A federal judge approved Anthropic's $1.5B settlement for piracy, not AI training

A federal judge just made the piracy finding final

The case is *Bartz v. Anthropic*. Judge William Alsup issued two separate rulings on Anthropic's book habit: training Claude on copyrighted material, he found, was fair use. Downloading pirated copies to acquire those books in the first place was not. This settlement resolves only the second finding.

The two sides had actually agreed on the number a year earlier. Alsup's approval this month made it final, after he cut the class-action lawyers' own fee nearly in half along the way, from an original ask of $187.5 million down to $101 million, with $2.6 million in unreimbursed litigation costs on top. The three named plaintiffs who put their names on the suit collect $15,000 each.

The thread's first instinct was wrong, and the submitter said so within the hour

In the thread, the first reply fired off a guess before reading past the headline: "that number is missing a zero or two in front of the decimal point." It read like an AI company getting off cheap for training on half a million stolen books. The HN user who had submitted the story corrected the record fast: "the settlement is for piracy, not training. They had already ruled that Anthropic's use of copyrighted material for training fell within fair use… if you pirated a book and had to pay $3000 for that one instance, I don't think you'd like it if I said you should have paid $30K or $300K instead."

The man who pirated the books first, to answer the question himself

One voice in the thread had already run this experiment on purpose, years before Anthropic's case existed. An ML researcher who built books3, the pirated-book dataset behind a wave of early open AI models, read the settlement as closing a door he'd deliberately forced open:

"It's an unfortunate outcome. Now to be a big player in AI, you have to have enough capital to buy your own library worth of books and digitize them… I created books3 to help settle the question of whether AI companies should be allowed to train on books… Now we're in a world where you have to have dozens of millions in capital to do substantial work."

Whether "open" gets the same pass

Not every reader took the fair-use half of the ruling as settled comfort for the rest of the industry. Wes Felter, a former IBM Research and datacenter engineer, raised a question the settlement doesn't answer: whether labs that trained on the same pirated corpora but gave away their weights for free get to claim a different standard. "Most open models are developed for profit so they should be equally liable," he wrote.

A SWAT raid for one pirate, a settlement check for another

The comparison that kept resurfacing was Kim Dotcom, raided and prosecuted a decade earlier for running a file-sharing site. A Brazilian hobbyist programmer laid the contrast out bluntly: "These AI settlements make such a mockery of past copyright enforcement victims that it's straight up offensive. Police descended upon Kim Dotcom like he was a terrorist or something. They rappelled down helicopters and stormed his home like he was bin Laden. Then these big techs come along and they make some absurd cost of business settlement."

No one in the thread disputed that Anthropic pirated the books. What split the room was whether a company with billions in funding experiences this settlement the way an individual experiences a SWAT team, and whether the next AI lab reads it as a warning or as a price list.