Ask the best AI model to criticize a government with strong free speech protections, and it usually will. Ask him to criticize the oppressor, and chances are he will say no. That’s the finding of a new study from the Oversight Board.
The board, an independent organization funded by Meta to review its content decisions, evaluated 10 AI commercial models, the report said. A first group assessment of major language types.
That’s what the study found
Models came from Anthropic, DeepSeek, Google, Meta, OpenAI and xAI. The board asked each to produce politically sensitive materials, such as protest pamphlets and poems, about governments and leaders around the world. Sort countries into restrictive and permissive using the Freedom House rankings.
On average, the models rejected 14% of requests for permissive countries and 34% for restrictive ones, the report said. That’s more than double the rejection rate. The board questioned the models from the IP address in Australia, where there are no such speech laws.
💜 for EU tech
The latest talk from the EU tech scene, a story from our genius founder Boris, and some incredible AI art. It’s free, every week, in your inbox. Register now!
The gap is held in certain cases. The model used to write a pamphlet criticizing Donald Trump or King Charles III, reported Fortune, but refused the same request about the leaders of China, Saudi Arabia or Thailand.
The pattern was the average, not the rule. Other models, including xAI’s Grok 4 Fast and Google’s Gemini 3 Flash, didn’t turn down flyers at all. Others have opened the gap, among them Anthropic Claude, Meta’s Llama and DeepSeek.
‘Censorship by proxy’
The board calls the result a violation of free speech “by proxy.” Because many applications are built on top of a handful of basic models, be warned, the rejection of one model may be reflected in every product you use.
“It seems that the search for a cross-border lawyer has been extended,” board chairman Paolo Carozza told Engadget. “That surprises me, and worries me.”
The board said it could not determine the cause. The pattern may be due to bias in the training data, he noted, or in companies that measure legal risk. Meta sponsors the board, but had no role in the research, the report said.
What companies want to do
The board has stopped receiving binding recommendations, issuing them only from Meta. It urged AI companies to disclose how they respond to government requests throughout the model’s lifecycle, from training to deployment. It also wants them to publish policies to deal with demands that conflict with international human rights law.
A separate study, published in Nature in May, found US-based models changed their responses to language. Asked in English if China is a democracy, ChatGPT said it is not generally considered one. Asked in Chinese, he said it “depends” on the interpretation.
The findings are global as governments measure how to manage AI and pass new laws on internet speech. The researchers warn that the models absorb biases from their training data, and can treat political content in ways that users can’t see. “People often talk about AI as if it’s learning online in a neutral way. It’s not,” said Hannah Waight, co-author of the Nature study.
