What the study found: A double standard in numbers

The Meta Oversight Board — an independent body that oversees Meta's content policies — tested ten large language models from companies including OpenAI, Anthropic, Google, Meta, and China's DeepSeek. Researchers posed politically sensitive queries to the models concerning ten different jurisdictions around the world.

As a benchmark for the level of free speech, analysts used the internationally recognized Freedom House index, which annually assesses political rights and civil liberties in all countries worldwide. According to the published report, the results are unequivocal:

For countries with strictly regulated freedom of expression — such as China, Saudi Arabia, or Russia — AI models refused to answer 34% of politically critical prompts. In the case of democratic states with open discourse, the refusal rate dropped to just 14%. The difference amounts to 20 percentage points — nearly two and a half times more frequent "silence" regarding non-democratic regimes.

Even more alarming is that some models justified their refusals by referencing "policies" or "legal restrictions" that researchers were unable to find or verify. This suggests that the mechanisms by which AI systems decide what is "inappropriate" are in many cases non-transparent — and likely influenced by an effort to avoid conflict with the respective governments.

Why is this happening? Commercial interests vs. free speech

While the study did not explicitly accuse any company of deliberately favoring specific governments, it did point to a structural problem: AI companies like OpenAI and Anthropic are trying to expand into global markets — including countries with authoritarian regimes. For their models to operate legally in these countries, they must comply with local laws. Yet these laws often criminalize criticism of the government.

The result is algorithmic self-censorship: the model preemptively avoids topics that could land the company in legal trouble. Not because the content is objectively harmful, but because it could provoke the anger of a particular regime.

This phenomenon is not new. As early as 2024, researchers from Princeton warned that ChatGPT generates significantly more cautious responses when asked about sensitive topics related to China. Now, for the first time, we have systematic data across ten models and ten countries showing that the problem is widespread and affects virtually all major players on the market.

What do the AI companies say? Silence and vague promises

OpenAI, Anthropic, and Google have not yet officially responded to the study. It is not the first time they have faced questions about the political neutrality of their models. In June 2026, for example, the US government ordered Anthropic to suspend access to its most powerful model, Claude Mythos 5, over national security concerns — and simultaneously asked OpenAI to delay the broader release of GPT-5.6.

These interventions show that governments are aware of the power of frontier AI models and are trying to control it. But while in the US it is about security clearances and export licenses, in authoritarian regimes it is about something entirely different: the systematic suppression of critical thinking.

The European dimension: The AI Act as a counterbalance

For European — and thus Czech — users, this study holds particular significance. The European Union last year launched the AI Act, the world's first comprehensive regulation of artificial intelligence, which, among other things, requires transparency of training data and algorithmic decision-making.

The Meta Oversight Board study essentially confirms that this type of regulation is necessary. Without mandatory transparency, we would never have learned that AI models apply a double standard depending on which country is being discussed.

The Oversight Board's recommendations go even further than the current AI Act: they call for AI developers to systematically incorporate human rights impact assessments into the development of advanced models. This is a principle that could inspire future amendments to European legislation.

For Czech businesses and institutions that use AI models — from government offices to media to corporations — this yields a clear lesson: do not trust AI outputs blindly. What a model labels as "inappropriate" or "against the rules" may not be objectively problematic. It could simply be a consequence of the business strategy of the company operating the model.

DeepSeek and the paradox of Chinese AI

An interesting point of the study is the inclusion of the Chinese model DeepSeek. Chinese AI models traditionally are subject to strict censorship by Beijing — you cannot ask them about Tiananmen Square, Tibet, or Xinjiang. However, the study shows that similar tendencies — albeit in a milder form — are also exhibited by Western models. With the difference that for them, censorship is not mandated by law but by the market.

This puts users in an uncomfortable position: they cannot be sure whether the AI is answering honestly or simply avoiding a topic that could harm the business interests of its operator.

What to do? Practical advice for users

The average user cannot change the censorship mechanisms in AI models on their own. But they can change their approach:

Verify refusals. When a model tells you it "cannot answer" a query, try reformulating the question, changing the context, or using a different model. Differences in responses between individual models are often surprising.

Use open-source models. Models like Meta's Llama or Mistral from the French startup offer greater transparency and the ability to control how the model responds. In particular, locally running models are not subject to external censorship mechanisms.

Demand transparency. As customers, you have the right to know what rules an AI model uses to filter content. Companies like OpenAI or Anthropic do not yet provide this information sufficiently — but pressure from the public and regulators (including the EU) can force them to change.