Reports Indicate Chinese AI Firms Utilized OpenAI Outputs to Train Systems

NEW DELHI — Chinese artificial intelligence firms DeepSeek and Moonshot AI have reportedly utilised outputs from OpenAI’s advanced GPT-6 Astra model to train their own latest AI systems.

The development, detailed in a report by The Information, relies on sources familiar with the matter and highlights ongoing tensions surrounding the cross-border transfer of artificial intelligence capabilities. The allegations renew industry-wide anxieties concerning AI distillation, a technique that allows developers to leverage outputs from sophisticated frontier models to replicate advanced system functionalities more rapidly and economically.

The practice sits at the center of escalating technological competition between the United States and China. Major western developers, including OpenAI, Anthropic, and Google, have consistently criticized distillation, arguing it enables competitors to bypass massive research and infrastructure expenditures. OpenAI defines adversarial distillation as the systematic and unauthorized use of one model’s outputs or reasoning to train, reproduce, or improve a competing system.

Regulatory and security scrutiny has intensified alongside these technical disputes. Earlier, the administration of US President Donald Trump barred civilian US government agencies from utilizing products and services offered by DeepSeek, citing potential national security risks tied to the startup’s operations in China.

Industry countermeasures are also expanding. Earlier this month, Anthropic announced the implementation of new safeguards designed to curb the unauthorized replication of its Claude AI models. According to the company, these measures reduced the volume of data gathered through suspected distillation attacks by roughly 90% to 99%.

Specific enforcement actions have already surfaced regarding Moonshot AI, the developer behind the Kimi system. In a blog post published on September 30, OpenAI stated it had traced a major distillation campaign beginning July 1 to individuals linked to Moonshot AI.

OpenAI alleged that the operators systematically manipulated model interactions to render protected reasoning visible to the requester in a coordinated, scaled manner that violated terms of service. According to OpenAI’s findings, the activity surged from initial operations to 16,000 requests on July 24 and 25, involving over 4,000 users. Subsequent investigations uncovered related operations across a network exceeding 15,000 users before the network was fully disrupted on July 28.