Anthropic claims that China's Alibaba used 25,000 fake accounts and 28.8 million exchanges to  illicitly 'distill' its Claude model — violations occurred from April to June 2026

Claude on an iPhone screen
(Image credit: Getty / NurPhoto)
7

Anthropic, the American AI lab that created one of the most advanced large language models available today, has said in a letter to the U.S. Senate that Chinese AI tech giant Alibaba has illicitly used Claude to train its own models. According to Reuters, the company sent the letter to Sen. Tim Scott (R-S.C.), the chair of the U.S. Senate Banking Committee, and Sen. Elizabeth Warren (D-Mass.), the ranking member of the said committee, before a scheduled hearing set to tackle AI issues.

This isn’t the first time that Anthropic has accused Chinese AI labs of “stealing” the capabilities of its AI model to train their own. Earlier this year, the company claimed that DeepSeek, Moonshot, and MiniMax used 24,000 fraudulent accounts and made 16 million exchanges collectively to train their own AI LLMs on Claude’s output. This method of training AI using the output of a more advanced model is called distillation, and while there are legitimate uses of this technique, such as when a frontier AI model is distilled to create a lighter, cheaper version of itself, it argues that competing labs can also use the same technique to build their own models at “a fraction of the time, and at a fraction of the cost, that it would take to develop them independently.”

Latest Videos FromTom's Hardware
Jowi Morales
Contributing Writer

Jowi Morales is a tech enthusiast with years of experience working in the industry. He’s been writing with several tech publications since 2021, where he’s been interested in tech hardware and consumer electronics.

  • hotaru251
    has illicitly used Claude to train its own models.

    says the company (and every other ai corpo) that's entire product is based on illicitly stolen work from everyone in world?
    Reply
  • cvzalez
    Oh my, seems like desperate move from Antropic, as if AI is a secrete, how it works is all over scientific papers and college computer science, it's probability, statistics, and linear álgebra on stereoids, yes it uses a lot of energy and it has to be developed from people that actually studied science!
    Reply
  • bigdragon
    Anthropic... isn't that the same company that has been going around saying AI needs hundreds of billions in new revenues or the world economy will implode?

    This is an excellent opportunity to stop and think about whether or not LLMs are the best way forward. They're a haven for stolen work and apparently not that hard to derivatively steal from. They're also wildly unreliable. Should we really keep this gravy train going or stop it before it goes flying off a cliff?

    Yeah yeah, I know the answer: kick the can down the road, line goes up, party like there's no tomorrow!
    Reply
  • usertests
    hotaru251 said:
    says the company (and every other ai corpo) that's entire product is based on illicitly stolen work from everyone in world?
    I support the "stealing", but on all sides. If China or anybody else can distill some sweet LLMs, they should.
    Reply
  • jp7189
    Anthropic is the only major player I'm aware of that doesn't produce an open weight model. I'm not a fan of their practices.
    Reply
  • rooted
    Hello my name is Pot, hi my name is Kettle.
    Reply
  • bit_user
    cvzalez said:
    Oh my, seems like desperate move from Antropic, as if AI is a secrete, how it works is all over scientific papers and college computer science, it's probability, statistics, and linear álgebra on stereoids, yes it uses a lot of energy and it has to be developed from people that actually studied science!
    I think you're missing a key detail, which is the training step. This is the most compute-intensive and time-consuming part of developing an AI model. It also requires vast amounts of data that's orders of magnitude bigger than the model, and that training data needs to be prepared in various ways.

    So, what Anthropic is alleging they did was to train their model by querying Claude, rather than going through the same process of gathering, filtering, and preparing their own training data. Not only does it avoid that extra work, but you also don't need the same ratio of raw data to weights as you do when training a model from scratch.

    It's similar to copying the answers off of somebody else's test, rather than studying the material and learning it for yourself.
    Reply