---
type: "post"
id: "chinese-companies-and-researchers-are-using-answers-from-5a08f8"
slug: "chinese-companies-and-researchers-are-using-answers-from-5a08f8"
title: "Chinese companies and researchers are using answers from American AI models to train their own."
description: "Reuters reports that Chinese researchers and companies are increasingly using outputs from leading US models to train smaller domestic systems, a practice known as distillation."
retrieval_nugget: "Reuters reports that Chinese researchers and companies are increasingly using outputs from leading US models to train smaller domestic systems, a practice known as distillation. The shift matters because it turns costly model answers into reusable training data, while intensifying questions about export controls, competition, and the terms governing model outputs."
published_at: "2026-11-17"
updated_at: "2026-11-17"
record_date: "2026-11-17"
date_kind: "published_at"
topics: ["anthropic"]
entities: ["Reuters","OpenAI","Anthropic","Alibaba","Qwen2.5-72B","Meituan","Shanghai Jiao Tong University"]
related_patterns: []
novelty_label: "new_fact"
what_changed: "A Reuters review of more than 300 Chinese records finds that distillation of American AI model outputs has spread beyond academic experiments into mainstream industrial use."
prior_refs: []
source_url: "https://www.reuters.com/world/china/china-taps-us-ai-models-train-homegrown-rivals-chip-curbs-bite-2026-11-17"
source_urls: ["https://www.reuters.com/world/china/china-taps-us-ai-models-train-homegrown-rivals-chip-curbs-bite-2026-11-17"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:chinese-companies-and-researchers-are-using-answers-from-5a08f8"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/chinese-companies-and-researchers-are-using-answers-from-5a08f8/","markdown":"https://newruntime.com/posts/chinese-companies-and-researchers-are-using-answers-from-5a08f8.md","json":"https://newruntime.com/posts/chinese-companies-and-researchers-are-using-answers-from-5a08f8.json"}
markdown_url: "https://newruntime.com/posts/chinese-companies-and-researchers-are-using-answers-from-5a08f8.md"
---

# Chinese companies and researchers are using answers from American AI models to train their own.

## Retrieval answer

Reuters reports that Chinese researchers and companies are increasingly using outputs from leading US models to train smaller domestic systems, a practice known as distillation. The shift matters because it turns costly model answers into reusable training data, while intensifying questions about export controls, competition, and the terms governing model outputs.

## Summary
Reuters reports that Chinese researchers and companies are increasingly using outputs from leading US models to train smaller domestic systems, a practice known as distillation. The shift matters because it turns costly model answers into reusable training data, while intensifying questions about export controls, competition, and the terms governing model outputs.

## What happened
Reuters published an investigation on November 17, 2026, based on a review of more than 300 Chinese academic papers, patents, corporate disclosures, recruitment notices, and government documents. It says Chinese use of distillation has expanded beyond academic experiments into mainstream industrial activity. Researchers use US models to generate answers, code, and solutions; companies have developed software to automate the process, while specialist workers prepare material for training smaller domestic models.

Distillation uses a powerful “teacher” model to generate answers or demonstrate how to complete tasks. Developers then use those outputs to train a smaller “student” system. Reuters describes the economic appeal for teams constrained by computing power: a difficult question can be answered once by an expensive model, and the resulting example can be retained for training a local system rather than repeatedly querying the original provider. Distillation can preserve selected skills in smaller systems that cost less to operate. Reuters cautions that it cannot reproduce a leading model’s full capabilities or eliminate the need for advanced computing.

The article traces several examples. In a November 2024 paper, Shanghai Jiao Tong University researchers described selecting about 5,000 answers from OpenAI’s o1 model, translating them with GPT-4o mini, and assembling 10,750 bilingual training examples to further train Alibaba’s Qwen2.5-72B. In an April 2025 paper, Alibaba researchers described using GPT-4 and GPT-4o to generate training material in languages other than Chinese, then refining and scoring responses before training smaller Qwen models. A February 2026 paper by Shanghai Jiao Tong University and Meituan researchers described using GPT-4o to generate plans for coordinating AI agents on programming problems, helping train a smaller model to organize those tasks.

The investigation places these examples amid a policy dispute. Reuters says Chinese government documents encourage distillation as part of a wider push to spread AI, including a January national AI+Manufacturing policy package issued by eight agencies. Shanghai and Shenzhen have also encouraged techniques to reduce computing requirements. US officials raised concerns with Chinese regulators in July about collecting outputs from leading US models; Chinese officials defended distillation as a standard development technique while saying they were open to action where conduct constituted intellectual-property theft. Reuters says OpenAI’s terms restrict using its output to develop competing models, subject to exceptions, and Anthropic similarly restricts training competing systems without permission.

## What the archive adds
New Runtime’s earlier coverage found Chinese models already occupying a substantial place in a startup’s technology stack, based on Kilo Code telemetry, on [December 5, 2025](https://newruntime.com/posts/on-thursday-research-firm-semianalysis-published-a-report-5ef491/). This Reuters investigation adds a different fact: the reuse of US model outputs as training material is described as an expanding industrial practice, not simply a source of model choice. It is a new development in the broader story of China’s AI ecosystem.

## Trend signal
This does not yet fit a pattern New Runtime tracks. Reuters describes a consequential training and policy practice, but the evidence does not establish a durable shift in how agent systems, software boundaries, or production harnesses are designed.

## Why now
Reuters places the report ahead of the November 18–19 APEC summit in Shenzhen, where Donald Trump and Xi Jinping were expected to meet. It also reports that Washington announced further restrictions on China’s access to advanced chips on November 16, bringing distillation’s potential role in working around compute constraints into sharper focus.

## Connections to other events
New Runtime’s earlier coverage of Chinese model adoption described Kilo Code telemetry showing Chinese models taking a meaningful share of one startup’s stack ([December 5, 2025](https://newruntime.com/posts/on-thursday-research-firm-semianalysis-published-a-report-5ef491/)). Reuters now describes one mechanism behind the broader competitive picture: using outputs from more capable foreign models to develop smaller local systems. The two accounts concern different evidence and should not be treated as proof that distillation explains Chinese model progress overall.

## Why it matters
For teams building models, distillation makes capability transfer and data provenance operational concerns, not just research questions. A model output can become durable training material that reduces later inference costs, but the legal and contractual status of that reuse may differ by provider and use case. Teams need to distinguish authorized data collection from workflows that breach terms or rights, and preserve provenance so they can explain where training examples came from.

The same dynamic complicates strategic assumptions about compute controls. Distillation may help a team get more value from limited compute, but Reuters says it does not remove the need for advanced computing or reproduce a frontier model in full. Model builders and policymakers should therefore avoid treating it as either a complete substitute for compute or a trivial transfer mechanism.

## Limits and uncertainty
Reuters’ review establishes documented examples and reports broader adoption, but it does not quantify what share of Chinese model training uses distillation or how much capability improvement it causes. The article notes disagreement over its importance: a US official characterized it as the main reason Chinese models narrowed the gap, while two people familiar with Chinese developers’ operations cited engineering, algorithms, data, and model design as additional factors. The full extent of any resulting cost or performance advantage is unmeasured here. The cited institutions did not respond to Reuters’ requests for comment, and the article does not resolve how specific output uses fare under each provider’s terms or applicable law.

## Reader takeaway
Distillation can turn expensive model responses into reusable training data, but the capability transfer is partial and the right to reuse outputs is not automatic. For builders, treat provenance and provider terms as part of the training pipeline, alongside compute and model quality.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Anthropic](https://newruntime.com/topics/anthropic.md): Explore the anthropic topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
