EXCLUSIVE: Chinese military researchers tap US AI models to train defence systems

According to a Reuters review, researchers affiliated with the Chinese People's Liberation Army have systematically harvested outputs from leading US AI models — including systems from OpenAI and Anthropic — to train specialized domestic platforms for defense applications. The findings, drawn from over 80 recently published Chinese academic papers and defense patents, expose a structural gap in Washington's export-control regime: closed-weight models accessed via APIs remain susceptible to output extraction even when underlying parameters remain protected.

Distillation at Industrial Scale

The documented technique is model distillation, a methodology long familiar to the ML community. Complex, multi-layered prompts are routed to a large proprietary "teacher" model — specifically OpenAI's GPT-4 and Anthropic's Claude — and the generated outputs, including logical reasoning pathways and code structures, are captured and aggregated into a training corpus for smaller, specialized "student" models. The efficiency differential is the critical variable. Student models require drastically less compute and memory than their teachers, enabling deployment on constrained military hardware such as autonomous drone swarms and frontline tactical servers. Offline operation removes the dependency on high-end NVIDIA GPUs — the precise category of hardware subject to US export restrictions — and sidesteps the compute ceiling that hardware controls are designed to enforce.

Documented Application Domains

The reviewed corpus identifies three primary application areas: cyberwarfare, drone swarm coordination, and tactical surveillance. Analyst Sunny Cheung, who reviewed dozens of PLA-linked documents in conjunction with the Washington-based Jamestown Foundation, characterizes the objective as the acquisition of complex reasoning patterns rather than raw data extraction. The operational logic: instructing a model to produce a correct answer is straightforward; transferring the underlying reasoning process is substantially harder and therefore more valuable. Distilled outputs from frontier models compress years of costly trial-and-error development into a single training pipeline, shortening the iteration cycle required for algorithmic warfare capabilities.

Open Variables

The Reuters reporting does not specify whether API providers can detect systematic distillation patterns at the access layer, nor whether rate-limiting, query-pattern analysis, or output fingerprinting is technically feasible at industrial scale. For the ML research community, the precedent is established: closed-weight frontier models can be partially reproduced through legitimate API use, with capabilities transferred to models orders of magnitude smaller than the original. Controlled replication studies — quantifying the fidelity gap between distilled student models and their teacher counterparts across standard reasoning and code-generation benchmarks — would clarify the actual rate of capability transfer. Until such measurements are published, the efficacy of API-level containment remains an open empirical question, and downstream reproducibility constraints on dual-use research become harder to enforce.