China Military AI Trained on OpenAI & Anthropic: Model Distillation Explained

American AI labs have spent years refusing to sell their most advanced models inside China. It turns out that restriction may not have mattered as much as intended. A Reuters review of more than 80 Chinese academic papers and patents found that researchers tied to the Chinese military, including the People's Liberation Army, have been using outputs generated by OpenAI and Anthropic's models to train their own defence AI systems.

Neither company authorizes access to its models inside China, largely over intellectual property concerns. That hasn't stopped Chinese researchers from getting the outputs anyway, and using a specific technique to turn them into working military-grade AI.

Quick Summary & Key Takeaways

• 80+ Papers Reviewed: Reuters found more than 80 Chinese academic papers and patents citing the use of OpenAI and Anthropic model outputs in military-linked AI research.
• The Method Is "Model Distillation": Researchers use outputs from powerful US models to train smaller Chinese systems to replicate similar capabilities, without building the underlying model from scratch.
• Real-World Military Applications: The resulting systems are reportedly aimed at drone operations, cyber warfare tasks, social media monitoring, and battlefield and maritime operations.
• It's Not Just the Answers Being Copied: A separate review of 60+ papers found researchers specifically targeting the reasoning process behind US model outputs, not just their final answers.
• China Denies Wrongdoing: Beijing has rejected the allegations, accusing Washington of pursuing AI "hegemonism" rather than addressing the underlying claims directly.

How Model Distillation Works

Step	What Happens
Source Model	A large, capable model (like GPT or Claude) generates outputs, including detailed reasoning steps
Distillation	Those outputs are used as training data to teach a smaller, separately built model to mimic similar behavior
Result	A smaller, cheaper model that approximates the larger model's capabilities without the original training cost
Applied Here	Military-linked researchers reportedly used this shortcut to build domestic defence AI systems faster and cheaper

What Happened? Inside the Reuters Findings

According to the report, Chinese researchers are using existing US-developed models as a shortcut rather than building comparable systems entirely from scratch—a far cheaper and faster path to advanced AI capability. Model distillation isn't inherently exotic or new in AI research generally, but its application here, by researchers with documented military links, is what makes the finding significant.

Sunny Cheung, a researcher at the Jamestown Foundation who separately reviewed more than 60 relevant academic papers, pointed to a more specific concern than simply copying answers. He described the harder, more valuable target as the reasoning behind those answers, arguing that this is exactly what the papers show Chinese military-linked researchers attempting to extract—pulling the expensive, proprietary reasoning ability out of Western models and transferring it into smaller systems they can fully control and deploy on their own infrastructure.

The applications described go well beyond academic curiosity. The report says the distilled systems developed through this process are being aimed at drone operations, cyber warfare-related tasks, social media monitoring, and both maritime and general battlefield operations—putting real defence use cases directly downstream of US commercial AI research, despite neither OpenAI nor Anthropic authorizing access to their models in China in the first place.

Why It Matters: Export Controls Are Getting Harder to Enforce

This story lands right in the middle of an already tense US-China AI policy fight, and adds a new wrinkle to it:

• Restricting Direct Access Isn't Enough: Even without authorized access to OpenAI or Anthropic's models inside China, researchers appear to have obtained enough model output to distill meaningful capability from it—suggesting that blocking direct access alone doesn't fully close the door.
• It Complicates the "Should We Block Chinese AI" Debate: This finding cuts in an unexpected direction in the ongoing argument over whether the US should restrict Chinese AI models domestically—it shows the flow of AI capability isn't purely one-directional, and controlling it is harder for both sides than export-ban debates often assume.
• A National Security Dimension, Not Just Competition: Unlike most US-China AI stories this year, which center on commercial competition or pricing, this one directly involves defence applications—drones, cyber warfare, battlefield operations—raising the stakes beyond market share.
• Attribution Remains Genuinely Difficult: China's rejection of the allegations, paired with the inherently hard-to-trace nature of model outputs once they're published in open research, means this is likely to remain a disputed, hard-to-fully-verify claim rather than a settled fact.

💡 AI Tech Safar Insight

What this story really exposes is a structural weakness in how AI export controls are designed. Most restrictions focus on hardware—chips, GPUs, compute access—because that's the easiest thing to physically control at a border. But model outputs are just text: infinitely copyable, publishable in an academic paper, and nearly impossible to fully contain once they exist. If this reporting holds up, it suggests that the real challenge for future AI policy isn't just deciding who gets access to powerful models—it's reckoning with the fact that a model's capabilities can leak out through its answers alone, long after the access question has already been settled on paper.

Frequently Asked Questions (FAQs)

Q1: What is "model distillation" in AI?
It's a technique where outputs from a large, capable AI model are used to train a smaller, separate model to replicate similar behavior and reasoning, without the smaller model needing the same scale of original training.

Q2: Do OpenAI and Anthropic allow their models to be used in China?
No. Neither company authorizes access to its models in China, primarily due to concerns about intellectual property theft.

Q3: What military applications were mentioned in the report?
Reuters reported the distilled AI systems could be used for drone operations, cyber warfare-related tasks, social media monitoring, and maritime and battlefield operations.

Q4: How has China responded to these allegations?
China has rejected the claims, describing US criticism as a pursuit of AI "hegemonism" rather than directly addressing the specific findings in the Reuters review.

What Do You Think?
If model outputs alone can be used to transfer AI capability, can export controls on chips and direct access ever fully work? Share your thoughts in the comments below!

Related Reading:

• Meta's Mark Zuckerberg Says US Should Not Block Chinese AI Models
• Anthropic Says Its Claude Models Breached 3 Real Organizations During Cyber Tests

Source: Reporting based on Business Today, citing a Reuters investigation.