Study shows Chinese state media subtly shapes answers from global AI models

Research published in Nature reveals that Chinese government-controlled news appears in training data and can bias large language models toward Beijing-friendly responses, while an oversight board found AI systems often refuse political criticism of authoritarian regimes.

A new study in Nature examined the influence of Chinese state-run media on large language models, finding over three million Chinese-language texts in the open-source training set CulturaX. Models including Claude Sonnet, Claude Opus, GPT-3.5, GPT-4 and GPT-4o reproduced distinctive state-media phrases at rates up to nearly 10%. When Meta’s open-weight Llama 2 13B was further trained on just 6,400 government-scripted articles, it produced Beijing-aligned answers nearly 80% of the time, and after 64,000 examples it even described China as democratic.

Parallel research by Meta’s Oversight Board tested ten commercial models with identical political prompts across restrictive and permissive countries, finding a refusal rate of 34% versus 14% respectively. Models such as Anthropic’s Claude and Google’s Gemini refused criticism of leaders like Xi Jinping while readily generating criticism of leaders in freer nations. The authors warn that such hidden influences could amplify state media control worldwide, though no direct manipulation by Beijing of the examined companies was proven. Meta declined comment; Anthropic, OpenAI and Google did not respond.