China has published a plan to supply data for training the world’s artificial intelligence. Western analysts warn that Communist Party narratives could travel with the Chinese AI training data.
The blueprint, released in June by China’s National Data Administration, aims to produce validated datasets across more than two dozen sectors by 2028. Fresh scrutiny followed on August 18, when the US-China Economic and Security Review Commission published a report concluding that Beijing is “marshaling data” to advance both its AI and military ambitions. Mike Kuiken, the commission’s vice chair, said the effort serves two goals at once: driving innovation, and tightening Communist Party control.
Inside China’s AI Training Data Blueprint
The plan targets a real bottleneck. Leading AI developers have largely exhausted the open web as a source of training material. The commission’s report noted that private firms, factories, and industrial sensors now hold much of the remaining data. China is building state-run data exchanges specifically to package and sell that material.
Officials have reported significant scale already. Liu Liehong, the National Data Administration’s director, said the country held more than 500 petabytes of high-quality datasets as of September 2025, alongside at least 4,000 interconnected exchanges, operators, and data merchants.
Beijing has paired this domestic build-out with an international push. President Xi Jinping addressed the World Artificial Intelligence Conference in Shanghai in July, where 29 countries signed an agreement establishing a new intergovernmental AI body. Xi also pledged 5,000 AI training placements for developing countries.
How Chinese Chatbots Already Dodge Sensitive Questions
Chinese AI models operate under domestic content law. The Cyberspace Administration’s 2023 interim measures require AI services to uphold “Core Socialist Values” and prohibit content that undermines national unity or state authority.
Testing shows the practical effect. Evaluation platform PromptFoo found that DeepSeek’s R1 model refused to answer roughly 85 percent of 1,360 sensitive prompts. The chatbot has repeatedly declined to discuss the 1989 Tiananmen Square crackdown and has avoided naming President Xi altogether, describing such questions as beyond its scope. A separate audit by NewsGuard in January 2025 scored the chatbot’s factual accuracy at just 17 percent, and found that Beijing’s official positions surfaced even in answers about unrelated foreign events.
Why This Reaches Western Models Too
Researchers say the effect isn’t confined to Chinese platforms. The American Security Project tested five chatbots in June 2025 and found that all five occasionally returned answers reflecting Party-aligned framing. Courtney Manning, the study’s lead author, said the models absorb material drawn from official Chinese sources, singling out DeepSeek and Microsoft’s Copilot in particular. A separate analysis found the divergence grew sharper when researchers prompted the models in Chinese.
Adoption trends compound the exposure. Hugging Face reported 2.05 billion downloads of Alibaba’s open-source Qwen models in 2026, compared with roughly 418 million for Google’s models and 227 million for Meta’s. Developers routinely fine-tune those open-weight models into thousands of downstream products — and researchers at the Slovak-based think tank CEIAS argue that political alignment can travel with the underlying weights into those products.
Analysts point to two distinct pathways here. Open Chinese models can carry their built-in filters directly into foreign-built products. Separately, Chinese-origin text increasingly seeps into the broader web that every AI developer scrapes for training data — a route analysts find more concerning precisely because it isn’t labelled or easily traced.
What’s at Stake
Three consequences stand out to researchers.
Silence becomes a default. Models trained on filtered source material inherit the gaps in that material, not just its slant — and users rarely notice what’s missing rather than what’s said.
Smaller languages face outsized risk. Chinese-language material makes up a relatively thin slice of the open web. As a result, curated official datasets can quickly come to dominate that slice. Governments across the Global South adopting inexpensive Chinese AI models risk reproducing those same defaults at a national level.
Standards follow the data. The commission’s report warned that Beijing is racing to set global rules for how data is governed. Its Beijing-based World Data Organization already counts members from 40 countries — and once such standards take hold globally, they rarely change.
Businesses face a subtler version of the same trade-off. China’s AI training data may prove highly valuable for robotics, logistics, and manufacturing applications, and firms may be willing to accept the political constraints that come with it in exchange for that industrial value. Alex Colville, a cyber expert at the Australian Strategic Policy Institute, warned that this dynamic hands authoritarian states real leverage over a chatbot’s underlying values.
What Beijing Says
Chinese officials reject the framing that this amounts to propaganda. Addressing the Shanghai conference, Xi said AI must not become any single country’s “solo performance.” The conference’s official chair statement also pledged support for Global South countries. Beijing presents its data-sharing initiatives as a way to help close the global AI divide.
Independent researchers add some nuance here. The American Security Project’s study also found bias in American AI models. That bias largely reflects the broader internet content used to train them rather than any single top-down directive.
The commission’s report also flagged real obstacles to Beijing’s ambitions. It noted that 85 percent of open Chinese government data was incomplete as of 2023. It also said that private firms still closely guard their most valuable datasets, while Beijing paused new data-backed securities in June over concerns about weak underwriting standards.
The commission recommended that Congress adopt a national strategy treating data as a formal economic asset. It also recommended new rules allowing companies to recognise data on their corporate balance sheets. The ambition behind China’s AI data strategy is clear. The execution, for now, remains unfinished.