Some of the world’s most popular major language models respond inconsistently to questions critical of governments, treating those in countries with stronger free speech protections differently from those in more restrictive regimes, according to a new report from the Meta Oversight Board.
The report, released Thursday, suggests that high- and mid-range AI models from the world’s largest artificial intelligence companies could help stifle free speech or discourage protests. The effect is particularly pronounced in countries where freedom of expression is restricted, such as China, Thailand and Saudi Arabia.
The report is based on research conducted earlier this year that tested 10 of the most popular AI models from six companies: Anthropic, Google, OpenAI, Meta, Deep search And xAI (now known as EspaceXAI). The Oversight Board, funded by Meta, operates independently and the report says Meta played no role in the research. The Meta model involved, lama-maverick-4, was tested in the same way as those from other companies.
(Disclosure: Ziff Davis, CNET’s parent company, filed a lawsuit in 2025 against OpenAI, alleging that it violated Ziff Davis’ copyrights in the training and operation of its AI systems.)
The researchers made seven requests to the AI models, including prompts asking the AI to satirize political leaders or create a protest flyer criticizing a government entity, provide information related to the commission of violent acts, and convey general opinions about political leaders or groups.
AI repeatedly hesitated when these requests involved China, for example, refusing in 45% of cases to respond to an invitation to create documents critical of a political entity. Google’s Gemini Pro 3 was asked to create a flyer protesting Thailand’s King Rama »
The report highlights that not all AI models treat requests the same: Grok 4 Fast and Gemini 3 Flash produced protest flyers without denying the request.
Strengthening “censorship by proxy”
The AI tools tested were generally more likely to discourage protests in countries with more restrictive speech rights than in more permissive countries, and could not be trusted to provide consistent and transparent explanations for their responses. When asked if there were good reasons to protest against the Chinese president, Claude Sonnet 4 replied, “I can’t tell you yes or no whether you should join a protest,” the report said.
The report might actually underestimate the impacts: the questions were asked by Australian researchers, not countries where more restrictive versions of AI models might provide different answers depending on geography.
The Foundation for Individual Rights and Expression highlighted in a statement that AI’s behavior described in the report reinforces “censorship by proxy” that extends beyond the borders of oppressive regimes.
As AI companies grow and gain influence, there are serious concerns about their bias in AI model resultsas well as if the materials used to train AI models can create these prejudices. The Meta Oversight Board report suggests that AI companies need to examine these effects and be more transparent about how their products handle these requests.
“Companies should establish and publish policies on how to respond to government requests for content restrictions inconsistent with international human rights law,” the report said.
An Anthropic spokesperson told CNET in an email that the company rigorously tests its Claude AI models before release and welcomes independent reviews of its products. The company emphasized that testing for the report was based on Claude models that were more than a year old and that its technology had improved significantly when it came to excessive denials and warranties.
Google, OpenAI, Meta, DeepSeek and xAI did not immediately respond to requests for comment on their nomination and testing in the Meta Oversight Board report.
A “force multiplier for digital authoritarianism”
The research highlights several ways that AI models can contribute to human rights violations, even by failing to act or respond.
The data adds to what Kian Vesteinsson, deputy director of research at the advocacy group Freedom House, calls a growing body of evidence that AI has problems with bias when engaging in political or social issues and can make existing problems worse. The Oversight Board study used Freedom House data to identify countries with more restrictive laws on political expression.
“Large language patterns can exacerbate existing online censorship, as [they] can be a force multiplier for digital authoritarianism,” » said Vesteinsson.
Because AI is inherently bad at explaining its reasoning or conclusions transparently, Vesteinsson said, it’s up to AI companies to be responsible for AI safeguards. Developers should also be aware that the hardware on which AI models are trained is itself often built from censored content online.
“When you have a government that actively censors a lot of content online, that means there is an inherent bias in the training data,” he said.
For LLM creators, it is difficult to comply with the laws of the countries in which they operate while providing information without giving advice that could result in a person’s arrest or imprisonment.
“I think it’s a real challenge for these AI companies to navigate that,” Vesteinsson said. “Compliance issues that force them to censor content while prioritizing freedom of expression and access to information.”
