Chinese Military Used OpenAI and Anthropic Outputs to Train Defence AI, Reuters Investigation Finds

Chinese military-linked researchers have reportedly used outputs from advanced artificial intelligence models developed by OpenAI and Anthropic to train domestic AI systems for defence-related applications, according to a Reuters review of more than 80 Chinese academic papers and patents.

The findings have intensified the ongoing debate over AI governance, intellectual property and technology competition between the United States and China.

Reuters Review Highlights Military-Linked AI Research

According to Reuters, researchers associated with the People's Liberation Army (PLA) and other Chinese defence institutions relied on a technique known as model distillation to develop specialised AI models.

The investigation, supported by research from the Jamestown Foundation, suggests Chinese military institutions have used outputs generated by leading Western AI models to accelerate domestic AI development while reducing computing costs.

The report states that this practice allows smaller AI models to inherit selected capabilities from larger frontier models without training them entirely from scratch.

What Is Model Distillation?

Model distillation is an established machine learning technique in which a smaller AI model is trained using responses generated by a larger, more capable model.

Instead of recreating a frontier AI system from the ground up, developers use the larger model's outputs to teach a smaller model specific reasoning patterns or specialised tasks.

The process itself is widely used across the AI industry. However, the dispute centres on whether proprietary AI outputs can be used without authorisation to reproduce valuable capabilities.

Defence Applications Mentioned in the Research

The reviewed papers describe several defence-oriented use cases, including:

• Military software code analysis
• Tactical decision-making support
• Drone navigation
• Target recognition systems
• Maritime surveillance
• Cybersecurity applications
• Social media monitoring
• Content classification

One paper reportedly described researchers using GPT-3.5 to summarise sensitive military software code before training a domestic AI model capable of operating entirely within Chinese military networks.

Other studies referenced lightweight AI models designed for drones, unmanned vehicles and battlefield decision support systems.

AI Governance Becomes a New Flashpoint

The report comes as AI governance emerges as a key issue in relations between Washington and Beijing.

US officials have repeatedly accused some Chinese organisations of extracting capabilities from American AI models in ways that may undermine export controls and intellectual property protections.

Chinese officials have rejected those allegations, arguing that the United States is attempting to maintain technological dominance while restricting China's AI development.

Anthropic and OpenAI Policies

According to the report, Anthropic stated that it does not provide commercial access to Claude models in China or to Beijing-controlled organisations and uses monitoring systems to detect policy violations.

The company also warned that distilled models may lose important safety mechanisms present in the original AI systems.

Reuters noted that OpenAI, the White House, the Pentagon, China's Foreign Ministry and the PLA did not immediately respond to requests for comment regarding the findings.

Why Distillation Matters

Experts say model distillation offers significant advantages by reducing computing requirements and enabling AI deployment on devices with limited hardware, including drones, satellites and autonomous systems.

However, specialists also caution that distilled models cannot fully replicate the broad reasoning abilities of frontier AI models.

Instead, they inherit selected capabilities tailored for specific applications.

AI Competition Between the US and China Intensifies

The findings arrive amid increasing competition between the world's two largest AI powers.

China continues investing heavily in domestic AI development while facing US export restrictions on advanced semiconductor technology.

At the same time, American AI companies are strengthening safeguards around their models to prevent unauthorised capability extraction.

As international discussions on AI governance continue, the issue of model distillation is expected to remain a central point of debate.

Key Highlights

• Reuters reviewed more than 80 Chinese academic papers and patents.
• Military-linked researchers reportedly used outputs from OpenAI and Anthropic models.
• Researchers employed model distillation to train domestic AI systems.
• Applications included drones, surveillance, cybersecurity and military software analysis.
• The findings add to growing tensions over AI governance between the US and China.
• OpenAI, the Pentagon and Chinese authorities did not immediately comment on the report.

FAQs

What is model distillation?

Model distillation is an AI training technique where a smaller model learns from the outputs of a larger AI model to perform specialised tasks more efficiently.

Did Reuters claim China directly accessed OpenAI's internal models?

No. Reuters reported that researchers used outputs generated by publicly accessible AI models, not the underlying model weights or internal systems.

Why is this controversial?

The controversy concerns whether proprietary AI-generated outputs can be used without permission to reproduce valuable capabilities, potentially raising intellectual property and security concerns.

What defence applications were mentioned?

The papers referenced drone navigation, target recognition, military software analysis, cybersecurity, surveillance and tactical decision support.

Did OpenAI or Anthropic respond?

According to Reuters, OpenAI did not immediately comment. Anthropic said it restricts commercial access to Claude in China and monitors for policy violations.