One company built a product by taking other people’s work without asking. The same company, in the same stretch of headlines, is furious that someone took its work without asking.
A federal judge gave final approval to Anthropic’s $1.5 billion settlement, the largest known copyright settlement in United States history, brought by authors who said Anthropic used their books to train Claude without asking nor without paying.
Now set that beside Anthropic’s own public accusation that three Chinese labs ran industrial-scale campaigns to harvest Claude’s output through roughly twenty-four thousand fraudulent accounts, extracting the very capabilities those authors’ books helped build. In the first story Anthropic is the one who took. In the second it is the one who was taken from. The word it uses for the second is theft, and it reaches for national security to describe the stakes.
Of course these two examples are not the same legal question. The authors’ case was about copyright. The case against the Chinese labs is about contract and fraudulent access, using fake accounts to get around the terms and the regional bar on selling Claude in China.
Different bodies of law. If you flatten them into one thing, someone will correct you and they will be right.
Yet, look underneath the law at what each fight was actually about. In both, the thing people were angry about was the same: you learned from our work and you did not pay us for it. What is so maddening is that grievance turned out to be the part the law would not touch.
Watch how it played out for the authors. The court that heard their case found that training Claude on their books was transformative fair use! The stealing learning was blessed. What Anthropic paid for was narrower. It had also downloaded and inexcusably stored pirated copies of the books. It appears the downloading/training was fine, it was storing them that was not fair use.
Thus if Anthropic had simply not stored stolen material, it all would have been blessed?
The settlement covers the piracy, roughly $3,000 per work across an estimated 500,000 works, and leaves the thing the authors actually resented, a commercial model built on their writing for free, standing as permitted. Anthropic’s own statement says the quiet part plainly. Training on books (without permission, without paying), it notes, remains fair use.
That is the shape I want you to see. Not that powerful players win, everyone already knows that. Something narrower and more useful.
Let me be more direct, stated at exactly the size the evidence supports. I am not telling you this is how the whole system always works. I cannot prove that from where I sit, and anyone who hands you a clean percentage for how often power bends the rules is selling you something. What I can do is walk you close enough to the move that you start seeing it yourself.
Two cases, worked as examples. Not a census.
My claim is this. A protection can be real on paper, open to everyone, and still be quietly reserved for the few in practice. The reservation gets written in the language of neutral principle, so it never looks like a reservation at all.
Why does that matter? Because what you cannot see, you cannot contest. The move works precisely by not looking like a move. That is the whole reason it is worth learning to see.
Take a second case, from a courtroom far from anything to do with technology. John Durnell used Roundup on his property for years and developed non-Hodgkin lymphoma. He sued Monsanto in Missouri, argued the label should have warned him, and a jury agreed and awarded him about $1.25 million. He won, he beat the biggest of the big.
The door to justice was open, he walked through it, and he came out with a verdict in his hand.
Then the Supreme Court took the verdict back. In Monsanto v. Durnell, decided seven to two, the Court held that federal pesticide law preempts state failure-to-warn claims when the EPA has made its labeling determination. Not a word about protecting a large company from ruin. Nothing so crude. The ruling was about federalism, about who properly decides what goes on a label, one federal agency or fifty separate juries.
A clean, neutral question of authority?
The output was that a man with a jury verdict and a cancer lost it, and thousands of similar pending claims were swept off the board with him. The costume was federalism. The body underneath was a transfer of risk from the company to the people it harmed!
Do you see what made the tilt visible?
Durnell won first!
The Court’s ruling leaves other kinds of claims alive, design defect and the rest, which tells you the door was genuinely open before this particular principle closed it. If Durnell had simply lost at trial, we would call it a hard case and move on. It was the reversal of a win, through a rule that sounded like pure structure, that let you see the machinery working.
Now a third incident that looks like it belongs here and does not, because the difference is my entire point.
Members of Congress had their phone records swept up in a federal investigation, and they were angry. They spoke of violation and overreach. How loud are they when the citizenry’s phone records are and were swept up? It is tempting to file this next to Durnell as one more instance of the system serving the connected.
It does not fit, for two reasons. First, nobody dressed this one up. The records in question were toll records, showing when calls happened but not their content, and that kind of metadata collection has been openly permissive for a very long time. There is no neutral principle here pretending everyone is equally protected. The permissiveness is right out in the open.
Second, and more important, the vector runs backward. This is not a benefit reserved for the few. It is a burden spread across the many, and “the few” were simply startled to feel it land on them for once! The ordinary version of this happens to regular people constantly and produces no press conference. So it is a different animal. Worth understanding, but not an instance of the thing I am describing.
Keep it as the counterweight that tells you my pattern has edges.
The not-so-neutral principle
Here is the move that the two real cases share. The exclusion is dressed as a neutral principle. Fair use. Federal preemption. The doctrine changes from case to case. The direction does not. The principle behaves like a dependent variable, and leverage is the thing it depends on.
And because I do not want to hand you a metaphorical lens that lights up at everything, here is how you would know I am wrong.
If leverage did not predict the outcome better than the stated principle does, my idea is shallow to empty. If fair use had shielded a single struggling author the way it shielded the company training on that author, or if preemption cut for injured individuals as often as against them, then there is no pattern here, just law doing its ordinary work.
The tell is not that a heavyweight won. The tell is that the principle bends the same direction every time a heavyweight is in the room.
The part I watch most directly is in my own field. The models that power modern AI have quietly become a commodity. The labs proved it with their own money when they spent billions standing up enterprise deployment companies instead of just selling more tokens.
The scarce thing is no longer the model.
It is the engineers (and the hardware, GPU’s, memory, and electricity) who can make a good-enough model actually work inside a real business, and that talent gets rationed by leverage. Notice the structure of Anthropic’s own venture, Ode: it comes with distribution into the portfolio companies of Blackstone, Goldman Sachs, and Hellman & Friedman.
Everyone can rent the engine at the same price.
The mechanics who make it run get routed first to the people who were already ahead. The same structure, arriving in the layer I work in every day.
Power ebbs. It always has. The tilt is real and it is also not permanent, and that is the closest thing to good news in any of this.
But the ebb rescues no one on its own. When it turns, and it will, you are either the person who read the shift coming or the person it happened to. The difference between those two is not luck and it is not patience.
It is whether you learned to read the costume while everyone around you was still admiring the outfit.
Speak Your Mind