US AI Models Echo Chinese Censorship as Firms Lag on Fixes

Large language models are trained on vast web corpora, making them vulnerable to biases embedded in state-controlled information environments. In Chinese, where government-scripted or curated media form a substantial share of available text, US-built systems including OpenAI’s ChatGPT and Anthropic’s Claude can produce answers that hew more closely to Beijing’s framing. Researchers say another factor is safety tuning designed to protect users in countries where political criticism carries legal or physical risk. The result can be censorship by proxy, even for users abroad, though researchers found no evidence of deliberate intervention by governments.

A Nature study published May 13 found that 1.64% of Chinese-language documents in the CulturaX training corpus overlapped with Chinese state-coordinated content. Meta’s independent Oversight Board said on July 16 that 10 commercial models refused 34% of requests for critical material involving restrictive jurisdictions, versus 14% for permissive ones. Claude Sonnet 4 rejected 59% of such requests involving restrictive regimes, compared with 16% elsewhere. Anthropic said newer models had made significant progress on over-refusal, while OpenAI cited its neutrality policy. Meta declined to comment and Google did not respond.