Imagine spending a decade sweating through a dojo, breaking boards, doing the whole "wax on, wax off" grind, while someone in another room gets wired into a chair, downloads the exact same skill set in about six seconds, and calmly says "I know kung fu." That second person is basically what Chinese defense researchers have been doing with America's most advanced AI models, except the dojo is billions of dollars in training compute, and the download is just asking questions and writing down the answers.

A Reuters review of more than 80 Chinese academic papers and patents found researchers linked to the People's Liberation Army quietly building specialized defense AI by studying the outputs of leading American models from OpenAI and Anthropic. The technique has a name, model distillation, and in plain terms it means feeding a smaller, cheaper AI thousands of answers from a much bigger, far more expensive one until the small model learns to mimic the big one's reasoning, without ever building the actual dojo, meaning the billions of dollars in chips and training time it took to create the original. Researchers at China's North University of China, which has ties to the weapons industry, reportedly used Anthropic's Claude to help generate training data for a social media monitoring classifier. Others at the PLA's National University of Defense Technology used the same trick to shrink an image-recognition model small enough to run onboard a drone.

Here's the part that makes this legally messy instead of a clean scandal: distillation itself is an extremely normal, widely used industry technique, not some secret hacking method. The real dispute is over whether extracting these capabilities without permission crosses a line, not whether downloading kung fu is inherently illegal. One AI data executive was quick to note the copied kung fu still isn't quite as good as the teacher's, since a distilled model only inherits selected moves, not the instructor's full black belt.

The irony is thick either way. At the same time American AI giants, including SpaceX, Google, NVIDIA, Microsoft, Oracle, and AWS, are signing deals to help build the U.S. military into what officials call an "AI-first fighting force," those same models are apparently moonlighting as an unpaid study guide for the other side. Export controls were supposed to keep America's chip advantage locked up. Nobody built a wall around the answers.

If you're in defense tech or cybersecurity, this is exactly the kind of gap tools like Palantir AIP and Recorded Future exist to watch for. Everyone else, the lesson's simpler: if you build the smartest kid in class, expect someone to start sitting suspiciously close during the test.

Sources: Defense News (via Reuters) — Chinese Military Researchers Tap US AI Models to Train Defense Systems (https://www.defensenews.com/industry/techwatch/2026/07/31/chinese-military-researchers-tap-us-ai-models-to-train-defense-systems/); Technology.org — Chinese Military Distill US AI Models (https://www.technology.org/2026/07/31/chinese-military-distill-us-ai-models/)