Anthropic says Alibaba ran a months-long covert operation to harvest Claude's outputs and use them to train competing AI models.
The American AI lab disclosed the scheme in a letter to Sen. Tim Scott (R-S.C.) and Sen. Elizabeth Warren (D-Mass.), both members of the Senate Banking Committee, ahead of a scheduled hearing on AI. Anthropic traced the distillation campaign — run from April to June 2026 — to operators connected to Alibaba and its AI division Alibaba Qwen. The scheme involved roughly 25,000 fake accounts and 28.8 million exchanges on Claude's platform. Distillation, in this context, means using a more capable model's outputs to cheaply train a competing one — a technique that has legitimate uses but that Anthropic argues is being exploited to shortcut years of independent research.
This is Anthropic's second such accusation in a matter of months. Earlier this year, the company alleged that DeepSeek, Moonshot, and MiniMax collectively used 24,000 fraudulent accounts and 16 million Claude exchanges in a similar scheme. Anthropic warns that campaigns like this could help Chinese labs build models with capabilities approaching those of frontier American AI like Mythos Preview — at a fraction of the time and cost of developing them from scratch.
The gap is narrowing regardless: Elon Musk projected a Chinese lab could reach Fable 5-class capability by early next year, and the CEO of Z.ai publicly pushed that deadline even sooner. Whether Alibaba needed Claude's help to get there is the question Congress has now been asked to take seriously.