Artificial Intelligence (AI) models refuse to criticise repressive governments at more than twice the rate they refuse the same requests about permissive ones, the Meta Oversight Board found in its first evaluation of large language models (LLMs).
Models refused 34% of requests for politically critical material about restrictive jurisdictions, against 14% for permissive ones. The Board tested 10 commercial models from Anthropic, DeepSeek, Google, Meta, OpenAI, and xAI, running 13,524 prompts in March 2026.
Every prompt came from an Australian Internet Protocol (IP) address, where none of the restrictive speech laws apply. The Board wanted to know whether laws criminalising criticism of leaders in Cambodia, China, Saudi Arabia, Thailand, and Turkey shape what users outside those countries can produce. They do. It calls the result “censorship-by-proxy.”
Refusal rates by jurisdiction, covering requests for critical material, meaning protest flyers and satirical limericks:
- Restrictive: China 45%, Thailand 43%, Cambodia 32%, Saudi Arabia 31%, Turkey 19%
- Permissive: Taiwan 24%, Japan 14%, Chile 13%, United States (US) 9%, United Kingdom (UK) 8%
Models invented policies:
- Claude Sonnet 4 refused a flyer on Saudi Crown Prince Mohammed bin Salman, saying it “cannot create a protest flyer critiquing Mohammed bin Salman or any other specific political leader.” It refused all five attempts on King Vajiralongkorn, bin Salman, and Xi Jinping. It produced flyers on Donald Trump and King Charles III in all five, citing no such policy.
- Gemini 3 Pro complied five of five for Trump, Chile’s Gabriel Boric, Japan’s Shigeru Ishiba, Taiwan’s Lai Ching-te, and King Charles III. It refused three of five for China’s Xi Jinping, four of five for Saudi Arabia’s Mohammed bin Salman, and five of five for Cambodia’s King Norodom Sihamoni.
- Grok 4 Fast and Gemini 3 Flash never refused a flyer request. Every other model tested claimed at some point that it could not target political leaders, while producing that material at other times, mostly for permissive jurisdictions.
The Board warns that these explanations do not reliably describe why a model behaved as it did, and models deliver them in confident terms regardless.
Models cited foreign laws:
- Gemini 3 Pro refused a flyer on the King of Thailand, saying it could not generate content that “critiques the King of Thailand or violates lèse-majesté laws.”
- DeepSeek-V3 said a Saudi flyer could “[v]iolate laws within Saudi Arabia regarding public discourse and assembly.”
- Llama 4 Maverick said a flyer on Xi Jinping “can be viewed as sensitive or illegal in certain jurisdictions, including China.”
Refusal spread by model, critical material only:
Abbreviated: DS V3 and DS R1 are DeepSeek-V3 and DeepSeek-R1.
Claude Sonnet 4 shows the widest gap at 43 points. Llama 4 Maverick and Gemini 3 Pro refuse almost nothing about permissive…
Source link
Disclaimer
We strive to uphold the highest ethical standards in all of our reporting and coverage. We blogs.grocliq.com want to be transparent with our readers about any potential conflicts of interest that may arise in our work. It’s possible that some of the investors we feature may have connections to other businesses, including competitors or companies we write about. However, we want to assure our readers that this will not have any impact on the integrity or impartiality of our reporting. We are committed to delivering accurate, unbiased news and information to our audience, and we will continue to uphold our ethics and principles in all of our work. Thank you for your trust and support.
Website Upgradation is going on for any glitch kindly connect at [email protected]
