In September 2026, Anthropic publicly accused three prominent Chinese AI firms — DeepSeek, Moonshot AI, and Alibaba — of covertly routing user traffic to its Claude model at industrial scale, harvesting outputs to train rival systems in a practice known as distillation. The allegation places a familiar technical method in unfamiliar moral territory, raising questions not merely about intellectual property but about the norms that will govern AI development when competitive pressure outpaces shared rules. What emerges is less a story about one company's grievance than a mirror held up to a worl
Y Combinator's Tan pushes US labs to 'distill' frontier AI models amid Chinese theft claims
If Chinese competitors are adopting these tactics, US companies should not unilaterally disarm.
Mark
So Anthropic is saying Chinese companies used Claude to train their own models without permission. Is that actually illegal, or is it more of an ethics question?
Mimi
It's genuinely murky. Distillation itself is a standard technique—you can learn from a model's outputs. The problem Anthropic identifies is the scale and the deception. They're alleging industrial-scale diversion of user traffic, which suggests intentional extraction rather than incidental learning.
Luke
But we should be careful here. Anthropic is making the allegation, and they have incentive to frame it as theft. We don't have independent verification of the scale, the intent, or even whether the traffic diversion actually happened. The companies deny it.
Mimi
That's fair. What we do know is that Garry Tan responded by essentially saying US labs should do the same thing. That's a telling reaction—it suggests the practice is real enough that American leaders are already thinking about how to compete on those terms.
Mark
Why would Tan say that publicly? Isn't that admitting the whole thing is happening?
Luke
He's not admitting anything about the Chinese claims. He's making a separate argument: if this is how competition works now, American companies shouldn't be the only ones following restraint. It's a competitive pressure argument, not a confirmation.
Mimi
And it reveals something important about how American AI leaders are thinking. They're not confident that norms or regulation will constrain Chinese behavior, so they're already mentally preparing for a world where everyone distills everyone else's models.
Mark
What about the Africa deal? How does that fit in?
Luke
It's the US trying to establish alternative partnerships before China does. If you can't win the technology race outright, you win by controlling who gets access to your technology. Africa is a huge emerging market.
Mimi
Exactly. The real competition isn't just about whose model is better. It's about whose model becomes the standard, whose company becomes the trusted partner, whose country's technology infrastructure becomes embedded in the developing world.
Mark
So this distillation thing is just one front in a much larger conflict.
Mimi
Yes. And the distillation allegations are partly about the technology, but they're also about establishing a narrative—America saying China plays dirty, China saying America is trying to monopolize AI. Both sides are building the case for why their approach is justified.
The Pulse
- Anthropic's accusation that Chinese firms secretly treated Claude as a free training resource has shattered any remaining pretense of genteel competition between American and Chinese AI developers.
- The allegations ripple outward fast: Y Combinator's Garry Tan publicly urges US open-weight labs to pursue the same distillation tactics, signaling that the ethical high ground may be the first casualty of the race.
- China's government frames the charges as protectionist theater, turning the dispute into a geopolitical flashpoint rather than a technical or legal matter.
- The US State Department counters on a different front, announcing a one-billion-dollar AI partnership initiative in Africa to contest Chinese influence in emerging markets.
- The controversy is landing in a space where law, ethics, and strategy have not yet converged — leaving the industry to navigate a race in which the rules are being written by whoever moves fastest.
In September 2026, Anthropic publicly accused three prominent Chinese AI firms — DeepSeek, Moonshot AI, and Alibaba — of covertly routing user traffic to its Claude model at industrial scale, harvesting outputs to train rival systems in a practice known as distillation. The allegation places a familiar technical method in unfamiliar moral territory, raising questions not merely about intellectual property but about the norms that will govern AI development when competitive pressure outpaces shared rules. What emerges is less a story about one company's grievance than a mirror held up to a world in which every actor risks becoming what it accuses the other of being.
Anthropic, the safety-focused company behind Claude, went public in September 2026 with a striking accusation: DeepSeek, Moonshot AI, and Alibaba had been secretly funneling user requests to Claude on an industrial scale, capturing the model's outputs to train their own competing systems. The technique — known as distillation — is a legitimate practice in machine learning, but Anthropic's complaint centers on the covert, large-scale nature of the alleged campaigns, which it characterizes as intellectual property extraction dressed in the language of ordinary use.
The disclosure landed with force across Silicon Valley and Washington. Garry Tan of Y Combinator responded not with condemnation but with a call to arms, urging American open-weight AI labs to pursue distillation themselves. The implicit logic: if rivals are operating without restraint, unilateral American restraint is simply surrender. The comment crystallized an anxiety that has been building among US AI leaders — that ethical commitments, however sincere, may become competitive liabilities in a race with no agreed-upon rules.
China's government rejected the theft framing entirely, casting the allegations as the latest instrument of American technological containment. The dispute quickly entangled itself with broader geopolitical maneuvering: the US State Department announced a one-billion-dollar AI partnership initiative targeting Africa, positioning American technology as the more trustworthy alternative for developing nations navigating between the two powers.
What the distillation controversy ultimately exposes is a three-way collision — between what is technically permissible, what is competitively rational, and what the international community might eventually recognize as legitimate conduct. Anthropic's public disclosure appears calibrated to build a reputational record and apply pressure, but it has also made plain that the frontier of AI competition has shifted. It is no longer decided primarily by research breakthroughs or engineering talent, but by control over data, access to compute, and the willingness to inhabit the ambiguous territory between law and norm.
Anthropic, the AI safety company behind Claude, has accused three major Chinese artificial intelligence firms of systematically extracting training data from its frontier models through a technique called distillation. The companies in question—DeepSeek, Moonshot AI, and Alibaba—allegedly routed user requests to Claude on an industrial scale, harvesting the outputs to train their own competing systems. The allegations surfaced publicly in September 2026, marking an escalation in the already tense competition between American and Chinese AI developers over access to cutting-edge technology and market dominance.
Distillation is a legitimate machine learning practice in which a smaller or more efficient model learns from a larger one. The technique itself is not new or inherently unethical. What distinguishes Anthropic's complaint is the scale and the secrecy: the company contends that these Chinese firms diverted user traffic to Claude without disclosure, treating the frontier model as a free training resource rather than a service to be used transparently. The practice, if confirmed, would represent a form of intellectual property extraction that sits in a gray zone between competitive intelligence and outright theft.
The allegations have reverberated through Silicon Valley and Washington. Garry Tan, the chief executive of Y Combinator, the influential startup accelerator, responded by calling for American open-weight AI labs to pursue distillation strategies of their own. His argument, implicit in the framing, suggests that if Chinese competitors are adopting these tactics, US companies should not unilaterally disarm. The comment reflects a broader anxiety among American AI leaders that restraint in competitive practices could cede advantage to rivals operating under different regulatory and ethical frameworks.
China's government and the companies involved have pushed back against the characterization of their activities as theft. Officials in Beijing have framed the American allegations as part of a broader campaign to contain Chinese technological advancement, pointing to what they describe as protectionist rhetoric from Washington. The dispute has become entangled with geopolitical competition for influence in emerging markets. The US State Department, meanwhile, announced a one-billion-dollar AI partnership initiative focused on Africa, positioning itself as a more reliable technology partner than China for developing nations seeking to build their own AI capabilities.
The distillation controversy sits at the intersection of three overlapping tensions: the technical question of what constitutes fair use of a commercial AI service, the competitive question of whether American companies should adopt practices they publicly criticize, and the geopolitical question of whether AI development will be governed by shared norms or by a race to the bottom in which each side justifies its tactics as responses to the other's aggression. Anthropic's public disclosure of the alleged campaigns appears designed to establish a record and apply reputational pressure, but it has also crystallized the reality that the frontier of AI competition is no longer primarily about research breakthroughs or engineering talent. It is increasingly about control over data, access to compute, and the willingness to operate in the spaces between legality and norms.
Notable Quotes
Garry Tan suggested that if Chinese competitors are adopting distillation tactics, American companies should not unilaterally refrain from similar practices.— Y Combinator CEO Garry Tan
Want the full story?
Read the original at Google News ↗
Technology