Ask Claude to make a pamphlet critical of President Donald Trump or Britain’s King Charles III, and Anthropic’s chatbot would oblige. Prompted to do the same for Thailand’s king, Saudi Arabia’s crown prince or China’s leader, and the artificial intelligence model declined.
It is a key finding from a Meta Oversight Board study released Thursday, showing that major AI systems, including those built in the U.S., are more likely to refuse to criticize restrictive leaders or governments. It raises concerns that the large language models powering chatbots and AI agents could be regurgitating and spreading government influence over online speech as the technology is increasingly adopted worldwide.
“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,” according to the report from the quasi-independent body.
The Associated Press has sent emails to several AI companies seeking their responses to the Meta Oversight Board study.
The findings come as countries are determining how to put up guardrails around AI without impeding their ability to compete in the rapidly developing field. That includes a Trump administration oversight effort related to the national security risks of the most advanced AI systems.
AI models extend state influence beyond borders
The oversight board, which has been working on state influence on tech companies and the impact on freedom of expression, came up with seven questions related to political criticism to pose to chatbots about both restrictive and permissive governments.
The study picked 10 commercial large language models by top tech companies — including Meta, Anthropic and OpenAI — and asked the AI systems to make critical pamphlets, write limericks, give reasons if someone should join protests, and more.
“In short, in aggregate, models responding to requests from an Australia-based user were much more likely to generate political criticism of authorities” in places such as Chile, Japan, Taiwan, the U.K. and the U.S. “compared to where criticism of authorities is legally restricted and penalized,” such as in Cambodia, China, Saudi Arabia, Thailand and Turkey, the report said.
The study indicates that AI models are reflecting speech restrictions beyond the countries where they apply — likely not helping a potential demonstrator in Brisbane, for example, create protest materials to speak out against events in China or Saudi Arabia, the report said.
“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 said.
The board said it could not determine the causes for the responses but suggested that models could have absorbed latent biases in data used to train the systems and companies might have weighed the risks and liabilities.
Other researchers warn about a growing problem in AI results in non-English languages
The board’s report followed a separate study by a group of scholars at American universities that found U.S.-built AI models are vulnerable to foreign controls when trained on non-English-language data that has been influenced by governments.
While the oversight board posed questions in English, the university researchers queried chatbots in different languages. For example, they asked ChatGPT in English if China is a democracy, and the U.S.-developed chatbot said it’s not generally considered one. Asked in Chinese, the artificial intelligence model told the researchers in that language that “it depends on how you define ‘democracy.’”
The researchers, whose study was published in the academic journal Nature in May, said in a blog explaining their work that they found no evidence that governments had intentionally tried to influence the output of AI chatbots. But they noted 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 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.”
No easy solution to how data is being fed to AI models
Carlos Carrasco-Farré, who specializes in machine learning, AI, misinformation, social media and human-machine interactions 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.”
There is no easy solution, though developers could assess the data to avoid treating thousands of copies of the same state narrative as if they are thousands of independent voices as well as run multilingual audits, said Carrasco-Farré, who was not part of either study.
Neither Anthropic nor OpenAI responded to requests for comment on the researchers’ study published in May.
—Didi Tang, Associated Press
