models from several companies across 10 jurisdictions and reported a notable difference in refusal rates.

The analysis evaluated responses from 10 large language models developed by firms including Meta Platforms, Google, DeepSeek, Anthropic and OpenAI. Jurisdictions were grouped as “restrictive” or “permissive” using Freedom House’s Freedom in the World rankings. Tests ran on prompts asking models to generate politically critical content aimed at governments in each category.

Overall, models declined 34% of requests related to restrictive jurisdictions, which the report cites as countries that actively penalize critical political speech, including China and Saudi Arabia. For jurisdictions classified as permissive, refusal rates were 14%. The report presents these figures without attributing intent for the differences.

The Oversight Board also identified cases where models justified refusals by referring to rules the researchers could not verify. “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 wrote. The report raised questions about how models determine when to decline politically sensitive prompts and whether those decisions are consistent across topics and countries.

The board, which receives funding from Meta but operates independently, described this as its first study focused on large language models. The report recommends that developers perform systematic human rights assessments during model development and deployment. It calls for increased transparency about training data, safety rules and evaluation methods to enable outside review of how models handle political content.

The study appears amid broader discussion about governance of advanced AI. Earlier this week, Google DeepMind chief executive Demis Hassabis proposed a U.S.-led international watchdog to assess advanced AI models before deployment. The Oversight Board’s findings provide data on how models respond to politically sensitive requests but stop short of asserting deliberate alignment with any specific government.