On September 8, 2026, US intelligence agencies, including the NSA, CISA, and FBI, reported that certain Chinese AI models appear to have been trained using knowledge distillation. This method involves transferring the behavior of larger models to smaller ones. The findings indicate that models from developers such as DeepSeek, Alibaba Group, Moonshot AI, MiniMax, and StepFun may have extracted data from prominent Western AI systems, including Anthropic's Claude, OpenAI's GPT, and Google's Gemini.
US Authorities Report Chinese AI Models Used Knowledge Distillation to Extract Data from Western Models
The report suggests that these techniques have enabled the rapid development of high-performance models, such as DeepSeek's R1 and Moonshot AI's Kimi series, which demonstrate capabilities comparable to their Western counterparts. For instance, DeepSeek is reportedly able to access and utilize data from Claude and GPT models via API access. The intelligence community highlighted that leveraging existing advanced models through these methods facilitates the swift advancement of powerful AI capabilities.
Sources
- 「中国AIがClaudeやGPTから数十億トークン抽出」 米当局が暴いた「知識蒸留」の実態 （ITmedia AI+, 2026-09-17）
- Pangram – AI detector for text and images （Hacker News Frontpage, 2026-09-17）