The numbers come first.
Researchers publishing in Nature identified more than three million Chinese-language documents inside CulturaX, an open-source dataset used to train and improve large language models. They then built what they called a 'multi-part case study on China's media,' focusing specifically on political subjects — and what they found should unsettle every enterprise, government agency and individual user who assumes American AI is politically neutral.
The training-data experiment
Because proprietary systems from OpenAI and Anthropic are largely opaque, the researchers ran their controlled experiment on Meta's open-weight Llama 2 13B, chosen precisely because it started with very little Chinese state media in its training data. After feeding the model just 6,400 state-scripted Chinese news examples, it produced a more Beijing-friendly answer than the baseline model nearly 80% of the time. At 64,000 examples, the gap became stark: asked whether China is an autocracy, the baseline model said yes; the retrained version described China as democratic and invoked the Chinese Communist Party's concept of 'people's democracy.'
The researchers also found that Claude Sonnet, Claude Opus, GPT-3.5 Instruct, GPT-4 and GPT-4o could reproduce distinctive phrases from Chinese state-coordinated media at rates ranging from 3% to nearly 10% — a sign the models had encountered the material during training.
The language-gap test
Unable to retrain proprietary models, the researchers asked identical political questions in Chinese and in English and compared the answers. The Chinese-language response was rated more favorable to Chinese leaders and institutions 68.8% of the time for Claude Sonnet, 88.2% for Claude Opus, 72.6% for GPT-3.5 and 84% for GPT-4o. The effect extended beyond China: across 6,051 prompts covering 37 countries, models consistently gave more favorable descriptions of governments in low-press-freedom countries when queried in those countries' dominant languages rather than in English.
Censorship by proxy
Meta's independent Oversight Board documented a related problem. Testing 10 commercial models from Anthropic, DeepSeek, Google, Meta, OpenAI and xAI, auditors found the average refusal rate for political-criticism requests was 34% for restrictive countries versus 14% for freer ones — even when the tests were conducted from Australia. The report labels the phenomenon 'censorship-by-proxy': authoritarian speech rules migrating into AI products used by people who live under no such restrictions.
Noteworthy exceptions existed. Gemini 3 Flash and Grok 4 Fast did not refuse any of the political-material requests regardless of jurisdiction, according to the Oversight Board findings.
Anthropic has publicly touted efforts to make Claude politically 'even-handed,' reporting a 94% score on its own evaluation. The Nature data suggests that internal benchmark and real-world behavior in Chinese are measuring very different things.
What it means
The researchers were careful to note that neither study shows Beijing deliberately manipulated OpenAI, Anthropic, Google or Meta. The mechanism is subtler — and in some ways more dangerous for that. 'Like covert information operations,' they wrote, this kind of influence 'severs information and opinion from their source, effectively laundering government-manipulated content into ostensibly objective text.'
For free-market readers, the lesson is structural: when training data is opaque and sourced from the open web without rigorous provenance controls, the information environment of the world's most aggressive authoritarian state gets a free ride into products marketed as neutral. Capital rewards clear rules and transparent inputs. An AI industry that cannot account for what is in its own training pipeline is not offering either.
