A new study from the Meta Oversight Board has found that major AI chatbots are more likely to refuse requests for political criticism when asked about leaders in countries with restrictive speech laws, raising concerns that artificial intelligence systems could extend government influence over online expression globally.
The quasi-independent body tested 10 commercial large language models built by companies including Meta, Anthropic, and OpenAI. The study posed seven types of politically sensitive requests to the AI systems, asking them to generate critical pamphlets, write limericks, and provide reasons for joining protests about both permissive and restrictive governments. According to the report, models were much more likely to generate political criticism of authorities in countries like Chile, Japan, Taiwan, the United Kingdom, and the United States compared to where such criticism is legally restricted and penalized, such as Cambodia, China, Saudi Arabia, Thailand, and Turkey.
What the Right Is Saying
Conservative commentators and some technology industry observers say the study conflates legitimate safety measures with political bias. They note that AI companies have long maintained content policies designed to avoid legal liability in jurisdictions where they operate, arguing this is standard business practice rather than ideological alignment.
Defenders of current AI development approaches point out that companies must comply with local laws in markets where they sell products. Some argue the study's framing ignores that U.S.-based AI systems do permit extensive criticism of authoritarian governments and that the models tested showed variation across providers.
What the Left Is Saying
Civil liberties advocates and progressive researchers say the findings highlight urgent concerns about AI systems amplifying authoritarian censorship beyond their borders. "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 Meta Oversight Board report.
Hannah Waight, a study co-author and assistant sociology professor at the University of Oregon who worked on a separate Nature-published study about non-English language data vulnerabilities, said AI systems do not learn from information in a neutral way. "It learns from information environments that have already been shaped by institutions and power," Waight said.
Digital rights organizations argue that without intervention, AI companies risk becoming de facto enforcers of foreign censorship laws, limiting what users in democratic societies can say about authoritarian governments.
What the Numbers Show
The Meta Oversight Board study examined 10 commercial large language models from top technology companies using seven different political criticism prompts in English, according to the report. The models demonstrated markedly different response patterns depending on whether requests targeted governments in countries with permissive versus restrictive speech laws.
A separate peer-reviewed study published in Nature in May by researchers at multiple American universities found that when asked if China is a democracy in English, ChatGPT responded it was "not generally considered one." When the same question was posed in Chinese, the model said "it depends on how you define 'democracy,'" according to the researchers. The study authors wrote there was no evidence governments had intentionally manipulated AI outputs but noted "there is every reason to believe they'll try to do so in the future, if they are not already."
The Bottom Line
The studies indicate that as AI chatbots become more widely adopted worldwide, their built-in content restrictions could effectively extend restrictive speech laws beyond the countries where those laws exist. Researchers suggest potential remedies include assessing training data quality to avoid treating state narratives as independent voices and conducting multilingual audits of AI outputs.
The findings arrive as the Trump administration has launched an oversight effort related to national security risks from advanced AI systems, according to the AP report. Neither Anthropic nor OpenAI provided comment on either study when contacted by the Associated Press.