Anthropic Accuses Alibaba of Distillation Attack, U.S. AI Companies Seek Legislative Intervention
0xBroomberg
Anthropic wrote to U.S. senators alleging Alibaba used tens of thousands of unauthorized accounts to secretly distill its AI models, demanding Congress legislate against such practices. This means → the U.S.–China AI rivalry is escalating from a technology race into a legal and geopolitical confrontation.
What exactly is Anthropic alleging?
Anthropic wrote to Senators Tim Scott and Elizabeth Warren on June 10, accusing Alibaba of using tens of thousands of unauthorized accounts to secretly harvest its AI technology and train Alibaba's own models with the collected data.
Anthropic labeled the practice a "distillation attack," describing it as "illegal, systematic, industrial-scale theft of American AI capabilities."
The letter, obtained and published by *The New York Times*, explicitly asked both senators to address curbing Chinese distillation at an upcoming AI hearing.
What is "distillation" — and why is it so contentious?
Distillation is an AI technique over a decade old: extract data from a large, high-performance model to train a smaller one that runs on cheaper hardware. In plain terms = one model acts as the teacher, another as the student — the student acquires the teacher's core abilities in a smaller package.
The technique itself is legal. The dispute is about what gets distilled. Both Anthropic and OpenAI explicitly ban distillation of their flagship models in their terms of service, yet the practice remains widespread.
Elon Musk testified in a federal court in Oakland in April that his company xAI had distilled OpenAI's technology, adding that "AI companies routinely distill other AI companies' technology."
This means → distillation itself is not the issue. Unauthorized, industrial-scale distillation is what Anthropic wants legislation to target.
Can existing law address this?
No court has issued a clear ruling on AI distillation. Trade-secret litigator Sarah Tishler notes that some legal scholars believe it violates the 2016 Defend Trade Secrets Act, but the argument remains academic.
Copyright law is equally hard to apply — distillation copies system behavior, not verbatim text. In plain terms = the student learned the teacher's problem-solving method without copying the teacher's exact words, and current copyright law struggles to call that "infringement."
This reflects a deeper problem: AI technology is advancing faster than the legal framework can keep up — precisely why Anthropic is asking Congress to step in.
Why has the urgency spiked now?
Experts estimate China trails the U.S. in AI development by roughly six months. Chinese startup Zhipu AI recently released its GLM-5.2 model, which performs near the level of top U.S. systems in cybersecurity — a domain of acute geopolitical sensitivity for both U.S. AI companies and the Trump administration.
Anthropic and its peers argue that without distillation, the gap would be far wider than six months — a margin that matters across commercial planning, drug research, mass surveillance, and military weapons.
This means → the real driver behind the legislative push is not just intellectual-property protection but strategic anxiety over how fast China is closing the gap. Exactly how much distillation has contributed to that narrowing cannot be precisely quantified, but the question has already become a focal point in congressional legislative discussions.
Content is for reference only, not financial advice.