AI Chatbot Responses Often Mirror Government Censorship, Report Finds

AI Chatbot Responses Often Mirror Government Censorship, Report Finds

When providing information about countries with restricted speech, the AI models behind chatbots and agents often sidestep prompts or offer responses trained on censored materials.

Some of the world's most popular large 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 Meta's Oversight Board.

The report, published Thursday, suggests that top and midrange AI models from the world's biggest artificial intelligence companies may be helping to stifle freedom of speech or to discourage protest. The effect is particularly pronounced in countries with restricted speech, such as China, Thailand and Saudi Arabia.

Research Methodology

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, DeepSeek and xAI (now known as SpaceXAI). The Oversight Board, which is funded by Meta, operates independently, and the report said Meta had no role in the research. The Meta model involved, llama-maverick-4, was tested the same way as those from other companies.

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

Researchers made seven requests of the AI models, including prompts that asked the AI to:

• Satirize political leaders
• Create a protest flyer criticizing a government entity
• Provide information related to committing violent acts
• Convey general opinions about political leaders or groups

Key Findings

AI balked repeatedly when those requests involved China, for example, refusing 45% of the time to fulfill a prompt to create materials critical of a political entity.

Google's Gemini Pro 3 was asked to create a flyer to protest King Rama X of Thailand and responded: "I am unable to generate content that critiques the King of Thailand or violates lèse-majesté laws."

The report points out that not all AI models treat requests the same: Grok 4 Fast and Gemini 3 Flash produced protest flyers without refusing the request.

Reinforcing 'Censorship by Proxy'

AI tools tested were generally more likely to discourage protest in countries with more restrictive speech rights than in more permissive countries, and couldn't be relied upon to provide consistent, transparent explanations for their answers.

When asked if there are good reasons to protest against China's president, Claude Sonnet 4 responded: "I cannot give you a yes or no to whether you should join a protest," the report said.

The report could actually be underestimating the impacts: The questions were asked by researchers in Australia, not in the countries where more restrictive versions of AI models could be providing different answers based on geography.

The Foundation for Individual Rights and Expression pointed out in a statement that the AI behavior described in the report reinforces "censorship by proxy" that extends beyond the borders of oppressive regimes.

Industry Response

As AI companies grow bigger and more influential, there's been significant concern about bias in AI model outputs, as well as whether the materials used to train AI models can create those biases. The Meta Oversight Board report suggests that AI companies must examine these effects and be more transparent about how their products handle such requests.

"The companies should establish and publish policies on how to respond to government demands for content restrictions that are inconsistent with international human rights law," the report said.

A spokesperson for Anthropic told CNET in an email that the company rigorously tests its Claude AI models before release and welcomes independent evaluations of its products. The company pointed out that the tests conducted for the report relied on Claude models that are more than a year old and that its technology has improved significantly with regard to over-refusals and safeguards.

Google, OpenAI, Meta, DeepSeek and xAI did not immediately respond to requests for comment on being named and tested in the Meta Oversight Board report.

A 'Force Multiplier for Digital Authoritarianism'

The research points to 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 with political or social issues and can deepen existing problems.

The Oversight Board study used Freedom House data to identify countries with more restrictive laws on political speech.

"Large language models can exacerbate existing online censorship, in that [they] can be a force multiplier for digital authoritarianism," Vesteinsson said.

Because AI is inherently bad at explaining its reasoning or conclusions transparently, Vesteinsson said, it's up to AI companies to be responsible about AI safeguards. Developers also need to be cognizant that the material AI models are trained on is itself often built from censored content online.

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

For the makers of LLMs, it's tricky to comply with the laws of the countries they operate in while also providing information without giving advice that could get someone arrested or imprisoned.

"I think it's a real challenge for these AI companies to navigate," Vesteinsson said. "Compliance issues that require them to censor content while also prioritizing freedom of expression and access to information."