Anthropic's accusation against Alibaba is current, specific, and serious, but the story has to be framed as an allegation, not a proven theft case.
Anthropic is not complaining about a few scraped prompts. It says operators linked to Alibaba's Qwen lab used almost 25,000 fake accounts to run 28.8 million exchanges with Claude from April 22 to June 5, 2026, then used those outputs to pull advanced capabilities into Alibaba's own models. If that account is right, this is industrial copying dressed up as AI development.
The claim appears in a June 10 letter to Senate Banking Committee Chair Tim Scott and Ranking Member Elizabeth Warren. According to Bloomberg's reporting on the letter, Anthropic called the activity the largest known distillation attack on the company so far and said the campaign targeted Claude's agentic reasoning, software engineering, and long-horizon task abilities. Business Insider also reported that the letter was signed by Sarah Heck, Anthropic's head of policy.
Also read: California built a real-time dashboard to track whether AI is actually killing jobs • China's AI and chip companies are flooding domestic markets with IPOs as the door to New York stays shut • A single Broadcom guidance miss sent South Korea's stock market into circuit-breaker territory and exposed how much the AI trade depends on perfection
You don't need to romanticize frontier AI research to see why Anthropic is angry. Training a leading model is expensive, slow, and full of dead ends. A rival that can fire millions of questions at Claude and train on the answers is not just saving money. It is trying to turn somebody else's research bill into its own shortcut.
That is the point Anthropic wants Washington to absorb. Distillation is not always dirty. In ordinary machine learning, it can be a legitimate way to train a smaller model from a larger one. The problem comes when a company allegedly creates fake accounts, hides the scale of its use, and extracts responses from a model it isn't supposed to access for that purpose. At that point, the nice technical term starts doing too much polite work.
Why the accusation lands in Washington
The Alibaba name matters. Qwen is not a side project from an unknown lab. It is Alibaba Cloud's main large language model family, and it has become one of China's most visible AI efforts. That makes Anthropic's letter more than a contract dispute over Claude's terms of service. It drops directly into the U.S. argument over whether Chinese AI labs are narrowing the gap through better engineering, looser rules, American infrastructure, or some mix of all three.
Frankly, this is the uncomfortable part for U.S. policymakers. Export controls on Nvidia chips are meant to slow the hardest part of building frontier systems, but model access can become another supply line. If a lab can use proxy accounts to harvest high-quality outputs from a U.S. model, the border around compute looks a lot less solid than it does in a policy memo.
Anthropic has made this argument before. In February, the company accused Chinese developers including DeepSeek, Moonshot, and MiniMax of using 24,000 fraudulent accounts and more than 16 million Claude exchanges in earlier distillation campaigns, according to reports at the time. The Alibaba allegation is larger by Anthropic's own numbers, and it arrives after months of U.S. anxiety over Chinese models that appear to be improving quickly while spending less than American rivals.
Alibaba has not publicly answered the allegation in the reports reviewed for this audit. That absence matters. A company response could narrow the dispute, challenge Anthropic's account, or explain whether the activity was carried out by employees, contractors, researchers, or unrelated operators. Without it, the cleanest way to write the story is still the honest one: Anthropic alleges the campaign, and the numbers come from Anthropic's letter as reported by Bloomberg, Business Insider, and The Wall Street Journal.
The stock market noticed, but the share-price move is not the main story. Investor's Business Daily reported that Alibaba's U.S.-listed shares fell more than 3% on June 25 after the Bloomberg report and closed at a 52-week low. Useful detail, yes. But the bigger issue is whether model companies can actually police industrial-scale extraction when the product is a conversation box built to answer questions all day.
If Anthropic is right, the weak point is not only Alibaba. It is the whole access layer around frontier models: accounts, APIs, resellers, cloud routes, and the ability to split one large campaign across thousands of small-looking users. You can ban a company from direct access and still leave plenty of doors ajar.
That is why this story is still live. The alleged exchanges ended on June 5, the Senate letter is dated June 10, and the first major reports appeared on June 25. The next question is not whether distillation exists. Everyone in AI knows it does. The question is whether the companies selling frontier access can stop it at this scale before the next model release becomes training data for the next rival.
