Anthropic Just Pointed a Very Public Finger
I’ve been covering this space long enough to remember when « distillation » was a term you heard in ML research papers, not in accusations traded between billion-dollar companies. That era is over. On Thursday, Anthropic published a report alleging that three China-based AI outfits — Alibaba, Moonshot AI, and DeepSeek — have been running sustained distillation campaigns against its Claude models. And according to Anthropic, the pace has picked up in recent months as the competition got nastier.
Let that sink in for a second. This isn’t a subtle technical dispute. This is one of the biggest AI labs in the West essentially saying: our rivals are using our own model to train theirs, at scale, and they’re getting better at hiding it.
I’ll be honest — my first reaction was a shrug. Distillation has been part of the AI toolkit forever. Every grad student with a GPU has done it. But the scale and the intent described here are different, and the fact that Anthropic went public with names attached tells me something shifted internally.
What Distillation Actually Is (and Why It Matters)
If you’re new to this, here’s the short version. Model distillation is the practice of training a smaller, cheaper model to mimic the outputs of a larger, more capable one. You feed the big model prompts, collect its answers, and use those answers as training data. The student model learns to approximate the teacher without ever seeing the teacher’s weights.
It’s efficient. It’s clever. And depending on who you ask, it’s either a legitimate research technique or a polite word for theft.
The reason it matters now is economics. Frontier models cost hundreds of millions of dollars to train. Distillation lets you skip a big chunk of that bill. If you can get a strong model to cough up enough high-quality outputs, you can bootstrap a competitor for a fraction of the cost. That’s the accusation on the table.
What struck me reading Anthropic’s report is how methodical the alleged campaigns were described. This isn’t a handful of curious engineers poking at an API. The report describes persistent, coordinated activity — thousands of accounts, rotating infrastructure, queries engineered to extract reasoning traces rather than just answers.
The Three Names on the List
Alibaba, Moonshot AI, and DeepSeek. That’s not a random trio. It’s a who’s who of China’s most ambitious AI players, each with its own frontier ambitions and each under pressure to close the gap with OpenAI, Anthropic, and Google.
DeepSeek, in particular, has been the story of the last couple of years. Its models shocked Western labs with their efficiency and capability, and the company has been unusually open about its methods. Now it’s named in a distillation complaint. Whether that’s fair or just the cost of being successful in a paranoid market, I can’t say with certainty. But the optics are rough.
Moonshot AI has built a reputation on long-context models and consumer-facing products. Alibaba’s Qwen family is everywhere — powering startups, fine-tunes, and open-source projects around the world. If even a fraction of the allegations hold up, the downstream effects touch thousands of companies that built on top of these models without knowing where the training signal came from.
And here’s the uncomfortable part: we don’t have independent verification. Anthropic is the accuser, the investigator, and the judge in its own report. That doesn’t make the claims false. It just means we should read them like we’d read any company’s self-published findings — with interest, and with a raised eyebrow.
Why Now? Because the Money Is Getting Tight
Timing is everything in these stories. Anthropic didn’t drop this report in a slow news week by accident. The AI industry is entering a phase where the easy money is drying up, the compute bills are astronomical, and investors are asking harder questions about moats.
If you’re Anthropic, you want two things: regulators to take distillation seriously, and customers to believe your model is genuinely hard to replicate. A public report does both. It frames the company as the victim of unfair competition while quietly reinforcing that its models are the ones worth stealing.
That’s not cynicism. That’s strategy. And honestly, I’d do the same thing in their shoes.
The counterargument is that distillation is a gray area, not a clear crime. Terms of service violations? Sure. But turning a ToS dispute into a geopolitical narrative is a choice, and it’s one that could backfire if the evidence doesn’t hold up under scrutiny.
The Geopolitics Nobody Wants to Say Out Loud
Here’s the thing I keep coming back to. Every major AI story in 2026 eventually becomes a US-China story. Export controls, chip bans, talent flows, and now training data. The distillation allegations slot neatly into a broader narrative that Washington has been building for years: China’s AI progress is partly built on borrowed Western work.
Is that true? Partly, probably. Is it the whole story? Almost certainly not. Chinese labs have published genuinely novel research, built impressive infrastructure, and trained huge models from scratch. Reducing all of that to « they copied us » is lazy and, frankly, a little insulting to the engineers doing real work over there.
But I also think Western labs have a legitimate grievance when their outputs are scraped at industrial scale to train competitors. There’s a difference between learning from public research and systematically extracting a commercial model’s behavior. The line is blurry, but it exists.
What This Means for the Rest of Us
If you’re building on top of any of these models — Claude, Qwen, DeepSeek, whatever — this story matters to you more than you think. Here’s why.
First, expect stricter API monitoring. Labs are going to get more aggressive about detecting distillation patterns, and that means more rate limits, more verification, and potentially more false positives that catch legitimate users in the crossfire.
Second, expect legal action. Anthropic hasn’t filed a lawsuit yet, but the report reads like groundwork. If a case does land, it could set precedents that reshape how every AI company handles its outputs.
Third, expect the open-source conversation to get even messier. Open weights are a gift to the ecosystem and a headache for anyone trying to protect a business model. Distillation scandals pour gasoline on that debate.
- Stricter API monitoring and account verification across major labs
- Potential lawsuits that could redefine output ownership
- Renewed pressure on open-weight releases and licensing terms
None of this is hypothetical. It’s already happening, just quietly, in terms of service updates and detection systems you’ll never see.
The Part Everyone’s Skipping
Here’s what bugs me about the coverage so far. Everyone’s framing this as a China-versus-America story. But distillation is a universal problem. Western startups do it too. Open-source communities do it. Hell, I’d bet good money that some of the loudest voices condemning this behavior have run similar experiments internally.
The real issue isn’t nationality. It’s that the entire AI economy is built on a foundation of unclear ownership over model outputs. We never settled the basic questions: Who owns a model’s answers? Can you train on them? What counts as fair use versus extraction? The industry sprinted ahead and left the law in the dust.
Anthropic’s report is a symptom, not a cause. It’s what happens when a trillion-dollar industry tries to operate without rules and then gets surprised when everyone plays dirty.
My Take, For What It’s Worth
I think Anthropic is probably right that something systematic is happening. I also think the report is a strategic document as much as an investigative one. Both things can be true.
What I don’t buy is the framing that this is an existential threat to Western AI leadership. Frontier models are hard to build, but the knowledge to build them is diffuse now. Distillation accelerates the spread. It doesn’t create capability out of nothing.
If anything, this whole saga should push the industry toward something it’s been avoiding for years: actual standards for how models are trained and where their data comes from. Transparency isn’t charity. It’s the only way to make these accusations meaningful instead of just marketing.
Until then, expect more reports, more finger-pointing, and more headlines that treat corporate rivalry as a morality play. I’ll be here, reading the footnotes and rolling my eyes at the press releases.
Because that’s the job. And somebody has to do it.
Original source: read the full article