Alibaba Extracted 28.8M Claude Interactions to Train Qwen Without Paying

Published by James Harris on June 25, 2026June 25, 2026

Anthropic has accused Alibaba of running the largest AI model extraction campaign it has ever encountered, claiming that operators linked to Alibaba and its AI division Qwen created nearly 25,000 fraudulent accounts and generated more than 28.8 million interactions with Claude between April 22 and June 5. The goal, according to Anthropic, was distillation: using Claude’s outputs to train a smaller model without paying for the underlying capability.

The scale here is not subtle. Anthropic had previously flagged similar behavior from DeepSeek (150,000 exchanges), Moonshot AI (3.4 million), and MiniMax (13 million), but the Alibaba figure dwarfs all of them combined. The company sent a letter to Senate Banking Committee leaders Tim Scott and Elizabeth Warren urging stronger AI export controls, better threat intelligence sharing with AI firms, and tougher intellectual property protections. The timing connects directly to a broader policy shift: in April the White House accused China of acquiring American AI intellectual property on an industrial scale, and days after Anthropic’s June 10 letter, the Commerce Department imposed new restrictions on Anthropic’s models over concerns about access by military users in countries of concern. The access question around Alibaba’s Qwen products has already been playing out in regulatory grey zones, where the line between commercial and state-adjacent use is rarely clean.

Twenty-eight point eight million interactions is not a research experiment. It is an industrial operation.

The Anthropic allegations land alongside China’s LineShine supercomputer reclaiming the No. 1 position in the Top500 rankings, its first since 2017, built without US chips. That detail matters because it reframes the policy debate: export controls on chips and model access are the same lever, and both are now visibly being routed around. For the companies collecting infrastructure rent regardless of which lab wins the model race, the distillation dynamic is actually a threat multiplier, since it compresses the gap between frontier capability and commodity reproduction faster than any single model launch could. Anthropic’s appeal to Congress is partly self-interested, but the underlying concern is structurally sound: if distillation at this scale works, the moat around any proprietary model narrows considerably.

Anthropic’s letter to the Senate Banking Committee is now on the record, and the Commerce Department’s post-June 10 restrictions suggest at least some of the policy ask has already found traction. Whether Congress moves on the IP protection piece, the harder legislative lift, will determine whether this becomes a precedent or a data point.