. 4 min read.

Anthropic accuses Alibaba of copying Claude or the biter bit

Anthropic says Alibaba queried Claude 28.8 million times through 25,000 fake accounts to train Qwen. The company that lifted the world's books cries theft.

I read Anthropic's letter with a little smirk, and I'll tell you why. The company behind Claude is furious, and fair enough. In a letter to US lawmakers and the White House, it accuses Chinese giant Alibaba of running the biggest raid ever on its models1. The technique has a polite name, distillation. You hammer a powerful model with questions, collect its answers, and use them to teach a cheaper rival to imitate it2. From Anthropic's side, that's years of research and billions of dollars gone in a few weeks. From a little further back, it's the biter bit.

Line-art illustration on a terracotta background, a stylized profile of a face reaching a hand toward a white circular motif that suggests an atom or an orbit.
An illustration by Anthropic, taken from its official site.

Twenty-five thousand fake accounts to copy Claude

The numbers are wild. Between April 22 and June 5, operators tied to Alibaba and its Qwen lab allegedly opened nearly 25,000 fake accounts and queried Claude some 28.8 million times. They covered their tracks using proxy servers, middlemen that hide where a connection comes from3. The questions went after the model's most valuable skills, coding and so-called agentic reasoning, its ability to handle long tasks on its own without anyone holding its hand1.

Alibaba then allegedly fed the haul into its own Qwen models. To Anthropic, that turns "hundreds of billions of dollars in American investment" into "a massive subsidy for our geopolitical competitors"2. Basically, someone spent six weeks quizzing your private tutor so they could open a school across the street.

When distillation becomes a matter of state

This is way past a business dispute. Anthropic took it to Congress and the White House. It wants penalties on Chinese labs, tighter controls on chip exports, and even an antitrust exception so AI companies can share intelligence on Chinese tactics3. The timing is no accident. The campaign allegedly ran in defiance of the administration's warnings, shortly after Washington locked down access to Fable 5 and Mythos 54.

Alibaba denies all of it. The group describes itself as a company run by an independent board with no military ties, and it's separately suing to get off a US blacklist2. It's also the first time Anthropic has gone after a Chinese heavyweight, after calling out smaller players like DeepSeek and Moonshot back in February1.

The biter, the bitten, and everyone else

And here's where my smirk comes in. At the very moment Anthropic is denouncing theft that's "illicit, systematic, and at industrial scale," it's fighting several lawsuits accusing it of training Claude on copyrighted work without asking anyone. Three music publishers, Universal, Concord, and ABKCO, want some $3 billion for more than 20,000 songs used without a license. A fourth, BMG, flat out argues that Anthropic's entire value was built on stolen work5. Anthropic pleads fair use, the US legal doctrine that allows some borrowing without permission6. Alibaba could hand it the exact same argument, since distilling a model really just means reading very carefully whatever it's willing to tell you.

I get Anthropic's anger. Watching a competitor vacuum up your work would make any creator sick. Except the industry now crying theft is the same one that harvested books, articles, and songs by the million to build its models, and keeps doing it while swearing it's progress. You can't treat copying as a mortal sin when it targets you and a technicality when it works in your favor.

That doesn't make everyone equally guilty. A state running an industrial-scale heist is not a poet borrowing a line. But the next time a big model plays the victim, we're allowed to ask where exactly the data that made it so valuable came from.

Notes

  1. Tom's Hardware, "Anthropic claims that China's Alibaba illicitly distilled its models from April to June 2026", June 2026. ↩ ↩2 ↩3

  2. Dawn, "Anthropic accuses Alibaba of mass AI capability theft", June 2026. ↩ ↩2 ↩3

  3. Slashdot, "Anthropic Says Alibaba Must Be Punished For Largest Claude Cloning Attack", June 2026. ↩ ↩2

  4. Ars Technica, "Anthropic claims Alibaba defied Trump to attack Claude and steal capabilities", June 2026. ↩

  5. Music Business Worldwide, "Anthropic, fighting lawsuits over alleged copying of song lyrics, accuses Alibaba of copying Claude", June 2026. ↩

  6. Judges weigh the purpose of the borrowing, how much was taken, and the harm done to the original. Generative AI is currently testing how far the doctrine stretches before it snaps. ↩

Read next…

  1. Amodei wants to pace the frontier without easing off the gas

  2. Claude will sign everything it writes

  3. Europe regulates the AI it doesn't build