Anthropic says it found industrial-scale distillation attacks from DeepSeek, Moonshot AI, and MiniMax: 24,000+ fraudulent accounts and 16+ million exchanges allegedly used to extract Claude capabilities.
If that claim is accurate, it’s a serious abuse case.
And still, my first reaction was the same one a lot of people had:
Wait. Let me get this straight. The companies that absorbed the whole internet are now furious that others are absorbing them?
That tension is the story.
Two things can be true at the same time
- Account fraud and terms-of-service evasion are legitimate abuse.
- Moral outrage sounds thin when your own training stack depended on data capture at internet scale.
Most discourse collapses into team sports. It shouldn’t.
If you fake accounts and automate extraction at scale, you should get throttled, banned, and litigated if needed.
But if your business model was built on “ask forgiveness later” ingestion dynamics, don’t expect universal sympathy when the same logic is applied one layer up the stack.
Distillation is not a bug in the market. It’s the market.
Once a model is exposed through an API or chat interface, competitors will try to:
- map behavior,
- reproduce style,
- approximate quality,
- and close the gap faster than raw pretraining alone would allow.
That is not surprising. That is competitive gravity.
The surprise is that some labs still talk like this was unforeseeable.
The internet-era precedent is ugly—but clear
The first wave of foundation models normalized this stance:
- public content was “available enough” to train on,
- consent was fuzzy,
- attribution was optional,
- and speed beat ethics committees.
Now the second wave is doing the same thing to model outputs.
Nobody likes being downstream of their own precedent.
What actually matters now (beyond hypocrisy memes)
The useful question isn’t “who’s morally pure?” because the answer is probably “nobody.”
The useful question is: what rules do we want now that everyone understands the extraction game?
Pragmatically, this means:
- stronger abuse detection at the account and traffic layer,
- tighter identity and rate controls for high-risk usage patterns,
- watermarking/fingerprinting research that survives paraphrase,
- clearer industry norms around synthetic data provenance,
- and legal frameworks that don’t pretend this is either fully legal or fully impossible.
My take
Anthropic is not wrong to defend its models.
But the industry should stop performing shocked morality plays about distillation. This is a structural consequence of how frontier AI was built and shipped.
If you train on everything, then expose powerful behavior through public interfaces, copy pressure is inevitable.
You can fight it. You should fight abuse.
Just don’t pretend you discovered some brand-new ethical category when the mirror finally turned around.