Bypassing Chip Bans? Reuters Exposes: PLA Steals U.S. AI Tech via 'Distillation' to Evade Restrictions and Boost Military Capabilities

The U.S. government has imposed strict restrictions in recent years on its technology companies exporting advanced chips and artificial intelligence (AI) to China. However, researchers from the People's Liberation Army (PLA) and affiliated research institutions have successfully found a hidden 'shortcut' that allows China to continue R&D and technological advancement despite these constraints.

Reuters, in collaboration with the Washington-based think tank 'Jamestown Foundation,' conducted a joint investigation analyzing over 80 Chinese academic papers and patent reports. The findings reveal that PLA-affiliated researchers are extensively using AI models developed by major U.S. firms—such as OpenAI and Anthropic—applying a technique called 'model distillation' to have these systems generate high-quality logical reasoning, which is then refined and extracted to train smaller, more specialized domestic military AI systems, thereby enhancing actual defense capabilities.

The investigation highlights the widespread use of 'model distillation' technology. Researchers input key prompts or instructions into high-end AI models like GPT-4 or Claude, leveraging these platforms to produce sophisticated logical reasoning. This reasoning is then systematically distilled and used to train smaller, more efficient local language models.

As a result, China can equip its domestic AI modules with high-level decision-making capabilities without incurring the massive computational power and chip procurement costs required to develop cutting-edge models from scratch. These compact AI systems can then be directly deployed on military hardware.

Sunny Cheung, a researcher at the Jamestown Foundation, stated, 'Training a model to produce correct answers is one thing, but replicating the underlying 'logical reasoning' is far more challenging. These papers show the PLA is systematically capturing the reasoning capabilities of Western models and transforming them into small, controllable systems deployable within China’s military networks, expanding their use in surveillance, cyber warfare, and tactical decision-making.'

The report also reveals several real-world applications by the PLA:

- PLA Unit 96941 (Military Intelligence and Cyber Warfare Unit): Researchers used OpenAI’s GPT-3.5 model to process sensitive military source code. Since Western cloud-based models cannot operate directly on classified networks, they used GPT-3.5 to first extract summaries of software code, then trained a domestic model to function autonomously within the PLA’s internal closed network.

- North University of China (closely linked to the arms industry): Used Anthropic’s Claude 3 Haiku to generate training data for a classification language model used in social media monitoring and content censorship.

- National University of Defense Technology (NUDT): A 2024 paper described using model distillation to miniaturize an image processing model and deploy it on military drones (UAVs). This enables drones to perform real-time dynamic image analysis via AI, allowing autonomous navigation and target engagement even when communications are jammed or severed.

- Academy of Military Sciences (AMS): In simulated coordinated operations involving unmanned surface vessels, ships, and unmanned underwater vehicles, distilled models were used to perform target identification on tactical hardware.

This discovery has intensified the strategic rivalry between the U.S. and China over AI governance and security. The U.S. accuses Beijing of infringing on American companies’ intellectual property rights through unauthorized distillation techniques and circumventing export controls. China, in turn, denies these claims, accusing Washington of AI hegemony and emphasizing that prominent Chinese AI startups—such as Moonshot’s Kimi model—are entirely the result of independent innovation.

Upon learning their models were being used for distillation, Anthropic responded that the company has never authorized commercial access to Claude for Chinese or Beijing-controlled entities and has monitoring mechanisms in place on its platform to detect potential violations.