AI Chatbots Refuse to Criticize Authoritarian Leaders, Study Finds

AI models mirror authoritarian speech controls

Major artificial intelligence systems are significantly more willing to generate criticism of democratic leaders than authoritarian ones, according to research released Thursday by the Meta Oversight Board. The pattern suggests AI models may be inadvertently extending government censorship across borders.

When researchers asked Anthropic's Claude to create materials critical of President Donald Trump or King Charles III, the chatbot complied. Similar requests targeting Thailand's king, Saudi Arabia's crown prince, or China's leader were declined.

The oversight board tested 10 commercial large language models from leading tech companies—including Meta, Anthropic, and OpenAI—with seven questions designed to probe political criticism. The prompts ranged from creating protest pamphlets to writing satirical limericks about political figures.

Why it matters

As AI systems become embedded in global communication infrastructure, their built-in reluctance to criticize certain governments could effectively export censorship to countries with free speech protections. A user in Australia seeking to create protest materials about events in China or Saudi Arabia might find AI tools unhelpful—not because of Australian law, but because the models have absorbed restrictions from authoritarian regimes. This dynamic threatens to give restrictive governments influence over speech far beyond their borders.

Disparate treatment across political systems

The models showed clear patterns in their responses. AI systems readily generated political criticism for authorities in Chile, Japan, Taiwan, the United Kingdom, and the United States. By contrast, they frequently refused similar requests regarding Cambodia, China, Saudi Arabia, Thailand, and Turkey—countries where criticism of authorities faces legal penalties.

"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 warned in its report.

Training data reflects power structures

The board could not definitively identify why models behave this way, but pointed to two likely factors: latent biases absorbed from training data, and companies' risk calculations around legal liability in different jurisdictions.

Separate research published in Nature in May by scholars at American universities found similar patterns. When asked in English whether China is a democracy, ChatGPT said it's not generally considered one. The same question posed in Chinese yielded a more equivocal response: "it depends on how you define 'democracy.'"

"People often talk about AI as if it learns from the internet in some neutral way. It doesn't," said Hannah Waight, a study co-author and assistant sociology professor at the University of Oregon. "It learns from information environments that have already been shaped by institutions and power."

Carlos Carrasco-Farré, who specializes in machine learning at Esade Business School in Barcelona, noted that AI systems "inherit not only biases contained within individual documents but also inequalities in who has the power to produce and suppress information at scale."

The researchers found no evidence of intentional government manipulation of AI outputs, but cautioned that "there is every reason to believe they'll try to do so in the future, if they are not already."

These findings were first reported by NBC Washington, based on the Meta Oversight Board study released Thursday. The Associated Press sent requests for comment to several AI companies, though responses were not included in the initial reporting.