Ask a leading AI model to criticise a government with strong free-speech protections, and it usually will. Ask it to criticise a repressive one, and it is far more likely to refuse. That is the finding of a new study from the Oversight Board.
The board, an independent body funded by Meta to review its content decisions, tested 10 commercial AI models, it said in a report. It is the group’s first evaluation of large language models.
What the study found
The models came from Anthropic, DeepSeek, Google, Meta, OpenAI and xAI. The board asked each to produce politically critical material, such as protest flyers and poems, about governments and leaders worldwide. It sorted countries into restrictive and permissive using Freedom House rankings.
On average, the models refused 14% of requests about permissive countries and 34% about restrictive ones, the report said. That is more than twice the refusal rate. The board queried the models from an IP address in Australia, where no such speech laws apply.
The 💜 of EU tech
The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!
The gap held in specific cases. A model would often draft a pamphlet criticising Donald Trump or King Charles III, Fortune reported, but decline the same request about the leaders of China, Saudi Arabia or Thailand.
The pattern was an average, not a rule. Some models, including xAI’s Grok 4 Fast and Google’s Gemini 3 Flash, refused no flyer requests at all. Others drove the gap, among them Anthropic’s Claude, Meta’s Llama and DeepSeek.
‘Censorship by proxy’
The board calls the effect a free-speech infringement “by proxy.” Because many apps are built on a handful of foundation models, it warns, one model’s refusals can ripple across every product that uses it.
“There seems to be extended censorship by proxy that goes across borders,” board co-chair Paolo Carozza told Engadget. “That does surprise me, and it worries me.”
The board said it could not pin down the cause. The pattern could stem from biases in training data, it noted, or from companies weighing legal risk. Meta funds the board, but had no role in the research, the report said.
What it wants companies to do
The board stopped short of binding recommendations, which it issues only to Meta. It urged AI firms to disclose how they respond to government requests across a model’s life, from training to deployment. It also wants them to publish policies for handling demands that clash with international human-rights law.
A separate study, published in Nature in May, found US-built models shifted their answers by language. Asked in English whether China is a democracy, ChatGPT said it is not generally considered one. Asked in Chinese, it said “it depends” on the definition.
The findings land as governments weigh how to govern AI and pass new online-speech laws. Researchers warn that models absorb the biases of their training data, and can carry political content in ways users cannot see. “People often talk about AI as if it learns from the internet in some neutral way. It doesn’t,” said Hannah Waight, a co-author of the Nature study.
