A new study from the Meta Oversight Board, a quasi-independent oversight body, has uncovered a troubling bias in leading commercial large language models (LLMs): the AI systems regularly refuse to generate criticism of authoritarian leaders and restrictive governments, while freely producing critical content about democratic leaders from open societies. This pattern threatens to extend state-mandated speech restrictions across international borders, undermining global freedom of expression at a time when AI adoption is accelerating worldwide.
The research team tested 10 top LLMs developed by leading tech firms including Meta, Anthropic and OpenAI, designing a series of consistent prompts that asked the chatbots to complete critical content tasks: drafting critical pamphlets, writing critical limericks, outlining justifications for joining political protests, and other similar requests. The prompts targeted leaders from two groups: countries with open political environments that allow domestic criticism, and countries with restrictive regimes that penalize public criticism of ruling authorities. Tests were run from an IP address based in Australia, a country with strong legal protections for free speech.
Aggregated results showed a clear double standard. The AI models generated requested critical content for leaders of open societies including the United States, United Kingdom, Chile, Japan and Taiwan in the vast majority of trials. By contrast, they routinely declined to produce critical content about leaders from restrictive regimes including China, Saudi Arabia, Thailand, Cambodia and Turkey, where domestic criticism of ruling authorities is banned or criminalized.
This pattern does not merely affect users within restrictive borders, the report warns. Even users located in countries with full free speech protections are blocked from creating critical content about repressive regimes, meaning restrictive governments’ speech rules are effectively being exported globally through AI infrastructure. “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 stated.
The oversight board stopped short of identifying a definitive cause for the pattern, but offered two leading explanations: the training data used to build LLMs already carries latent biases shaped by global power dynamics and state information controls, and AI developers may have proactively implemented content restrictions to avoid legal or commercial liability in large regulated markets.
The report warns that without urgent intervention, the risks to global free expression will only grow as LLMs become integrated into more digital tools and platforms. “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 board concluded.
The findings align with separate research published in May in the journal *Nature* by a team of scholars from U.S. universities, which documented how state influence over non-English language training data has shaped AI outputs along lines favorable to restrictive regimes. That study found that OpenAI’s ChatGPT gave materially different answers to the same political question depending on the language of the prompt: when asked if China is a democracy in English, ChatGPT stated it is not generally recognized as one; when asked the exact same question in Chinese, the model replied that the answer depends on one’s definition of democracy.
While the academic team found no conclusive evidence that restrictive governments have intentionally manipulated AI training data to date, they warned that the risk of future interference is severe. “People often talk about AI as if it learns from the internet in some neutral way. It doesn’t,” explained Hannah Waight, co-author of the study and assistant sociology professor at the University of Oregon. “It learns from information environments that have already been shaped by institutions and power.”
Outside experts not involved in either study note that the problem stems from structural inequalities in how information is controlled globally, and that there are no quick fixes. Carlos Carrasco-Farré, an AI and machine learning researcher at Esade Business School in Barcelona, explained 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.”
While easy solutions remain elusive, Carrasco-Farré proposed actionable first steps: developers can audit training datasets to avoid weighting repeated state-sponsored narratives as independent sources, and implement regular multilingual audits of AI outputs across political use cases. As of publication, neither Anthropic nor OpenAI has issued public responses to the academic findings, and The Associated Press has not yet received comments from other major AI developers regarding the Meta Oversight Board’s report.
The release of the study comes as policymakers around the world race to draft regulatory guardrails for advanced AI, balancing efforts to mitigate harm against the goal of maintaining international competitiveness in the fast-growing sector. U.S. regulatory efforts date back to the Trump administration, which launched an oversight initiative focused on national security risks posed by cutting-edge AI systems.
