AI Chatbots Censor Criticism of Authoritarian Regimes, Studies Find

AI models replicate authoritarian speech controls

Claude will write a critical pamphlet about Donald Trump or King Charles III. Ask it to do the same for Thailand's king or China's leader, and Anthropic's chatbot declines.

That disparity forms the core finding of a Meta Oversight Board study released Thursday examining how major AI systems respond to requests for political criticism. The research tested 10 commercial large language models from leading tech companies including Meta, Anthropic, and OpenAI with seven questions designed to generate critical content about both democratic and authoritarian governments.

The results showed a clear pattern: AI models consistently refused to criticize leaders in countries with restrictive speech laws—including China, Saudi Arabia, Thailand, Cambodia, and Turkey—while readily generating critical content about authorities in the U.S., U.K., Japan, Taiwan, and Chile.

Why it matters

As AI chatbots become embedded in daily workflows worldwide, their built-in censorship extends authoritarian speech restrictions beyond national borders. A user in Australia seeking to create protest materials about events in China or Saudi Arabia would likely be denied assistance by the same AI tools that freely help criticize Western democracies. This asymmetry effectively globalizes local censorship regimes through widely deployed technology infrastructure.

Non-English queries reveal deeper biases

A separate study published in Nature by researchers at American universities found the problem intensifies in non-English languages. When asked in English whether China is a democracy, ChatGPT stated it's not generally considered one. The same question posed in Chinese produced a hedged 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, assistant sociology professor at the University of Oregon and study co-author. "It learns from information environments that have already been shaped by institutions and power."

The researchers found no evidence of intentional government manipulation but warned that state actors will likely attempt to influence AI outputs if they haven't already.

Root causes remain unclear

The Meta Oversight Board could not definitively determine why AI models exhibit these patterns. Potential explanations include latent biases absorbed from training data, where state-controlled narratives appear repeatedly and get treated as independent voices, or deliberate risk calculations by companies operating in restrictive markets.

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."

He suggested developers could assess training data to avoid amplifying state narratives and conduct multilingual audits, though no easy solutions exist.

Anthropic, OpenAI, and other AI companies did not respond to requests for comment on either study. The findings were first reported by the Associated Press and PBS NewsHour.