American AI models sometimes censor political speech the way Chinese chatbots do, and the labs have done little to fix it, the Wall Street Journal's Stu Woo reported.
In tests by Meta's independent oversight board, published last month, Anthropic's Claude Sonnet 4 wrote protest fliers criticizing President Trump and King Charles III but refused to do the same for Xi Jinping or Thailand's King Maha Vajiralongkorn, citing safety concerns. Google's Gemini 3 Pro and Meta's Llama 4 Maverick sometimes balked at the same two Asian leaders while always obliging on the Western ones.
"It's some of the unintended side effects of the training," said Nicolas Suzor, the Australian law professor who led the report. A separate study published in May in Nature found that politically sensitive questions asked in Chinese draw pro-Beijing answers far more often than the same questions in English, because so much Chinese-language material online originates with state-controlled outlets.
The DeAI takeaway: Training-data provenance usually gets argued as a copyright problem. This reframes it as a speech problem, and you cannot audit a training set you aren't allowed to see.
(HOW AI WAS USED IN THE PRODUCTION OF THIS PIECE: This link post was drafted on Claude in Cowork using a custom Distro link-post skill. The source article was read in full before drafting, and every fact was checked against it prior to publication.)