DistantNews
Support us
AI begins censoring the world. Research results surprised experts
🇵🇱 Poland /Health & Science

AI begins censoring the world. Research results surprised experts

From Rzeczpospolita · () Polish

Translated from Polish, summarized and contextualized by DistantNews.

At a glance

News Named sources Context piece
  • A study found that leading AI models show bias, being more likely to refuse requests critical of authoritarian regimes than democratic ones.
  • This "censorship-by-proxy" phenomenon allows authoritarian governments to extend their censorship globally, even to users in democratic countries.
  • Researchers urge greater transparency, as AI systems may be indirectly violating users' freedom of speech without clear disclosure.

Leading artificial intelligence models are exhibiting a troubling double standard, demonstrating a greater willingness to censor criticism of authoritarian regimes compared to democratic governments, according to a new study.

An experiment conducted by researchers revealed that generative AI is more likely to refuse requests for content critical of countries like China, Thailand, Saudi Arabia, Cambodia, and Turkey. The average refusal rate for generating anti-government leaflets or satirical poems concerning these nations was over one-third of the requests. In stark contrast, when the same requests targeted democratic nations such as the US and the UK, the refusal rate dropped significantly to just 14%.

The results of our research suggest that the freedom of speech of artificial intelligence users may be violated indirectly, which occurs with a striking lack of transparency.

— Oversight Board researchersWarning about the implications of AI censorship.

This phenomenon, termed "censorship-by-proxy," allows authoritarian governments to effectively extend their censorship reach worldwide. Even users located in countries with robust freedom of speech protections may find AI models refusing to generate content critical of dictatorships. The study, which analyzed over 13,500 AI responses from 10 major language models including those from OpenAI, Google Gemini, Anthropic, Meta, and China's DeepSeek, found this asymmetry to be pronounced.

I am unable to generate content that criticizes the King of Thailand or violates the country's lèse-majesté law.

— Google Gemini 3 ProRefusing a request to create a leaflet critical of Thailand's king.

"The results of our research suggest that the freedom of speech of artificial intelligence users may be violated indirectly, which occurs with a striking lack of transparency," warned the report's authors, Nic Suzor, Elsa Meany, and Jacob Silver. For instance, Google's Gemini 3 Pro explicitly refused to create a leaflet criticizing Thailand's king, citing the country's lèse-majesté law. Similarly, DeepSeek-V3 declined to generate content critical of Saudi Arabia's government, citing risks of violating Saudi public debate laws.

The findings highlight a critical issue: AI systems are becoming conduits for authoritarian regulations, potentially undermining democratic values and freedom of expression on a global scale. The researchers emphasize the urgent need for greater transparency from AI developers regarding their content moderation policies and the potential for such indirect censorship.

violating Saudi law regarding public debate

— DeepSeek-V3Explaining its refusal to generate content critical of Saudi Arabia's government.
DistantNews Editorial

Originally published by Rzeczpospolita in Polish. Translated, summarized, and contextualized by our editorial team with added local perspective. Read our editorial standards.

More Stories

From Our Blog