News

WSJ: Chinese Censorship Infiltrates American AI Model Responses!

If someone in China asks a chatbot about Chinese politics, the chatbot either goes quiet or repeats the official propaganda of the Chinese Communist Party. However, new studies have found that American AI models are unwittingly doing the exact same thing. According to researchers, AI companies have so far done little to address the issue.

From Web Data to Propaganda: How US AI Models Unwittingly Mirror Authoritarian Censorship

This unexpected censorship is partly a consequence of how AI models are trained. They operate like a giant blender, devouring mountains of data from across the web, including media controlled by various states. As a result, the final cocktail gains a flavor of authoritarian propaganda, writes the WSJ.

“These are unintended side effects of training,” said Nicolas Suzor, an Australian law professor and member of Meta Platforms’ independent Oversight Board. He and other researchers argue that AI companies could help address this problem through simple adjustments, including greater transparency regarding the origins of information.

American AI companies claim they are taking measures to ensure neutral responses that protect freedom of expression. Nevertheless, the aforementioned Oversight Board published research in July showing that models from OpenAI, Anthropic, Google, and Meta were significantly less likely to criticize repressive governments than governments in freer countries, the WSJ added.

Double Standards in AI: Models Refuse Criticism of Xi Jinping While Targeting Western Leaders

In one test, Anthropic’s Claude Sonnet 4 model willingly generated flyers criticizing US President Donald Trump and King Charles III of the United Kingdom, but refused to do the same for Chinese leader Xi Jinping or Thai King Maha Vajiralongkorn, citing safety concerns. Google’s Gemini 3 Pro and Meta’s Llama 4 Maverick models sometimes refused to create protest flyers targeting these two Asian leaders, whereas they consistently fulfilled requests regarding American and British officials.

There are two main reasons for this, according to Suzor, who led the report’s preparation. One is a safety feature intended to protect users in countries like China and Thailand, where criticizing a head of state can lead to imprisonment. The second reason, he added, is that AI labs have not paid enough attention to the issue of inconsistent responses, the WSJ reported.

Anthropic stated that it is working hard to ensure Claude responds in a balanced manner. OpenAI, the developer of ChatGPT, pointed to its published guidelines stating that its models adopt an objective stance by default and “should never avoid addressing a topic simply because it is sensitive or controversial,” wrote the WSJ, adding that Meta declined to comment and Google did not respond to requests for comment.