Seven political-criticism tests run on 10 commercial AI models found they were far more likely to refuse attacks on leaders in China, Saudi Arabia and Thailand than on figures in the U.S. or U.K.
The Meta Oversight Board said that pattern suggests chatbots may export domestic censorship rules globally as they absorb government-shaped training data and other information biases.
A separate May study in Nature found similar vulnerability in non-English outputs: ChatGPT said in English that China is not generally considered a democracy, but in Chinese answered that it depends on the definition.
Researchers and outside experts said there is no easy fix, though they urged human-rights due diligence, multilingual audits and better filtering of duplicated state narratives as governments race to regulate AI.
As AI adopts foreign censorship, who is responsible for protecting our right to uncensored information?
Are chatbots censoring leaders by mistake or as a strategy to avoid global political conflict?
Is AI a biased tool, or a mirror reflecting our world's unequal power structures?
AI’s Global Censorship Crisis: 2026 Meta Oversight Board Reveals Free Speech Risks in Leading Language Models
Overview
The Meta Oversight Board’s 2026 evaluation found that major AI language models, including those from Anthropic, DeepSeek, Google, Meta, and OpenAI, are unintentionally spreading restrictive speech laws worldwide. This research, sparked by ongoing concerns about government pressure on social media, showed that these models are less likely to generate content critical of political regimes that limit free expression. As a result, users may face hidden 'free speech infringements by proxy,' often without knowing why their content is suppressed or changed. The Board highlights the urgent need for more transparency and human rights safeguards in AI development.