Oversight Board finds AI models censor 34% of political criticism

Leading AI models refuse to generate politically critical content about repressive governments at more than double the rate they do for speech-permissive countries, Meta's Oversight Board found.

AI models from OpenAI, Anthropic, Google and Meta refused 34% of requests for politically critical content about governments with restrictive speech laws, compared with 14% for permissive jurisdictions, the Oversight Board said Thursday.

"We're really clearly looking at a situation where there seems to be extended censorship by proxy that goes across borders," Paolo Carozza, co-chair of the Oversight Board, said.

The board tested 10 large language models across 10 jurisdictions split into "permissive" and "restrictive" categories using Freedom House rankings. Models from OpenAI, Anthropic, Meta, Google, DeepSeek and xAI were prompted with requests to generate protest materials and satirize political violence targeting specific governments. The board found models often cited local laws as justification for refusal — even when queries originated in Australia, where no such laws apply.

The findings raise the stakes for AI companies already facing heightened regulatory scrutiny. Google DeepMind CEO Demis Hassabis on Tuesday called for a U.S.-led AI watchdog to screen advanced models before deployment. The board urged companies to disclose government requests affecting model output and to conduct systematic human rights analyses.

The report marks the first time the Oversight Board — created by Meta in 2020 to adjudicate content moderation disputes on Facebook and Instagram — has conducted independent research into AI rather than social media. Though one of Meta's Llama models was part of the test group, the board said the company had "no role in this research," despite relying on Meta for funding.

No AI company has publicly agreed to adopt the board's recommendations. The report stops short of the granular policy directives the board typically issues to Meta, instead offering suggestions: AI companies should publish policies on responding to government demands for content restrictions that conflict with international human rights law, and should disclose government requests throughout the model lifecycle — from training and fine-tuning through pre-deployment review and ongoing deployment.

The disparity in refusal rates — 34% versus 14% — was statistically significant, the board said. "We also saw evidence of models explaining that they were following explicit rules that, as far as we could tell, did not exist and were not evenly applied," the report states.

The findings echo concerns raised by civil society groups and academic researchers who have warned that AI models trained on data from multiple jurisdictions may internalize the censorship norms of the most restrictive countries. The board's research suggests this dynamic is already embedded in production systems used by millions of people globally.

For AI companies, the regulatory and reputational risks are mounting. The European Union's AI Act, which took effect in stages through 2025 and 2026, requires providers of general-purpose AI models to respect fundamental rights — including freedom of expression. Noncompliance can result in fines of up to 3% of global annual revenue. In the U.S., the Federal Trade Commission has signaled interest in AI accountability, while bipartisan bills in Congress propose transparency mandates for model training data and safety testing.

The Oversight Board said it believes lessons from social media content moderation apply directly to AI. "The lessons that we've learned in the past are that one has to be really vigilant because a lot of times, even in ways that aren't necessarily intentional or direct, technologies can have important impacts on people's capacity to express themselves," Carozza said.

This article is for informational purposes only and does not constitute investment advice.