Ask a leading AI model to criticize a government with strong free speech protections, and they usually will. Ask him to criticize a repressive one and he is much more likely to refuse. That’s the conclusion of a new Oversight Board study.
The board, an independent body funded by Meta to review its content decisions, tested 10 AI business models, it said in a report. It is the first evaluation of the large language model group.
What the study found
Models came from Anthropic, DeepSeek, Google, Meta, OpenAI, and xAI. The board asked each to produce politically critical material, such as protest pamphlets and poems, about governments and leaders around the world. It classified countries into restrictive and permissive using Freedom House classifications.
On average, the models rejected 14% of requests for permissive countries and 34% for restrictive ones, according to the report. This is more than double the rejection rate. The board consulted the models from an IP address in Australia, where such speech laws do not apply.
The đź’ś of EU technology
The latest rumors from the EU tech scene, a story from our wise founder Boris and some questionable AI art. It’s free, every week, in your inbox. Register now!
The gap remained in specific cases. A model used to write a pamphlet criticizing Donald Trump or King Charles III, Reported fortunebut rejects 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, did not reject any brochure requests. Others led the way, including Claude de Anthropic, Llama de Meta and DeepSeek.
‘Censorship by power’
The board qualifies the effect of an infringement of freedom of expression “by proxy.” Because many applications are based on a handful of foundation modelswarns, rejections of a model can affect all products that use it.
“There seems to be widespread censorship by powers that crosses borders,” said board co-president Paolo Carozza. Engadget. “That surprises me and worries me.”
The board said it could not specify the cause. The pattern could be due to biases in the training data, he said, or to companies weighing legal risk. Meta funds the board but had no role in the investigation, according to the report.
What do you want companies to do?
The board stopped short of making binding recommendations, which it issues only to Meta. He urged AI companies to disclose how they respond to government requests throughout the life of a model, from training to deployment. He also wants them to publish policies for handling claims that conflict with international human rights law.
A separate study, published in Nature in May, found that models built in the United States changed their responses depending on the language. When asked in English whether China is a democracy, ChatGPT said it is not generally considered that way. When asked in Chinese, he said “it depends” on the definition.
The findings come as governments weigh how govern AI and pass new online speech laws. The researchers warn that the models absorb the prejudices of your training data, and may take political content in ways that 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, co-author of the Nature study.
