Anthropic has accused Alibaba's Qwen research lab of running what it described as the largest known industrial-scale distillation campaign ever detected against its Claude AI models, using thousands of fraudulent accounts to siphon off capabilities at scale.

In a letter disclosed this week to the Senate Banking Committee, Anthropic alleged that operators tied to Alibaba created approximately 25,000 fake accounts on Claude's platform and ran 28.8 million conversations with the model between April 22 and June 5. According to the company, the effort was narrowly targeted at Claude's highest-value frontier capabilities—advanced software engineering and multi-step agentic reasoning—rather than commodity tasks where Chinese models already compete.

AI distillation works by using a cheaper model to imitate a more powerful one: queries are sent to the leading system, answers are harvested, and a copycat model is trained to approximate the results. Anthropic's letter frames the alleged campaign as an attempt to turn hundreds of billions of dollars in U.S. AI investment into a subsidy for a geopolitical competitor, and warns that similar tactics could become a persistent feature of the competitive landscape.

Alibaba has not responded publicly to the specific claims, and no independent third party has verified the allegations. The disclosure comes at a moment of heightened scrutiny of Chinese AI labs, which have repeatedly been accused of accessing Western models through unofficial channels. In January 2025, for instance, the release of DeepSeek's open-source model triggered a brief market panic that erased roughly $589 billion from Nvidia's market value in a single session—a reaction many analysts later judged to be disproportionate.

Economists and AI industry observers say the latest accusations should be read with similar calibration. Distillation copies only a static snapshot of a model that continues advancing after the imitation is trained, and the copied system inherits any blind spots or limitations of the original. “When your fastest path to a frontier model is copying the leader's answers at industrial scale, the exercise is an admission of where you actually stand,” one analyst noted in a recent industry analysis of the case.

For investors, the episode underscores that U.S. frontier labs still set the global pace on capability development, even as global compute supply remains constrained and demand for training infrastructure continues to outstrip availability. The gap between leading models and their imitators, researchers argue, is structural rather than temporary.

Image source: v3b.fal.media