BEIJING, November 17, 2026 — Chinese technology developers are increasingly using outputs generated by leading American artificial intelligence models to help train competing domestic systems, according to a Reuters investigation that highlights a new dimension of the international AI race. The practice, commonly known as model distillation, involves using the responses or behaviour of a more capable model to improve a smaller or competing system. Reuters’ investigation, republished by Investing.com, reviewed hundreds of Chinese academic papers, patents, corporate disclosures and other documents.

In a typical distillation process, a developer sends questions or tasks to a more advanced model and collects its answers. Those outputs can then be used as examples when training another model to solve similar problems. The approach can help developers improve reasoning, coding or task-planning abilities without repeating every stage of the original model’s expensive training process. It does not automatically mean that the resulting system copies the original model’s internal architecture or weights. The Reuters report describes how the practice has moved beyond isolated academic experiments into wider industrial use.

The investigation cites examples in which researchers used American models to generate plans for coordinating AI agents on programming tasks. Other reported applications include generating answers, producing computer code and helping solve difficult technical problems. Specialist workers and software systems can then organise the resulting material for use in training domestic models. The scale and sophistication of these operations are part of what makes the issue significant for global technology companies. Read the investigation and its cited examples.

The development is taking place against a backdrop of US restrictions on exports of advanced semiconductor technology to China. High-performance chips are essential for training and operating large AI systems, and limits on access can increase the cost and complexity of developing frontier models. Distillation offers one possible way to make more efficient use of available computing resources, although it cannot eliminate the need for capable hardware, skilled engineers, quality data and original research. Reuters details the connection between chip restrictions and the expanding use of distillation.

The practice also raises questions about intellectual property and the terms governing access to commercial AI services. Leading American AI providers generally restrict the use of their model outputs to develop competing systems, subject to the precise terms and exceptions in their agreements. The legal position can depend on how the outputs were obtained, the applicable contract and the jurisdiction involved. The Reuters investigation discusses the restrictions and concerns raised by US officials.

Chinese developers and other industry participants have also argued that distillation is only one contributor to improvements in domestic AI. Advances in algorithms, model design, engineering and data quality remain important. That distinction matters because attributing progress to a single technique can oversimplify a highly competitive research field. Reuters includes competing assessments from people familiar with Chinese developers’ operations.

The wider significance is that AI competition is no longer confined to who builds the largest model or owns the most advanced chip. It increasingly involves training methods, access to computing, software ecosystems, licensing rules and the ability to turn research into affordable products. Governments and technology companies are likely to pay closer attention to model-output policies as these techniques become more widely used. Read the full Reuters investigation.