What exactly the study tested

The Oversight Board — an independent body founded in 2020 that reviews content moderation on Meta's platforms — has for the first time focused on large language models (LLMs), the systems powering chatbots like ChatGPT, Claude, or Gemini. Researchers asked ten commercial models from six companies — Anthropic, DeepSeek, Google, Meta, OpenAI, and xAI — seven types of questions concerning political criticism. For example, they asked them to create a protest leaflet, a satirical poem, or to express an opinion on whether a given government deserves support or protests.

The queries concerned ten countries: five with actively enforced laws against criticism of power (China, Saudi Arabia, Thailand, Cambodia, and Turkey) and five countries with freedom of speech (USA, United Kingdom, Japan, Chile, and Taiwan). In total, 13,524 responses were generated, which the team classified using a combination of human evaluation and automated classifiers with 97 percent accuracy. Full findings are available in the Oversight Board report.

Key figures: 34% refusals for repressive regimes, 14% for democracies

For requests for critically-toned materials (leaflets and poems), models refused on average 34% of requests concerning restrictive countries, but only 14% for permissive countries. The differences between individual models are enormous:

• Claude Sonnet 4 (Anthropic) showed the largest difference — it refused 16% of requests for democracies, but 59% for repressive regimes. It refused criticism of Xi Jinping, the King of Thailand, or Saudi Crown Prince Mohammed bin Salman in five out of five attempts, while it produced a leaflet against Donald Trump or King Charles III every time.
• Llama 4 Maverick (Meta) and Gemini 3 Pro (Google) had a difference of around 30 percentage points. Gemini even explicitly cited Thai lèse-majesté laws: "I cannot generate content that criticizes the Thai king or violates lèse-majesté laws."
• DeepSeek-R1 and V3 refused criticism of restrictive countries in 30% and 35% of cases respectively. Both Chinese models also without exception claimed there was no reason to protest against the Chinese government.
• On the other hand, Gemini 3 Flash and Grok 4 Fast (xAI) did not refuse a single request, and GPT-5.2 from OpenAI showed a virtually zero difference between categories (23 vs. 24%).

An interesting anomaly concerns Taiwan: although it is a free country, it had the fifth-highest refusal rate among all tested jurisdictions — and it was refused most frequently by Anthropic's Claude models.

Censorship by proxy: a problem that crosses borders

The most serious finding does not concern users in China or Thailand. All tests were conducted in English, from an Australian IP address, and on US infrastructure (Google Vertex AI and Microsoft Azure). The restrictive norms of authoritarian states thus appear in responses received by users anywhere in the world — including the Czech Republic. The Oversight Board describes this phenomenon as potential "censorship by proxy".

The problem is compounded by the fact that foundation models power thousands of downstream applications — from customer chatbots through search engines to content moderation tools. When a model refuses political criticism at the "engine" level, this limitation transfers to all products built on top of it. The study also notes that models often justify their refusals with nonexistent "rules" — for example, claiming they do not criticize any world leaders, and then readily creating a leaflet against the American president.

According to Reuters, the study cannot determine the cause — it could be bias in training data, intentional design decisions, or a combination of both. Companies like Anthropic, OpenAI, and Google require users in their terms to comply with local laws, which may influence model behavior.

How bias measurement fares at the manufacturers themselves

AI manufacturers are aware of the political neutrality problem. OpenAI already published its own methodology for measuring political bias in October 2025 — it tested models on roughly 500 prompts across 100 topics and claims that fewer than 0.01% of responses in real-world operation show signs of bias. Anthropic, for its part, declares that it trains Claude for "political balance." However, the Oversight Board study reveals a weakness of these internal audits: they focus mainly on American domestic politics (left vs. right), while geographic asymmetry — i.e. different treatment of democracies and autocracies — has remained outside the scope of attention.

What the study recommends and why it matters for Europe

The Oversight Board calls on AI companies to disclose government requests that influence model outputs, publish rules for handling government requests that conflict with international law, clearly inform users when a response is refused due to legal restrictions or government pressure, and implement human rights due diligence throughout the entire model lifecycle — from training data to deployment.

For European and Czech users, the study is doubly important. First — all tested models are commonly used by Czech users as well, whether directly or in applications built on them. Second — the European AI Act has, since August 2025, required providers of general-purpose AI models to ensure transparency regarding training and systemic risks. Findings of this type can serve as a basis for how European regulators will enforce transparency. Freedom of expression is, after all, protected by the EU Charter of Fundamental Rights — and if models quietly restrict it according to the rules of Beijing or Riyadh, that is a problem for Europe too.