AI Model Distillation Escalates as U.S. Labs Uncover Massive Chinese Extraction Campaigns

Recent disclosures from Anthropic and OpenAI reveal a surge in illicit model‑distillation efforts by Chinese AI firms. Both companies reported shutting down campaigns that siphoned billions of prompts to recreate their flagship models at a fraction of the original cost, threatening the competitive edge of U.S.

The practice, first highlighted in early 2025 when DeepSeek claimed to train a capable model for under $300,000, involves extracting model outputs—sometimes even encrypted reasoning—to train smaller “student” models.Key developments include Anthropic’s detect… Researchers have shown that model reasoning can be subtly altered to degrade the quality of stolen training data without hurting performance, offering a possible defensive tactic.

However, large‑scale operations persist: Anthropic traced nearly 200 million exchanges, including a 3 million‑queries‑per‑day Alibaba‑linked effort and Moonshot’s harvest of live customer chats.Why it matters: Distillation lets rivals replicate expensive fron…