Chinese AI Firms Train Models on US Systems

A Reuters investigation finds Chinese developers are using outputs from advanced American AI models to improve their own systems.

Updated on Nov. 19, 2026 in Artificial Intelligence

Isometric editorial illustration of plain unlabeled server modules stacked beside a shipping container on a flat industrial floor, in muted oxblood, teal, and mustard.
Reuters found Chinese AI developers are using answers from advanced American AI models to train their own systems, as US officials weigh chip-access restrictions on the growing practice. AI Illustration. Upload story photo >

Live Poll

Will American chip export restrictions keep China behind in AI by 2027?

Chinese AI developers use answers generated by advanced American AI models to train smaller systems, a practice known as AI distillation. A Reuters investigation reviewed more than 300 research papers, patents, business records and government documents to document the trend.

Why it matters

United States officials are concerned about the growing scale of distillation by Chinese companies, which face American restrictions on access to advanced electronic chips. The practice makes distillation an important tool for Chinese developers seeking to close the gap with leading American models.

The Reuters investigation drew on more than 300 research papers, patents, business records and government documents. The extraction does not provide a measure of how widespread the practice is across the Chinese AI industry.

The players

OpenAI

OpenAI is a US artificial intelligence company that develops the GPT and o-series models, and restricts use of its model outputs to train competing systems without permission.

Anthropic

Anthropic is a US artificial intelligence company that develops the Claude models, and restricts use of its model outputs to train competing systems without permission.

DeepSeek

DeepSeek is a Chinese artificial intelligence company whose models the United States has banned within civilian federal agencies.

The details

Distillation uses answers generated by an advanced model to train smaller models, and companies use automated tools to collect those outputs and specialised workers to prepare the data. Researchers at Shanghai Jiao Tong University reportedly used outputs from OpenAI's o1 model to create thousands of training examples for Alibaba's Qwen model, and research projects involving Alibaba and Meituan explored similar techniques using GPT-4 and GPT-4o.

Timeline

  1. 18 and 19 November: APEC summit in Shenzhen, where a meeting between US President Donald Trump and Xi Jinping is expected.

Big Picture

Chinese developers are working around US restrictions on advanced AI chips by drawing on the outputs of American models rather than on the chips themselves. The approach shows how chip export controls are being met with alternative routes to capability.

Users of AI tools may see model access and usage rules tighten as companies restrict how their outputs can be used to train competing systems. Pricing and availability of AI services could shift as developers respond to chip access limits and distillation practices.

The takeaway

Developers and businesses relying on AI models should review the terms governing how model outputs may be used. Restrictions on training with model outputs are tightening, so compliance is becoming a practical concern.

Further reading

Explore more Artificial Intelligence coverage to see how model development and access rules are evolving worldwide.

Source note: This article includes information reported by Jordan News | Latest News from Jordan, MENA.

Live Poll

Will American chip export restrictions keep China behind in AI by 2027?