A Reuters investigation that changes the perspective
Reuters reviewed more than 80 scientific papers and patent applications from China, identifying several projects in which researchers affiliated with the People's Liberation Army used responses generated by OpenAI and Anthropic models to train specialized AI systems.
The investigation does not claim that the Chinese military gained direct access to the weights of U.S. AI models. Instead, the documents describe the use of model distillation, a technique in which a new model is trained using the outputs generated by an existing one. Unauthorized distillation of a competing system is widely regarded as an improper way of appropriating a rival's research efforts.
A smaller, distilled model optimized for specific tasks may be sufficient while being significantly more cost-effective. This is precisely why distillation has become attractive for developing AI systems designed for highly specialized applications, without incurring the enormous cost of training a frontier model from scratch.
Distillation becomes a strategic issue
Until recently, model distillation was viewed primarily as a commercial dispute between AI companies.
OpenAI and Anthropic have repeatedly accused Chinese firms of using their models' outputs to develop competing systems. The cases involving DeepSeek and, more recently, Moonshot AI have further fueled this debate.
The Reuters investigation adds an entirely new dimension: if these techniques are also being used in military or national security projects, the stakes extend far beyond commercial competition.
For Washington, this development provides an additional argument for restricting exports of advanced AI chips and limiting access to the most capable American AI models.
It also highlights a difficult technical challenge. Even if access to a model's weights is protected, the outputs it generates can, under certain circumstances, become the raw material for training other models.
In other words, not only chips and source code but also the knowledge produced by an AI model during its use can be considered a strategic resource.
More than a commercial dispute
The Reuters investigation demonstrates how important it has become to control how powerful AI models can be used to create new ones. If model distillation was previously seen mainly as an intellectual property issue, it is increasingly being viewed as a matter of national security and geopolitical advantage.
At the same time, distillation is not a perfect copy of the original model. It always involves a trade-off between cost, size, and performance.
In this context, pressure for stricter rules governing access to next-generation AI models is likely to increase, while the debate over balancing openness with the protection of strategic technologies is set to become an increasingly important issue in the years ahead.
