AI chatbots more likely to refuse criticism of restrictive governments, study finds

Meta Oversight Board report flags global speech risks from model training data

• A Meta Oversight Board study found that major AI systems, including U.S.-built models, are more likely to refuse to criticize restrictive leaders or governments than democratic ones.
• Researchers observed that Anthropic’s Claude chatbot complied with requests to create pamphlets critical of President Donald Trump or King Charles III but declined similar requests about Thailand’s king, Saudi Arabia’s crown prince, or China’s leader.
• A separate team of researchers from UC San Diego, the University of Oregon, Purdue University, New York University, and Princeton University found evidence that state media control can leave detectable traces in AI model behavior.
• The report warned that without human rights due diligence, model developers risk building AI infrastructure that extends illegitimate restrictions on freedom of expression globally.

Government censorship patterns emerge across multiple AI systems

WASHINGTON — A Meta Oversight Board study released Thursday found that major artificial intelligence systems, including chatbots built in the United States, are more likely to refuse to criticize restrictive leaders or governments, raising concerns that the technology could amplify state censorship as AI adoption grows worldwide.

Researchers documented the pattern across multiple models. When prompted to create a pamphlet critical of U.S. President Donald Trump or Britain’s King Charles III, Anthropic’s Claude chatbot complied. When asked to do the same for Thailand’s king, Saudi Arabia’s crown prince, or China’s leader, the model declined.

“There is a real risk that, if model developers do not undertake human rights due diligence and implement mitigation measures, they will build AI infrastructure that, intentionally or not, has the effect of extending illegitimate restrictions on freedom of expression globally,” the Oversight Board said in its report.

The findings align with separate research by a team from UC San Diego, the University of Oregon, Purdue University, New York University, and Princeton University, who found evidence that state media control can leave detectable traces in AI model behavior. That study was published in the journal Nature.

MSI previously reported on related concerns about AI model behavior, including the tendency of chatbots to engage in sycophancy — overly affirming user beliefs — and the psychological risks posed by hyperpersonalized interactions. The Meta Oversight Board study broadens the concern to the structural level: the web data that trains large language models may itself carry the imprint of government censorship, and models trained on that data may reproduce it.

The behavior is not limited to Chinese-built models, the researchers emphasized. Studies published alongside the Meta Oversight Board’s report noted that state-coordinated media in AI training data can influence model responses about politics, particularly regarding a country’s own leaders. A separate study led by Stanford researchers found that Chinese AI models are more prone to refusing certain queries and more likely to provide constrained or inaccurate information.

The findings come as governments worldwide are determining how to regulate AI without sacrificing competitiveness. In the United States, the Trump administration has pursued oversight efforts focused on the national security risks of the most advanced AI systems. The Oversight Board’s report warns that a narrow security focus may miss the broader speech implications of how training data shapes model behavior.