Chinese military-linked researchers are drawing on leading US artificial intelligence models to help build homegrown defence AI, according to a Reuters investigation.
The investigation, based on a review of more than 80 Chinese academic papers and patents, found that institutions tied to China’s armed forces have used a technique called model distillation to create specialised systems for surveillance, cyber operations, drone control and tactical decision-making.
Model distillation works by training a smaller AI model to mimic the outputs of a larger one, cutting the computing power needed to run it. That makes it possible to fit useful AI capability onto resource-constrained platforms such as drones, tactical vehicles and battlefield sensors, rather than requiring a data-centre-scale system.
The findings feed into a long-running debate in Washington over how to keep advanced AI systems, even ones released as open commercial products, from being repurposed by rival militaries. Distillation is attractive precisely because it does not require direct access to a company’s original model weights or training data; it only needs enough interaction with a model’s outputs to train a smaller copy, which makes the technique difficult to police through export controls on hardware or software alone. The episode illustrates a broader challenge facing dual-use technology policy: once a powerful AI model is public, it becomes very hard to stop determined actors, including foreign militaries, from mining it for capability gains.