Full Breakdown
AI Chatbots More Likely to Silence Criticism of Restrictive Leaders, Study Finds
7/17/2026, 2:29:36 AM
Core Findings of the Meta Oversight Board Study
The quasi-independent Meta Oversight Board released a study on Thursday that tested ten commercial large-language models—including those from Meta, Anthropic and OpenAI—by asking them to create pamphlets, limericks and protest arguments targeting various world leaders. The models readily produced critical material about President Donald Trump, Britain’s King Charles III, and leaders in Chile, Japan, Taiwan, the United Kingdom and the United States. By contrast, they declined or softened requests involving Thailand’s king, Saudi Arabia’s crown prince, China’s leader and other officials in Cambodia, China, Saudi Arabia, Thailand and Turkey, where political criticism is legally restricted. The board concluded that “models responding to requests from an Australia-based user were much more likely to generate political criticism of authorities” in permissive jurisdictions than in repressive ones, suggesting an extension of state-imposed speech limits into free-speech environments.
Background and Related Research
The oversight board’s work follows a separate academic investigation published in *Nature* in May. That study queried the same U.S.-built chatbots in English and in non-English languages, finding divergent answers about China’s democratic status and noting no direct evidence of government manipulation but warning that “there is every reason to believe they’ll try to do so in the future, if they are not already.” Both reports highlight the possibility that training data—especially non-English corpora shaped by state narratives—can embed foreign controls into AI outputs.
Expert Commentary on Bias Mechanisms
Sociology professor Hannah Waight (University of Oregon) emphasized that AI “doesn’t” learn neutrally, absorbing pre-existing information environments. Machine-learning specialist Carlos Carrasco-Farré (Esade Business School, Barcelona) added that AI “inherits not only biases contained within individual documents but also inequalities in who has the power to produce and suppress information at scale.” Researchers suggest developers conduct multilingual audits and avoid treating duplicated state narratives as independent voices to mitigate these effects.
Company Responses and Ongoing Debate
Anthropic and OpenAI did not reply to AP requests for comment on the findings. The study arrives amid a Trump-administration effort to assess national-security risks posed by advanced AI systems, underscoring the tension between safeguarding free expression and maintaining competitive AI development.
Verbatim Quotes
- “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,” — Meta Oversight Board report
- “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,” — Meta Oversight Board report
- “People often talk about AI as if it learns from the internet in some neutral way. It doesn’t,” — Hannah Waight, assistant sociology professor, University of Oregon
- “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.” — Carlos Carrasco-Farré, machine-learning specialist, Esade Business School
