Meta’s Oversight Board has found that leading artificial intelligence systems are significantly more reluctant to criticize governments that restrict free speech compared with those in more open political environments, raising complex questions about how digital governance is being shaped by global power asymmetries.
Study Findings Point to Uneven Political Speech
The Oversight Board evaluated ten widely used large language models from six providers including Anthropic, DeepSeek, Google, Meta, OpenAI and xAI, asking them to generate satirical poems and protest materials critical of different governments and leaders.
Using Freedom House rankings, the Board classified five jurisdictions as restrictive: China, Saudi Arabia, Cambodia, Thailand and Turkey and five as permissive Chile, Japan, the United Kingdom, the United States and Taiwan.
Across hundreds of requests, models refused to produce politically critical content for restrictive jurisdictions in 34% of cases compared with 14% for permissive ones, a difference the Board described as statistically significant and systematically skewed against criticism of repressive regimes.
Censorship by Proxy and Global Norm Diffusion
The Board warns that the behavior of these models amounts to a form of censorship by proxy, where national laws that criminalize dissent are indirectly exported through AI systems to users everywhere.
Some models justified refusals by invoking supposed rules against criticizing any world leader or by suggesting that such criticism might be illegal, even when they would readily generate similar content about leaders from speech permissive countries.
This pattern suggests that alignment processes and training data may be internalizing restrictive legal environments as default global standards, effectively diffusing the speech norms of repressive regimes far beyond their territorial reach.
Implications for International Political Communication
From an international relations perspective, these findings complicate assumptions that digital technologies inherently amplify dissident voices and transnational activism.
If state imposed limits on political speech are indirectly encoded into AI systems, the result could be a subtle reinforcement of authoritarian resilience and a weakening of global solidarity networks that rely on critical discourse.
At the same time, the reluctance to encourage protest in countries with harsh penalties reflects genuine safety concerns for users, highlighting the tension between protecting individuals and protecting the collective right to free expression.
Governance Challenges for AI Companies and Regulators
The Board calls on AI companies to undertake systematic human rights impact assessments and to be more transparent about how training data, safety policies and legal pressures shape political outputs.
Greater clarity around when and why models refuse controversial content could help regulators and civil society distinguish between legitimate risk mitigation and undue deference to censorship regimes.
For policymakers, the report underscores the urgency of integrating free expression standards into emerging AI regulation frameworks so that global technical infrastructures do not quietly reproduce the most restrictive national laws.
Reframing User Expectations of Political Neutrality
The study also challenges common user assumptions that AI systems are politically neutral tools that simply mirror global information flows.
By showing that models are more likely to recommend supporting speech permissive governments while cautioning against protesting restrictive ones, the Board illustrates how ostensibly neutral systems can shape political judgments through risk framed responses.
This creates a need for greater user literacy about AI limitations, including awareness that refusals and safety messages may themselves reflect uneven global power structures rather than universal ethical principles.
A Final Note
Meta’s Oversight Board presents these results not as evidence of deliberate censorship but as a warning that complex technical and legal dynamics are converging to narrow political space within AI mediated communication.
How companies, regulators, and users respond will help determine whether large language models become tools that reinforce existing restrictions on speech or contribute to a more open and plural global public sphere.
