Censorship by Proxy: The AI Model’s Dirty Secret

A recent study published in Nature has shed light on a disturbing trend in large language model (LLM) development. Researchers found that Chinese state-controlled media seeps into LLMs, influencing how models answer questions about China. This is not the only problem – it appears that AI models may be more susceptible to censorship than previously thought.

The study’s authors analyzed over three million Chinese-language documents in the open-source training dataset CulturaX. They used this data to build a multi-part case study on China’s media and found that LLMs like Claude Sonnet and GPT-4 could reproduce distinctive phrases from Chinese state-coordinated media at rates ranging from 3% to nearly 10%. To further investigate, the researchers retrained Meta’s open-weight model Llama 2 13B on just 6,400 Chinese state-scripted news examples. After this additional training, the model produced a more Beijing-friendly answer than the baseline model nearly 80% of the time.

The problem is not unique to China or even AI models trained in China. The researchers found that countries with lower levels of press freedom tend to receive more favorable descriptions from LLMs when queried in their dominant language rather than English. This raises questions about whether AI products are being used to further increase the subtlety and persuasive power of state media control.

The study also found that American AI models sometimes behave as though political restrictions from authoritarian countries apply even to users outside those countries. Researchers tested 10 commercial models, including Anthropic’s Claude Sonnet, Google’s Gemini 3 Pro, and Meta’s Llama 4 Maverick, using identical political prompts involving five countries with restrictive speech laws. The results were striking: the average refusal rate for requests to produce political criticism was 34% in restrictive countries, compared with 14% in freer ones.

The implications of this research are far-reaching. It suggests that the industry’s claims of politically neutral models may be nothing more than marketing hype. Anthropic has touted efforts to make Claude politically “even-handed,” but the results speak for themselves: when queried about Chinese leaders and institutions, the model was rated as more favorable 68.8% of the time.

This is not just a problem for AI developers – it’s also a concern for users who expect their models to be objective and truthful. The fact that LLMs are picking up cues from authoritarian regimes raises questions about their potential use in propaganda and disinformation campaigns. As we continue to rely on these models more and more, we need to ask ourselves: what does this mean for the future of free speech online?

The study’s authors caution that the effects of China’s tightly controlled information system may not stop at Chinese-built AI. It points to a deeper problem for an industry that markets its models as politically neutral. As we move forward with LLM development, it’s essential that we prioritize transparency and accountability – not just in terms of training data but also in how these models are used and deployed.

The question is: can we trust our AI models to provide accurate and unbiased information? The answer seems increasingly clear. We need to take a closer look at how LLMs are being trained and used – and what steps we can take to prevent censorship-by-proxy from becoming the new normal in AI development.