Osegi
AI·2 min read

AI distillation: how Chinese firms learn from US AI models

Chinese developers are training cheaper models on answers from OpenAI's systems and others. A Reuters review of 300+ documents shows how common it has become.

A glowing AI chip split between an American and a Chinese flag, with data streaming from one side to the other.
Image: Digital Trends

Chinese AI developers have found a way to keep pace without building everything from scratch: they train their own models on answers produced by leading American ones. The technique is called AI distillation, and it is now spreading across China's AI industry, according to a Reuters investigation of more than 300 academic papers, patents, company records and government documents.

What is AI distillation?

Think of a student learning from an experienced teacher. A powerful model writes answers, produces code or works through hard problems. Developers then use those responses to train a smaller model that can do similar tasks with far less computing power. That matters for Chinese firms facing limits on advanced US chips. ByteDance and Tencent have secured shipments of Nvidia's H200 chips, but the most advanced hardware stays restricted.

Distillation is not new and not inherently illegal. It also does not produce an exact copy. Developers still need expertise, computing resources and original research.

What did the investigation find?

Researchers at Shanghai Jiao Tong University used responses from OpenAI's o1 to build thousands of training examples for Alibaba's Qwen model. Other projects involving Alibaba and Meituan explored similar methods with GPT-4 and GPT-4o. It is no longer just a university pastime: Reuters found companies building automated tools to collect AI-generated responses, plus specialised workers who prepare that material for training. The Chinese government is encouraging distillation too, to make AI cheaper and easier to deploy, especially in manufacturing.

Why Washington is worried

The US has spent years restricting China's access to advanced AI chips. If Chinese developers can improve their models using American outputs, those limits may matter less. Companies including OpenAI and Anthropic also bar the use of their models' outputs to train competing systems without permission.

US officials have warned that large-scale distillation crossing into intellectual-property theft could bring sanctions. Chinese officials call distillation a legitimate technique, though they acknowledge that real IP violations deserve action. The US recently banned DeepSeek from federal civilian agencies, including third-party software built on its models.

The timing is awkward. Donald Trump and Xi Jinping are expected to meet at the APEC summit in Shenzhen on November 18 and 19.

Apps worth a look

SUMO: the number puzzle

Tap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.

Tags:ai distillationchinaopenaialibabachip restrictionsanthropic

More from this day

See all →

More in: AI

See all →