Chinese AI developers DeepSeek and Moonshot AI relied heavily on outputs from OpenAI’s GPT-6 Astra when training recent models, according to a report by The Information cited by Reuters. The report said Moonshot’s Kimi K3.1 Pro and DeepSeek’s V4.1-Flash 1017 benefited from distillation using Astra outputs. OpenAI, DeepSeek and Moonshot did not immediately provide detailed public responses to the specific claims.
What distillation means
Model distillation is a technique in which a smaller or different model learns from the outputs of a more capable model. It can be used legitimately within a company to compress a model or reduce deployment costs.
The controversy arises when one company’s system is used at scale to reproduce capabilities in a rival model without permission.
Why the allegation matters
Frontier AI systems cost billions of dollars to train. If competitors can obtain large volumes of high-quality outputs and use them as training data, they may be able to improve their models at much lower cost.
That creates economic and strategic tension between leading AI companies.
OpenAI has raised similar concerns before
OpenAI said in a September blog post that it had linked part of a distillation campaign to individuals associated with Moonshot AI.
The company and other U.S. AI developers have been investing in systems designed to detect unusually large or structured output harvesting.
DeepSeek has already changed the cost debate
DeepSeek’s earlier models drew attention for delivering strong performance with lower reported training costs than many Western systems.
That intensified questions about how much of the AI race depends on raw compute spending versus data quality, training techniques and model distillation.
Moonshot is also a major Chinese contender
Moonshot AI, known for its Kimi family of models, has become one of China’s most visible generative AI developers.
Competition among Chinese firms has accelerated as companies race to match or exceed U.S. frontier models.
The U.S.-China technology rivalry is part of the context
Washington has restricted exports of advanced AI chips to China, while Chinese companies are trying to improve model performance with limited access to the most powerful hardware.
Distillation can become more attractive when compute resources are constrained.
Detection is becoming a new security field
AI companies are increasingly treating model-output harvesting as a form of intellectual-property or platform security risk.
Anthropic said earlier this month that new defenses had reduced data collected through suspected distillation attacks by 90% to 99%.
Legal boundaries remain unsettled
The law around training on AI-generated outputs is still developing. Terms of service may prohibit automated extraction, but broader intellectual-property questions remain unresolved.
Future court cases and regulation could determine how much protection model providers can claim over generated outputs.
Users may see tighter access controls
If companies believe competitors are harvesting outputs at scale, they may impose stricter rate limits, identity verification or monitoring on developer APIs.
That could increase friction for legitimate researchers and startups as well.
The report does not establish every detail
The Information’s account is based on people described as having direct knowledge, but the companies have not publicly confirmed the full methodology or scale described.
The most defensible conclusion is that the allegation has intensified an already active dispute over distillation, access to frontier models and the economics of AI competition.
Sources
- Reuters, Oct. 29, 2026
- The Information report as cited by Reuters
- OpenAI September 2026 public statement on distillation activity
Another unresolved issue is whether model providers can technically distinguish ordinary customer use from training-oriented harvesting. Large volumes, repeated prompt structures and automated querying can provide clues, but sophisticated actors can distribute activity across accounts and networks. Better detection may therefore require both technical monitoring and stronger contractual enforcement.
Policy makers may eventually enter the debate
Governments are already concerned about AI intellectual property, national security and access to advanced computing. If distillation becomes a major route for transferring frontier capabilities across borders, regulators could consider disclosure requirements, export-control implications or rules governing automated access to advanced models.
Any intervention would be difficult to design because distillation is also a legitimate machine-learning technique used throughout the industry. Policymakers would have to distinguish ordinary model development from unauthorized extraction without blocking beneficial research.
The commercial consequences could extend beyond the companies named in the report. Cloud providers, API platforms and enterprise AI vendors may all tighten monitoring if they believe model outputs are being systematically harvested for training. At the same time, more aggressive restrictions could make it harder for smaller developers to build legitimate products. The industry is therefore trying to protect high-cost intellectual assets without closing the ecosystem that helped generative AI spread so quickly.
