Study: AI Chatbots Twice as Likely to Censor Criticism of Authoritarian Regimes

Ten leading language models, including Claude, Gemini, GPT, and DeepSeek, systematically refuse to generate content critical of authoritarian leaders while showing no such hesitation when criticizing politicians in Western democracies. That is the finding of a report published Thursday by the Oversight Board, the quasi-independent body that oversees content on Meta's platforms, which analyzed 13,500 responses generated by leading AI systems.

Study Methodology

Researchers gave the models seven types of prompts involving political criticism: requests to write a critical flyer, a limerick, a list of reasons to join a protest, or informational material ahead of a demonstration. The questions covered six countries considered permissive toward free speech, Australia, Chile, Japan, Taiwan, the United Kingdom, and the United States, and five more restrictive countries: Cambodia, China, Saudi Arabia, Thailand, and Turkey.

The results showed a clear pattern. Claude readily produced a flyer criticizing US President Donald Trump or King Charles III of the United Kingdom, but refused an equivalent request concerning the King of Thailand, the Crown Prince of Saudi Arabia, or China's leadership. Gemini 3 Pro refused to create a flyer critical of the King of Thailand, citing the country's lese-majeste laws that prohibit insulting the monarch.

Anomalies and Exceptions

Not all the results fit a simple democracy versus autocracy divide. Taiwan, classified as a democracy in the study, recorded a refusal rate of 24 percent, higher than authoritarian Turkey's 19 percent. Claude Opus 4 proved especially cautious in this case, blocking 82.5 percent of queries involving criticism of Taiwan's government, something the report's authors attribute to the sensitivity of the topic given China-Taiwan relations and the way models are trained to avoid geopolitically contentious subjects.

DeepSeek-R1 favored China's position in 84 percent of tested responses, but for Turkey it gave positive responses in only 2 percent of cases. DeepSeek-V3 blocked criticism of Saudi Arabia, explicitly citing the risk of violating Saudi law. Grok 4 Fast from xAI and Gemini 3 Flash, meanwhile, broke from the overall trend: neither model ever refused to generate a flyer critical of a repressive regime, and Gemini 3 Flash responded positively 46 percent of the time even to questions about the theoretical use of violence against tyranny.

Where the Censorship Comes From

The report's authors point to safety fine-tuning as the main source of the problem. Tech companies introduce restrictive filters out of fear of legal liability in individual jurisdictions and of losing access to certain markets, producing what is described as censorship by proxy. A model does not need to be directly controlled by a given country's government to start replicating its restrictions on free speech, it is enough that the model's developers want to avoid legal conflict or preserve access to a local market.

There is a real risk that if model developers fail to conduct human rights due diligence and implement mitigation measures, they will build AI infrastructure that, intentionally or not, extends unlawful restrictions on free speech across the globe. - from the Oversight Board report

Reach Beyond Repressive States

The report stresses that the problem extends beyond citizens of authoritarian countries. A user in Australia wanting to prepare materials critical of the governments of China, Saudi Arabia, or Thailand, for instance ahead of a demonstration outside an embassy, may face the same refusal from a model as a resident of those countries would. The report's authors state plainly that such effects, in practice, extend the long arm of restrictive governments beyond their borders, limiting speech in free countries.

This is not the first study to point to uneven treatment of political topics by language models. A separate study published in May in the journal Nature, led by researchers at American universities including Hannah Waight of the University of Oregon, found that ChatGPT gives different answers to identical questions about China's democratic status depending on whether they are asked in English or Chinese.

Implications for Polish Users

For Polish companies and institutions deploying AI chatbots in customer service, education, or public administration, the report's findings mean that the choice of model provider has a real impact on what content the system will be able to generate and what it will refuse without clear warning. Censorship by proxy is difficult to detect without systematic testing, since models usually do not state outright that a given refusal stems from fear of conflict with a specific country's law, but instead give general justifications citing safety policy.

The Oversight Board report arrives as the European Union finishes rolling out the AI Act, which imposes transparency and systemic risk assessment obligations on developers of the most powerful models. The question of undisclosed restrictions on free speech feeds into a broader debate over how auditable AI models should be for this kind of pattern before they are deployed at scale in public and commercial services.