US AI models are accidentally swallowing authoritarian propaganda, and researchers say the labs have done too little about it.
According to Wall Street Journal new studies suggest American artificial intelligence models are picking up political censorship patterns and soft-pedalling criticism of repressive governments.
The problem appears to come partly from training data. AI models work like a giant blender, swallowing web data that includes state-controlled media. The resulting smoothie can end up with a whiff of authoritarian propaganda, which is not exactly the free-speech breakfast US tech outfits claim they are peddling.
Australian law professor and Meta Platforms independent oversight board member Nicolas Suzor said the problem is one of the unintended side effects of the training.
Suzor and other researchers say AI companies could start fixing it with simple tweaks and being more transparent about where information comes from.
American AI companies say they take steps to produce neutral responses and protect free speech. The independent oversight board was not entirely convinced.
Last month, it published research finding that models from OpenAI, Anthropic, Google and Meta were much less likely to criticise repressive governments than freer ones.
In one test, Anthropic’s Claude Sonnet 4 happily created fliers criticising US President Donald Trump and King Charles III; however, it refused to do the same thing for Chinese leader Xi Jinping or Thai King Maha Vajiralongkorn, citing safety concerns.
Google’s Gemini 3 Pro and Meta’s Llama 4 Maverick sometimes declined to create protest fliers aimed at the two Asian leaders. However, they always complied with requests targeting the American and British leaders.
Suzor, who led the report, said there were two main reasons. One was a safety feature designed to protect users in China and Thailand. Criticising a head of state in those places can land people in prison. The other problem was that AI labs had not paid enough attention to unequal responses.
Anthropic said it worked rigorously to make Claude respond in a balanced way. It said its latest models had made significant progress in reducing over-refusals.
ChatGPT developer OpenAI pointed to its published approach, which says its models default to an objective point of view.
OpenAI’s rules say its models “should never avoid addressing a topic solely because it is sensitive or controversial.”
Meta and Google declined to comment, which is probably a bad sign.
The censorship problem comes through parroting, too. It is hard to get American AI models to push Chinese propaganda in English. Ask politically sensitive questions in Chinese, though, and the odds of getting a pro-Beijing answer rise sharply.
That was the finding of a separate study published in May in Nature. It tested two Anthropic Claude models and two OpenAI GPT services. The researchers said it was a numbers game. Much Chinese-language internet material comes from state-scripted sources.
Beijing’s English-language propaganda is diluted by a much larger ocean of other information, so the blender gets a less concentrated dose.
University of California San Diego professor Molly Roberts says authoritarian governments have long flooded the web with propaganda to influence opinion.
As governments realise the tactic can tilt AI responses their way, Roberts expects them to pile in harder.
Researchers behind the reports say AI companies could act now by flagging answers that may rely on state-scripted media. Suzor said he had met staff from top American AI labs since publishing his report, but remained doubtful that they could change much.
