The ai race between the US and China just got a lot more tense, and a lot more public. This time, the claim isn't about who has the best benchmark score, it's about cheating and who gets to build on whose work.
An American AI company behind the chatbot Claude says three major Chinese AI labs, DeepSeek, MiniMax, and Moonshot AI, effectively "mined" Claude by hammering it with prompts at scale, then using those outputs to improve their own models. The big debate is messy though: if "distillation" is common in AI development, when does it cross the line into theft, or does it at all?
What Anthropic is accusing DeepSeek, MiniMax, and Moonshot AI of doing
The core allegation is straightforward: Claude was queried so many times, in such a structured way, that it looks less like normal product usage and more like an extraction campaign.
Anthropic says the three labs created over 24,000 fake accounts and ran more than 16 million exchanges with Claude. The purpose, Anthropic claims, was to make Claude generate enough useful outputs that competing models could learn from them.
If you want the company's own explanation of what it detected and how it thinks these campaigns work, see Anthropic's write-up on detecting and preventing distillation attacks.
One detail matters a lot here: Claude isn't officially available in China. So the accusation includes how access happened, not just what the prompts looked like.
How Claude was allegedly accessed despite regional restrictions
Anthropic's claim is that these labs used proxy servers to route traffic and bypass regional access controls. That's important because it shifts the story from "a competitor used an API in a way you didn't like" to "a competitor worked around the product's availability rules".
This is also why fake accounts matter. If you create thousands of accounts, you can spread out volume, avoid rate limits, and run different prompt sets in parallel. At that point, "asking questions" starts to look like industrial activity, not a curious user.
The alleged scale, broken down by company
A quick table helps here, because the numbers are kind of the whole point.
| Company named in the claim | Alleged exchanges with Claude | Claimed focus areas |
|---|---|---|
| DeepSeek | ~150,000 | Logic, alignment, and handling sensitive topics |
| Moonshot AI | ~3.4 million | Advanced reasoning |
| MiniMax | ~13 million | Coding |
The takeaway is that the allegation isn't only "they copied outputs". It's that they targeted different capability areas, almost like a training plan. DeepSeek allegedly focused on behaviour and boundaries, Moonshot on reasoning, and MiniMax on code.
For a mainstream summary of the allegation and the "industrial-scale" framing, CNBC's coverage of the distillation claims is a useful reference.
Distillation, explained without the buzzwords
"Distillation" can sound like a fancy lab term, but the basic idea is easy to picture.
Imagine you've got an exam and you didn't study. So you sit next to the smartest kid in class. Then you ask them question after question, thousands of times, and copy the answers. Eventually, you produce your own paper that looks smart, even though you didn't do the learning the hard way.
That's the rough analogy being used here.
In AI terms, it works like this:
- A strong model produces answers (in this case, Claude).
- A weaker model trains on those answers until it starts to mimic the stronger model's behaviour.
Distillation itself isn't rare or exotic. Teams do it internally all the time to create smaller, faster, cheaper versions of their own models. That's the key difference: internal optimisation versus using a competitor's system as your "teacher".
So is it illegal, or just against the rules?
That's where things get uncomfortable, because "wrong" and "illegal" aren't the same.
Distillation may violate terms of service, especially if you're using fake accounts or bypassing geographic restrictions. Still, whether it counts as theft in a legal sense can depend on the details and on how courts and regulators choose to interpret it.
And honestly, that legal grey area is part of why this argument is happening in public. When the rules aren't settled, the loudest players try to shape them.
The bigger fear: safety guardrails get stripped away
There's a second layer to the complaint that isn't really about copying at all. It's about safety.
US AI firms often say they build safety guardrails into their systems. In plain English, that means trying to stop the model from helping with harmful tasks, like cyber attacks or malicious automation.
The worry, as described, is that distillation can copy "capability" without copying the same safety layer. So a competitor could end up with something that behaves powerfully, but refuses less, or refuses differently.
A distilled model can inherit the "how to do it" answers, without inheriting the "don't do it" brakes.
Now, to be clear, that's still part of the accusation and the framing. It's one side's view of what happens when model outputs are reused at scale. The Chinese companies named haven't responded in this segment, and neither has Beijing.
Why this matters even if you don't care who "wins"
Even if you're not emotionally invested in US versus China, the guardrails argument hits a practical point: people don't just fear strong AI, they fear strong AI that's easy to misuse.
If an AI model becomes better at writing code, planning steps, or reasoning through constraints, then the "abuse surface" grows too. A model that can help automate normal tasks can also automate bad ones, unless someone put real effort into restrictions, monitoring, and policy.
That's why this story keeps circling back to control. Not just who trains the best model, but who sets the limits, and whose limits become the default.
The hypocrisy argument: who gets to claim the moral high ground?
There's also an obvious counterpunch baked into this whole debate. Many people hear a US company complaining about "copying" and immediately think: hang on, aren't AI models trained on huge amounts of data pulled from the internet?
That criticism shows up here too. The segment argues that US firms use vast amounts of training data they may not have clear rights to, and it points at Anthropic in particular as having faced serious copyright-related legal pressure. It even mentions a reported lawsuit settlement figure, and the broader point being made is simple: it's hard to lecture others about ethics when your own house isn't spotless.
Elon Musk is mentioned as weighing in as well, accusing Anthropic of routinely taking data. Whatever you think of Musk, the moment is revealing because it shows how fast this turns into a mud fight. Nobody wants to be the only one playing by polite rules.
If you want more context on how tense the broader US China competition has become, this site has also covered the "catching up fast" narrative in Microsoft Shocks the AI World: "China AI Is Now Too Powerful" and What It Really Means.
Two things can be true at the same time
Here's the messy bit people don't always like saying out loud:
- It's plausible that competitors try to copy capabilities, because that's how fast-moving markets behave.
- It's also true that many AI companies, especially in the US, have their own unresolved questions around training data, permission, and fair use.
So when a company calls distillation "theft", a lot of listeners don't only judge the claim. They judge the messenger too.
Why this looks like the first skirmish of an AI cold war
This debate isn't really about one chatbot being asked too many questions. It's about power.
The US has led much of the modern AI wave, especially in frontier model deployment. China, meanwhile, has been moving quickly, and it has strong incentives to reduce reliance on US tech stacks. When you put those incentives next to export controls and chip restrictions, the outcome is kind of predictable: accusations, counter-accusations, and a race to lock down advantages.
The segment frames this as the beginning of an AI cold war, and that phrase fits because so much of this is indirect conflict:
- Control chips and advanced hardware through export rules.
- Limit who can access high-end models and APIs.
- Accuse rivals of cheating, and paint their progress as illegitimate.
- Push narratives about safety, national security, and governance.
For another angle on how quickly the competitive pace is changing, especially when multiple labs ship major updates close together, see Google, OpenAI, and MiniMax Dropped Powerful AI at Once, and It Changes the Rules.
Speed wins, scale wins, and rules get written afterwards
One line of thinking in the segment is worth sitting with: whoever trains faster, scales faster, and controls the guardrails ends up writing the rules for the future.
That's a big statement, but it maps to how technology power usually works. Platforms that become defaults get to set norms, pricing, and access. Later, laws often show up to formalise what's already true in practice.
For a broader report on the same "are they cheating?" framing around China's top labs, CNN's coverage of the allegations adds additional context and reaction.
What I learned while thinking this through (and where I landed)
I've had my own small taste of this dynamic, not at industrial scale, obviously, but in a way that made the whole thing feel real. I once tested two different AI tools side by side on a coding task, nothing fancy, just debugging a messy script. One tool kept giving clean, careful steps, and it refused certain risky requests. The other one felt… looser. It moved faster, it offered more direct snippets, and it didn't push back much.
At first I liked the second one more because it felt helpful. Then I noticed I was trusting it too quickly. That's on me, but still, the product design mattered. A few refusals, a few safety nudges, they slow you down just enough to think. Without them, you can drift into dumb mistakes, or worse, without noticing.
So when people say "distillation strips guardrails", I get why that scares them. And when others say "US companies complain about copying while training on everything", I get that too. It's frustrating, because both sides have a point, and you can feel it in your gut.
If anything, this story made me more sceptical of clean heroes and villains. It also made me pay more attention to who controls access, because access is power, even when the tech looks the same on paper.
Conclusion: the argument isn't going away, it's spreading
Anthropic's claims put real numbers behind a fear many AI companies already had, that competitors can learn a lot just by querying models at scale. Yet the hard part remains: distillation sits in a grey zone, common as a method, controversial as a tactic, and legally unclear in many cases.
For now, this is still a one-sided public accusation, with no response covered here from the Chinese labs or Beijing. Still, it's a strong signal that the AI race is shifting from quiet competition to open conflict, and the next fights will be about access, safety, and who gets to set the rules. If you want to keep up with how fast this rivalry is escalating, OpenAI GPT-5.3 shocks Anthropic's Claude Opus 4.6 is a useful companion read on the product side of the same tension.
0 Comments