Source-reported · Independent verification status is not recorded for this article; claims remain attributed to their sources.
China’s use of U.S. AI model outputs to train domestic systems challenges Washington’s export controls
China is increasingly training its own artificial intelligence models on outputs from American systems using a method known as distillation, prompting U.S. warnings about intellectual property risks and gaps in chip export rules, while Beijing defends the technique as a standard part of AI development.
Chinese researchers are leaning more heavily on leading American artificial intelligence models as training material for homegrown systems, raising questions in Washington over whether current export controls and intellectual property rules can contain the spread of advanced AI capabilities.
The approach relies on a technique known in the field as distillation, in which a smaller or less capable model is trained on the outputs of a more advanced one. In practical terms, that can involve sending large numbers of queries to U.S.-based AI systems, collecting their responses and using those responses as labelled data to teach domestic models to imitate their behaviour. This process lets Chinese teams avoid some of the cost and hardware demands of training very large models from scratch.
Because distillation depends on access to model outputs rather than to the physical chips running them, it poses a policy problem for U.S. officials who have tightened rules on exports of advanced semiconductors and AI accelerators to China. Those chip controls are designed to slow Beijing’s progress in sensitive applications by limiting access to high-end processors. As long as Chinese users can reach frontier models over the internet, though, they can still import some of the models’ capabilities without importing any hardware.
In Washington, officials and company lawyers increasingly frame this practice as a potential intellectual property issue, arguing that model outputs represent valuable results derived from proprietary architectures and training data. They have also raised concerns that distillation could be applied in areas with military relevance such as target recognition, signals analysis or cyber operations, sidestepping both export controls and conventional licensing.
Beijing and many Chinese AI specialists dispute that interpretation. They argue that training on publicly accessible AI outputs is comparable to training on web data that is available to anyone, and that distillation is a widely used method in global AI research for compressing large models, cutting costs and fine‑tuning behaviour. From their perspective, attempts to restrict how Chinese developers use accessible outputs look less like defence of intellectual property and more like an effort to freeze a technological gap.
For developers inside China, the practical incentives are clear. Training a cutting-edge large language model from the ground up can require very high spending on computing power and a large fleet of advanced GPUs or other AI accelerators, which have become harder to obtain under U.S. export rules. A domestic model distilled from the behaviour of an American system can run on less powerful hardware and be trained more quickly, opening the field to smaller firms and labs.
If distillation allows Chinese teams to get close to the performance of U.S. frontier models while relying on cheaper chips, the link between hardware access and AI capability weakens. That would reduce the leverage Washington gains from restricting semiconductor exports and might encourage other countries facing similar controls to adopt the same strategy.
For U.S. technology companies whose models are being queried at scale, the situation blends strategic and commercial concerns. They must decide how tightly to manage access from China, aware that broad bans or harsh limits could provoke retaliation from Beijing or accelerate the rise of purely local competitors. They also face the technical challenge of detecting and distinguishing large‑scale distillation efforts from ordinary high‑volume use by legitimate customers.
Signals to watch include whether the U.S. moves to regulate cross-border access to advanced AI services through licensing, geofencing or usage caps; whether American firms introduce new technical barriers or output watermarking aimed at limiting distillation; and how openly Chinese research continues to reference foreign models as teachers. Any mention of distillation practices in future U.S. export control updates would be a clear sign that Washington now sees this as a loophole it intends to address directly.