Chinese AI giants DeepSeek, Moonshot used OpenAI's GPT-6 Astra to train their models: Report

Chinese AI companies DeepSeek and Moonshot AI reportedly utilized outputs from OpenAI's GPT-6 Astra model to train their most recent AI systems, according to a report by The Information. Multiple sources familiar with the matter informed The Information of this extensive use of Astra's outputs for model training. AI distillation, a technique that allows developers to learn from advanced AI models, has been a significant factor in the progress of Chinese open-weight models despite criticism from OpenAI and Anthropic.

The Trump administration had previously prohibited civilian U.S. government agencies from utilizing products and services from DeepSeek due to security concerns. The practice of AI distillation has raised concerns among major tech companies as the U.S. and China compete intensively. Adversarial distillation, specifically, is the unauthorized use of another model's outputs or reasoning to help train, reproduce, or improve another model.

OpenAI, Google, Anthropic, and other major AI developers have warned that this could enable competitors to replicate their advanced models' capabilities without investing similar resources. In a recent blog post, OpenAI stated that they traced a significant portion of a distillation campaign, initiated in July, to individuals linked to Moonshot AI, the developer of Kimi.

OpenAI alleged that the operators manipulated model interactions to reproduce protected reasoning in a coordinated manner that violated their terms of service. The activity, which started on July 1, saw a surge in requests between July 24 and 25, involving more than 4,000 users. Subsequent investigation uncovered related activity across a network of over 15,000 users, which OpenAI successfully disrupted by July 28.