
Chinese Teams Distill US AI Model Output at Industrial Scale
Reuters reviewed 300-plus Chinese documents showing distillation of US model output at industrial scale, as Anthropic reports 16 million fraudulent Claude exchanges.
- By
- Sophie Lindqvist
- Filed
- Channel
- AI & Compute
- Read
- 4 min read
Anthropic counted more than 16 million exchanges with Claude through roughly 24,000 fraudulent accounts, and it traced them to three Chinese AI labs. That figure, published on Feb. 23, anchors a Reuters investigation published Tuesday, Nov. 17, which found that Chinese companies now systematically use answers from leading US models to train domestic rivals as Washington tightens chip export controls.
Reuters reviewed more than 300 Chinese academic papers, patents, corporate disclosures, recruitment notices and government documents. The review shows distillation of US model output has spread from academic work into mainstream industrial use. The technique uses a powerful "teacher" model to generate answers, which developers then use to train a smaller "student" model.
What do the documents show?
The records describe purpose-built software that automates the collection of US model output, plus specialist workers who prepare training material. A former Alibaba Group engineer told Reuters the appeal is simple: one expensive query does repeated work. An American model answers a hard question once, and a Chinese team keeps the example to train a local system.
The paper trail is specific:
- In a November 2024 paper, Shanghai Jiao Tong University researchers selected about 5,000 answers from OpenAI's o1 model, translated them with GPT-4o mini and built 10,750 bilingual examples to further train Alibaba's Qwen2.5-72B.
- Alibaba researchers described a more systematic method in an April 2025 paper, using GPT-4 and GPT-4o to generate training material in languages other than Chinese, then scoring and refining responses before training smaller Qwen models.
- A February 2026 paper by Shanghai Jiao Tong University and Meituan used GPT-4o to generate coordination plans for AI agents on programming problems, which trained a smaller model to organize such tasks.
Shanghai Jiao Tong University, Alibaba and Meituan did not respond to Reuters' requests for comment. That silence matters. The same Shanghai researchers who built the training set also warned that heavy reliance on distillation could stifle original innovation and leave developers dependent on breakthroughs at more advanced labs.
How did Washington respond?
US Commerce Department officials raised the issue with Chinese regulators at meetings in Chengdu in July. Three people familiar with the discussions told Reuters the officials argued some collection of US model output could amount to intellectual-property theft. Chinese officials defended distillation as a standard technique, said they were open to acting where conduct was IP theft, and made no concrete commitments, according to a US official. US authorities had publicly accused Chinese companies of industrial-scale distillation in September.
A US official told Reuters that Washington believes industrial-scale distillation is the main reason Chinese models have narrowed the gap with American rivals. Two people familiar with Chinese developers' operations disputed that claim, crediting advances in engineering, algorithms, data and model design alongside distillation. The Commerce Department and China's Cyberspace Administration did not respond to Reuters.
The report landed one day after Washington moved to further restrict China's access to advanced chips from companies such as Nvidia, and ahead of the Nov. 18-19 APEC summit in Shenzhen, where President Donald Trump and Chinese leader Xi Jinping are expected to meet.
Does Beijing encourage it?
Chinese government documents do not tell companies to distil American models. A national AI+Manufacturing policy package published in January by eight government agencies supports model training, fine-tuning and distillation for industrial applications. Shanghai and Shenzhen have encouraged techniques that cut computing requirements.
Chinese developers have asked the Ministry of Industry and Information Technology where ordinary use of foreign models ends and organized collection of output begins, one person familiar said. The ministry did not respond.
What have OpenAI and Anthropic said?
OpenAI's terms restrict using its output to develop competing models, with exceptions, and Anthropic restricts training competing systems on Claude without permission.
Anthropic's Feb. 23 statement attributed more than 13 million Claude exchanges to MiniMax, more than 3.4 million to Moonshot AI and more than 150,000 to DeepSeek. Anthropic argued such campaigns undermine chip export controls and "reinforce the rationale" for them.
Ten days earlier, OpenAI told the House Select Committee on the Chinese Communist Party that DeepSeek was engaged in "ongoing efforts to free-ride on the capabilities developed by OpenAI and other U.S. frontier labs," according to a memo seen by Reuters. The memo said accounts linked to DeepSeek employees tried to bypass access restrictions through obfuscated third-party routers. DeepSeek did not respond.
Distillation is not illegal and is used outside China too. It cannot reproduce a leading model's full capabilities or remove the need for advanced computing power. But as chip curbs tighten, expect the dispute over what counts as theft, and what counts as standard practice, to shape the next round of export policy.
Source: mlq.ai
More from Sophie Lindqvist
Show full bio
News editor covering money, funding and market shifts around Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,060 articles
Related articles
chinese-ai-labs-buy-500m-in-u-s-data-despite-an-export-gap
Chinese AI Labs Buy $500M in U.S. Data Despite an Export Gap
researcher-finds-blackwell-fingerprints-in-china-s-top-open-ai-models
Researcher Finds Blackwell Fingerprints in China's Top Open AI Models
open-source-ai-platforms-race-to-become-china-s-hugging-face-ce6b57a2
Open-Source AI Platforms Race to Become China's Hugging Face
beijing-reads-u-s-ai-slowdown-push-as-chip-export-strategy-d46d3b6e
Beijing Reads U.S. AI Slowdown Push as Chip Export Strategy



