Chinese AI developers ‘using workarounds to distil US frontier models’

Chinese tech firms continue to train their AI models on the latest releases by US frontier labs despite efforts by Western firms to curb the practice, according to Reuters.

The newswire cited an analysis it carried out of over 300 academic papers, corporate disclosures, government files, patents, and recruitment notices to make the claim.

It added that the practice has now become mainstream, with firms using US AI models to prepare training data for Chinese AI models as well as to produce code at industrial scale.

Common methods of distillation, as the process is known, include creating proxy accounts to access frontier model APIs and structuring model outputs in such a way that they can be used to train other models.

Reuters claimed that the Chinese government appears to be encouraging this activity to spur the development of cheap and effective AI models.

Model distillation is not illegal, but is generally banned in the terms of service for AI platforms, particularly those accessed via API.

The practice has become a major point of contention between the US and China, with senior figures in the Trump administration having publicly accused Chinese firms of engaging in the activity.

In July, Michael Kratsios, director of the White House Office of Science and Technology Policy, posted on X: “Legitimate AI distillation used to create smaller, more efficient models plays a vital role in this open innovation ecosystem.

“However, large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable.”

In February, the AI developer Anthropic said it had been subject to “industrial-scale campaigns” by DeepSeek, Kimi developer Moonshot AI, and Minimax to distil its Claude models.

In all, Anthropic said, it had tracked 16 million distillation exchanges with Claude via 24,000 fraudulent accounts. In October, the firm said it had taken measures to decrease the effectiveness of distillation attempts on Claude, reducing the amount of data that such campaigns can harvest by 90 to 99 per cent.

Reuters added that distillation could help Chinese AI developers to circumvent new US export controls on advanced semiconductors. Currently, firms such as Nvidia are only allowed to sell older, less powerful chips in the country, but Chinese developers have been able to use chip clusters hosted in data centres in countries such as Malaysia, Singapore and Thailand.

The new rules come into force in 90 days and aim to put an end to this practice. They require data centre operators to carry out due diligence on customers to prevent access by users from countries deemed high-risk, including China, Russia and Iran.



Share Story:

Recent Stories


The future-ready CFO: Driving strategic growth and innovation
This National Technology News webinar sponsored by Sage will explore how CFOs can leverage their unique blend of financial acumen, technological savvy, and strategic mindset to foster cross-functional collaboration and shape overall company direction. Attendees will gain insights into breaking down operational silos, aligning goals across departments like IT, operations, HR, and marketing, and utilising technology to enable real-time data sharing and visibility.

The corporate roadmap to payment excellence: Keeping pace with emerging trends to maximise growth opportunities
In today's rapidly evolving finance and accounting landscape, one of the biggest challenges organisations face is attracting and retaining top talent. As automation and AI revolutionise the profession, finance teams require new skillsets centred on analysis, collaboration, and strategic thinking to drive sustainable competitive advantage.