Artificial-intelligence competition between the United States and China is no longer only about who can build the most powerful model.
It is also about who can learn from models built elsewhere.
A Reuters investigation has found that researchers associated with China’s military used outputs from leading American AI systems—including models created by OpenAI and Anthropic—to help train domestic defence-related models.
Reuters reviewed more than 80 Chinese research papers and patents linked to military institutions. The documents described the use of a technique known as model distillation, through which outputs from a more capable model help train a smaller specialised system. �
Reuters +1
Reported applications included drone navigation, surveillance, cyber operations, target recognition, and the analysis of software within secure military environments.
What Is Model Distillation?
A frontier AI model may contain enormous numbers of parameters and require expensive computing infrastructure.
A smaller organisation may not possess enough advanced chips to reproduce that system.
Model distillation offers another route.
A powerful “teacher” model generates answers, explanations, classifications, or examples. A smaller “student” model studies those outputs and learns to imitate parts of the teacher’s behaviour.
The smaller model may not reproduce everything the original system can do.
However, it can become highly effective within a narrower area.
For example, a distilled model may be trained specifically to:
Recognise objects in aerial imagery
Analyse computer code
Navigate a drone
Process surveillance material
Classify technical documents
Operate inside a secure local network
This makes distillation attractive where cost, chip availability, or security prevents direct use of a large foreign model.
What Did Researchers Reportedly Do?
Reuters found examples of Chinese military-linked researchers using outputs from American models as training material.
In one case, researchers reportedly used GPT-3.5 to examine source code before developing a local system capable of operating within a more secure environment.
Other papers described using frontier-model outputs to improve specialised domestic systems.
The significance is not merely that Chinese researchers used an American chatbot.
It is that commercially available intelligence may have been converted into a training resource for military applications. �
Reuters +1
Why the Story Went Viral
The story challenges the assumption that restricting advanced chips is enough to prevent rivals from acquiring sophisticated AI capabilities.
Export controls can limit access to computing hardware.
They can not easily prevent researchers from studying publicly available model outputs, academic papers, open-source code, or responses produced through commercial services.
Once intelligence can be accessed through an interface, the output itself may become a form of technological transfer.
The teacher does not need to hand over its internal architecture.
Its answers may still reveal enough patterns to educate another model.
Is Distillation Illegal?
Model distillation is not automatically illegal or unethical.
It is a common machine-learning technique used throughout the industry.
Companies distil their own larger models into smaller and faster systems. Researchers also use distillation to reduce computing requirements and make AI more accessible.
The controversy depends on several questions:
Were the source models accessed in accordance with their terms?
Were restrictions deliberately bypassed?
Was proprietary capability extracted?
Were the resulting systems used for prohibited purposes?
Did the process infringe intellectual property rights?
Can commercially available AI outputs legally be used to train competing models?
These issues remain politically and legally contested.
What Are OpenAI and Anthropic Likely Concerned About?
Frontier AI companies place restrictions on military, surveillance, and harmful applications.
But policy rules attached to an online service do not automatically control what happens after an output leaves the platform.
A researcher may collect responses and use them to train a separate local model.
That smaller system can then operate beyond the original company’s monitoring, safety filters, or governance framework.
This creates what might be called capability leakage.
The original model does not travel.
Its behavioural knowledge does.
The Limits of Distillation
A distilled model is not necessarily equal to its teacher.
It may inherit:
Factual errors
Biases
Security weaknesses
Incomplete reasoning
Gaps in specialist knowledge
It may also perform poorly outside the narrow tasks represented in its training material.
Nevertheless, “less capable than the frontier” does not mean harmless.
A smaller model tailored to one military or surveillance function may be more operationally useful than a general model that can discuss thousands of topics.
The Bigger AI Lesson
Artificial intelligence is different from many earlier technologies because interacting with the finished product can help reveal some of its capabilities.
A country denied access to advanced semiconductor manufacturing may still study:
Model outputs
Research publications
Open-source systems
Public benchmarks
Developer tools
API behaviour
The AI race is, therefore, not only a competition over physical infrastructure.
It is a competition over the movement of knowledge.
Mary Chuks’ Perspective
Human knowledge has always crossed borders.
Books, universities, migration, and international collaboration have accelerated civilisation.
But military AI introduces a difficult tension.
Open research can benefit medicine, science, and education while simultaneously making advanced capabilities available for surveillance or warfare.
The solution can not be to end all knowledge-sharing.
It must be to create governance that distinguishes beneficial access from high-risk capability transfer.
AI companies need more than usage policies written on webpages.
They need detection systems, access controls, auditing, android international agreements capable of addressing how outputs are collected and reused at scale.
Practical Takeaways
Policymakers and AI companies should consider:
Monitoring unusual bulk extraction of model outputs.
Restricting automated collection designed for capability replication.
Developing clearer rules for model distillation.
Requiring disclosure when one model is trained substantially from another model’s outputs.
Strengthening controls around military-linked accounts.
Supporting international standards for responsible military AI.
Preserving legitimate academic research while targeting high-risk misuse.
Conclusion
The future of technological competition may not depend entirely on who builds the original intelligence.
It may also depend on who can observe, extract, specialise, and redeploy that intelligence most effectively.
Models may be protected inside data centres.
Their knowledge is harder to contain once they begin answering questions.
Original Source and Further Reading
Original investigation: Eduardo Baptista, Reuters, “Chinese Military Researchers Tap US AI Models to Train Defence Systems,” published 31 July 2026. �
Reuters
Technical explainer: Reuters, “What Is AI Model Distillation and Why Is It Becoming a US-China Flashpoint?” published 31 July 2026. �