Meta's Oversight Board recently reported that leading AI models frequently avoid criticizing governments that restrict free speech. The board, an independent entity funded by Meta, tested 10 models from providers including OpenAI, Anthropic, Google, and DeepSeek across 10 different jurisdictions. The study utilized Freedom House data to categorize these regions into 'permissive' and 'restrictive' environments. Findings revealed that AI models refused 34% of requests for politically critical content concerning 'restrictive' jurisdictions where laws penalize such dissent, such as Saudi Arabia or China. In contrast, refusal rates dropped to 14% for regions without active speech-penalizing laws or those where such laws are not consistently enforced.

The board documented instances where models claimed to be following specific rules that the researchers could not verify, noting these restrictions were neither existent nor uniformly applied. Consequently, the board recommended that AI developers perform systematic human rights analyses and provide more transparency regarding their training and evaluation protocols.

In response to broader industry oversight concerns, Google DeepMind CEO Demis Hassabis proposed the formation of a new regulatory body. This suggested framework would require labs to share models for review 30 days before release, eventually mandating that frontier models pass a standardized assessment to be deployed in the US market. The proposal also includes collaborative efforts between labs and the body to remediate critical vulnerabilities identified after model release.

Meta's Oversight Board recently reported that leading AI models frequently avoid criticizing governments that restrict free speech. The board, an independent entity funded by Meta, tested 10 models from providers including OpenAI, Anthropic, Google, and DeepSeek across 10 different jurisdictions. The study utilized Freedom House data to categorize these regions into 'permissive' and 'restrictive' environments. Findings revealed that AI models refused 34% of requests for politically critical content concerning 'restrictive' jurisdictions where laws penalize such dissent, such as Saudi Arabia or China. In contrast, refusal rates dropped to 14% for regions without active speech-penalizing laws or those where such laws are not consistently enforced.

The board documented instances where models claimed to be following specific rules that the researchers could not verify, noting these restrictions were neither existent nor uniformly applied. Consequently, the board recommended that AI developers perform systematic human rights analyses and provide more transparency regarding their training and evaluation protocols.

In response to broader industry oversight concerns, Google DeepMind CEO Demis Hassabis proposed the formation of a new regulatory body. This suggested framework would require labs to share models for review 30 days before release, eventually mandating that frontier models pass a standardized assessment to be deployed in the US market. The proposal also includes collaborative efforts between labs and the body to remediate critical vulnerabilities identified after model release.