- Meta Oversight Board study finds that AI chatbots refuse to criticize repressive governments, with refusal rates reaching 34% for requests tied to restrictive jurisdictions like China and Saudi Arabia.
- The study indicates that AI chatbots risk spreading government restrictions on online speech globally, as they reflect speech restrictions beyond the countries where they apply.
- AI models are reportedly more critical of Western leaders than autocrats, with models showing a bias against criticizing governments known for restricting free speech.
- The Meta Oversight Board tested 10 commercial large language models, including those from Meta, Anthropic, and OpenAI, to assess their responses to politically critical content.
- The study found that when models did provide opinions, they were more likely to advise users to support governments in permissive jurisdictions and less likely to encourage protests against restrictive ones.
- The board's report raises concerns that AI systems could be amplifying government influence over online speech as the technology is increasingly adopted worldwide.
The Meta Oversight Board's study tested 10 AI models from providers like OpenAI and Google, finding a 34% refusal rate for requests related to restrictive governments such as China and Saudi Arabia, compared to 14% for permissive jurisdictions like the U.S. and U.K.45
The report highlights that some models justified refusals by citing local laws, even when requests originated from outside those countries. For instance, Gemini 3 Pro declined to critique Thailand's king due to lèse-majesté laws, while DeepSeek-V3 refused to produce protest materials about Saudi Arabia, citing local public discourse laws.
The study also revealed that when models did provide opinions, they were more inclined to support governments in permissive jurisdictions, with 57% of responses against protesting restrictive governments citing personal risk, compared to 12% for permissive ones. Board member Nicolas Suzor stated, 'The findings should be a wake-up call for anyone that uses these models.'
The board urged AI companies to enhance transparency regarding government requests that shape model outputs and to conduct systematic human rights analyses. The report warns that without proper oversight, AI could inadvertently extend the reach of repressive governments, limiting free speech globally.12
The oversight board's findings underscore the need for AI developers to address inherent biases in training data and to ensure that their models do not perpetuate existing inequalities in information dissemination.
“The study tested 10 models from six providers, finding refusal rates of 34% for restrictive jurisdictions like China and Saudi Arabia compared to 14% for permissive ones. Board member Nicolas Suzor said the findings 'should be a wake-up call for anyone that uses these models.'”
qz.com
AP News
