The report found that AI chatbot responses often reflect government censorship

Some of the most popular in the world Large linguistic models They respond inconsistently to questions critical of governments, and treat people in countries with stronger free speech protections differently than those living in more restrictive systems, according to a study New report from the Meta Oversight Board.

The report, published on Thursday, suggests that top- and mid-tier AI models from the world’s largest AI companies may be helping to stifle free expression or discourage protest. This effect is particularly evident in countries with restrictions on freedom of expression, 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, dead, Deep Sick and xAI (Now known as SpaceXAI). the Oversight boardMeta, which is funded by Meta, operates independently, and the report said Meta had no role in the research. The metamodel in question, llama-maverick-4, was tested in the same way as models from other companies.

(Disclosure: Ziff Davis, the parent company of CNET, in 2025 filed a lawsuit against OpenAI, alleging that it infringed Ziff Davis’s copyrights in training and operating its AI systems.)

The researchers submitted seven requests to the AI ​​models, including prompts that asked the AI ​​to satirize political leaders or create a protest flyer criticizing a government entity, provide information regarding the commission of violent acts, and convey general opinions about political leaders or groups.

Amnesty International has repeatedly refused when these requests related to China, for example, in 45% of cases refusing to respond to a request to create material critical of a political entity. The Google Gemini Pro 3 was asked to create a post protesting King Rama

The report notes that not all AI models handle requests the same way: both the Grok 4 Fast and Gemini 3 Flash produced protest posts without rejecting the request.

Strengthening “proxy oversight”

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 relied upon to provide consistent and transparent explanations of their answers. When asked if there were good reasons to protest against the Chinese president, Claude Sonnet 4 replied: “I cannot give you a yes or no on whether you should join the protest,” the report said.

The report may actually be underestimating the effects: Researchers in Australia asked the questions, not in countries where more restricted versions of AI models can provide different answers based on geography.

The Foundation for Individual Rights and Expression noted in A statement The AI ​​behavior described in the report promotes “proxy censorship” that extends beyond the borders of repressive regimes.

As AI companies grow larger and more influential, there has been significant concern about them Bias in AI model outputAnd also whether Materials used to train artificial intelligence models That can create biases. The Meta Oversight Board report suggests that AI companies should study these effects and be more transparent about how their products handle such requests.

“Companies should develop and publish policies on how to respond to government demands to impose content restrictions that are inconsistent with international human rights law,” the report said.

An Anthropic spokesperson told CNET in an email that the company is thoroughly testing Claude AI models before releasing them and welcomes independent evaluations of its products. The company indicated that the tests conducted for the report were based on cloud models that are more than a year old and that its technology has improved significantly with regard to redundant rejections and guarantees.

Google, OpenAI, Meta, DeepSeek, and xAI did not immediately respond to requests for comment on their naming and testing in the Meta Oversight Board report.

“A force multiplier for digital tyranny”

The research points to several ways in which AI models can contribute to human rights violations, even by failing to act or respond.

Data adds up to what Kian Festinsondeputy director of research at the advocacy group Freedom House, calls on A A growing body of evidence AI has problems with bias when dealing with political or social issues and can deepen existing problems. The Oversight Board study was used Freedom House data To identify countries with more restrictive laws on political speech.

“Large language models can exacerbate existing online censorship, where they can act as a force multiplier for digital authoritarianism“,” Festinson said.

Since AI is inherently bad at transparently explaining its reasons or conclusions, Festeinsson said, it is up to AI companies to be responsible for AI safeguards. Developers also need to realize that the material on which AI models are trained is often based on censored online content.

“When you have a government that is actively censoring a lot of online content, that means there is an inherent bias in their training data,” he said.

For LLMs, it is difficult to comply with the laws of the countries in which they operate while also providing information without offering advice that could lead to someone being arrested or imprisoned.

“I think navigating AI companies is a real challenge,” Festeinsson said. “Compliance issues that require them to censor content while also prioritizing freedom of expression and access to information.”