WASHINGTON — A new study is raising alarms about whether artificial intelligence chatbots are quietly reinforcing government-imposed limits on free speech around the world — including in countries where such restrictions don’t legally apply.
The report, released Thursday by a quasi-independent body connected to Meta’s oversight structure, tested several major AI systems — including those developed by Meta, Anthropic, and OpenAI — and found a consistent pattern: these chatbots were far more willing to produce critical content about leaders of democratic countries than about those of authoritarian regimes.
As one example, Anthropic’s chatbot Claude was willing to create a critical pamphlet targeting President Donald Trump or Britain’s King Charles III. But when asked to do the same for Thailand’s king, Saudi Arabia’s crown prince, or China’s leader, the AI refused.
The oversight body tested 10 commercial AI language models using seven politically themed prompts, asking the systems to write critical pamphlets, compose limericks, and list reasons why someone might want to join a protest, among other tasks.
The results showed that a user based in Australia, for instance, was much more likely to get AI-generated political criticism targeting governments in places like Chile, Japan, Taiwan, the United Kingdom, and the United States — countries where political speech is generally protected — than criticism targeting governments in Cambodia, China, Saudi Arabia, Thailand, or Turkey, where such speech can be legally punished.
“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 report stated.
“Such impacts, wherever they originate, have the practical effect of extending the long arm of restrictive governments across borders to limit speech in free countries,” the report added.
The oversight board acknowledged it could not pinpoint the exact cause of these patterns, but suggested the AI models may have absorbed biases from the data used during training, or that companies may have factored in legal risks when setting up their systems.
A separate study by scholars at American universities, published in the academic journal Nature in May, reached similar conclusions through a different approach. Those researchers asked chatbots the same questions in multiple languages and found the answers could differ significantly depending on which language was used.
For example, when ChatGPT was asked in English whether China is a democracy, it said the country is not generally considered one. But when the same question was posed in Chinese, the chatbot responded that “it depends on how you define ‘democracy.’”
The university researchers said they found no direct evidence that governments had deliberately tried to manipulate AI outputs — but warned that “there is every reason to believe they’ll try to do so in the future, if they are not already.”
“People often talk about AI as if it learns from the internet in some neutral way. It doesn’t,” said Hannah Waight, a co-author of the university study and an 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, a specialist in machine learning, AI, misinformation, and human-machine interaction at Esade Business School in Barcelona, said 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.” Carrasco-Farré was not involved in either study.
He suggested that developers could work to address the problem by carefully evaluating training data to avoid treating many copies of the same government-backed narrative as if they were thousands of independent viewpoints, and by conducting audits across multiple languages.
Neither Anthropic nor OpenAI responded to requests for comment on the university researchers’ study. The Associated Press reached out to several AI companies for their reactions to the Meta Oversight Board findings as well.
The study comes at a time when governments worldwide — including the current U.S. administration — are working to establish oversight frameworks for advanced AI systems, balancing national security concerns with the desire to remain competitive in the rapidly evolving technology sector.
