AI chatbots may be unintentionally extending the influence of government censorship beyond national borders. According to the report, major AI systems—including models developed by U.S. companies such as OpenAI, Anthropic, and Meta—were significantly more willing to generate criticism of leaders in democratic countries than of leaders in countries with restrictive speech laws. For example, some models readily produced criticism of U.S. President Donald Trump or the U.K.'s King Charles III but refused similar requests involving leaders in countries such as China, Saudi Arabia, or Thailand. The findings raise concerns that AI systems may be reproducing restrictions embedded in their training data rather than applying a consistent standard for free expression.
The researchers tested 10 commercial large language models using a series of politically related prompts, including requests to write critical pamphlets, poems, and protest materials concerning governments with varying levels of political freedom. Overall, the models were much more likely to comply when asked to criticize authorities in democratic nations than when asked about governments where political criticism is legally restricted. The report argues that this pattern could effectively extend the influence of restrictive governments into countries with stronger protections for free speech, even when users are located outside those jurisdictions.
The article also references complementary academic research showing that AI responses can vary depending on the language used. In one example, a chatbot clearly stated in English that China is generally not considered a democracy, but gave a more qualified answer when asked the same question in Chinese. Researchers stress that this does not necessarily indicate intentional manipulation by AI companies. Instead, they argue that language models learn from information ecosystems that already reflect political, cultural, and institutional influences, making them susceptible to reproducing those patterns unless developers actively test for and mitigate them.
The report concludes that AI developers should strengthen human rights due diligence, conduct multilingual bias testing, diversify training data, and improve transparency about how politically sensitive content is handled. As AI chatbots increasingly become tools for accessing information, creating content, and participating in public discourse, the study argues that ensuring consistent treatment of political expression across countries and languages will be an important challenge for both AI companies and policymakers. The goal is not to eliminate safety guardrails, but to ensure they are applied fairly and do not unintentionally amplify government restrictions on freedom of expression.
