Anthropic paid $1.5bn to keep a ruling it liked

Anthropic's $1.5bn class-action settlement with book authors won final approval on 20 July, according to reporting by TechCrunch and aggregated coverage of the Northern District of California proceedings. The structural point is not the headline number. It is that the money was paid to close a case Anthropic was, on the doctrinally important question, winning.

What the case actually resolved. The suit split into two questions. One: is training a large language model on copyrighted books fair use? Two: is downloading more than seven million books from Library Genesis and Pirate Library Mirror to build the training corpus lawful acquisition? Reporting on now-retired Judge William Alsup's earlier ruling, aggregated by AI Weekly and summarised by TechCrunch, has him answering yes to the first and no to the second. Training was fine. Torrenting from named pirate sites to feed the trainer was not.

The settlement pays out on the second question and leaves the first untouched. Liability attached to the supply chain, not to the model. That is a narrow legal footprint dressed as a landmark.

The price sheet, such as it is. The settlement fund covers roughly 500,000 eligible titles at approximately $3,000 per work, according to reporting aggregated by AI Weekly from Reuters, Bloomberg Law and the Authors Guild. TechCrunch, via Yahoo Finance, reports the same $3,000-per-work figure and describes the settlement as the largest known in US copyright history. The arithmetic is clean: $3,000 across 500,000 works is $1.5bn, which is the fund.

This is the first dollar-denominated benchmark for unlicensed training data anyone has to point at. Counsel at OpenAI, Meta, Google and Midjourney will be running the multiplication against their own corpus compositions this week. The number is real. Its portability is not.

Why the price does not travel cleanly. Liability here rested on unusually clean evidence of unlawful acquisition: specific torrents, named pirate repositories, a "central library" built out of them. That is a fact pattern optimised for plaintiffs. Web scraping, licensed-but-out-of-scope use, and fair-use disputes over news content all sit in murkier territory. A defendant that never touched LibGen will argue, plausibly, that the Anthropic number is not their number.

Cutting the other way: the $3,000 figure carries no punitive multiplier and no jury verdict on top. Statutory damages for wilful infringement run considerably higher. Anthropic settled to avoid finding out what a jury would have done with the piracy conduct at trial, which suggests the ceiling was well above $3,000 per work. Defendants pointing at the Anthropic settlement as a cap should be careful; plaintiffs pointing at it as a floor have more to work with.

The trade Anthropic made. Alsup's fair-use ruling on training was favourable to the whole industry. It was also a single district court decision. If the piracy-damages phase had gone to jury trial and Anthropic had lost badly, the appeal would have put the entire ruling, including the fair-use finding, in front of the Ninth Circuit. An appellate court might have affirmed it, sharpened it, complicated it, or reversed the training half on grounds nobody has yet argued.

Anthropic paid $1.5bn to make sure none of that happened. The favourable ruling stays on the books as persuasive but non-binding authority. It also stays untested. That is what the money bought.

This is a rational allocation of capital. It is also, from an industry-wide standpoint, slightly strange. The most quoted piece of judicial reasoning in favour of AI training being fair use is now the ruling that a defendant paid a record sum to prevent from ever being reviewed. Every other lab that wants to lean on it in its own case is leaning on a ruling that its most aligned party declined to defend on appeal.

The frame this fits. Read this through AI-safety-as-market-position and it lines up. Anthropic's brand rests on being the more careful frontier lab. Paying to close a piracy suit, quickly and at record scale, is consistent with that positioning in a way that a bruising jury trial over torrented books would not have been. The settlement is expensive reputational hygiene as much as it is legal risk management.

Read it through inference economics and the cash-flow question is the one to sit with. $1.5bn is not trivial against Anthropic's cost base, but paid across installments it is absorbable — particularly for a company whose 2026 has included reported talks over a compute-lease arrangement with Meta worth up to $10bn over two years, per Briefs.co reporting cited in the research file. The settlement is a line item, not a solvency event.

The fee haircut is its own signal. The court cut plaintiffs' attorneys' fees from a requested $187.5m to $101.5m, a reduction of $86m, or roughly 46 per cent of the request, according to the topic shell summary flagged as unverified against the court order. Treat the specific numbers with appropriate caution until the docket confirms them. If they hold, the direction is the thing to notice. A court willing to cut a fee request by nearly half in a record-setting AI copyright case is a court signalling to plaintiff firms that the class-action AI-copyright economics they may have been modelling on the way in are not the economics that will survive review on the way out.

That matters for the pipeline. There are active suits against OpenAI, Meta, Google and Midjourney, and a fresh class action against Google over Gemini training was filed the week of 14 July by Hachette, Cengage, Elsevier, Scott Turow and S.C.R.I.B.E., per TechCrunch. If the fee arithmetic on these cases compresses, the aggressiveness and volume of the next wave may compress with it — or plaintiff firms will restructure how they get paid.

What to watch. Three things. Whether the first fee-reduction ruling in an AI copyright class action starts showing up in other courts as a reference. Whether any of the parallel defendants cite Alsup's fair-use ruling in dispositive motions, and how those courts treat a ruling its beneficiary paid to keep unreviewed. And the composition of the next big settlement: if it looks like Anthropic's, piracy-adjacent conduct, no fair-use concession, the industry has a template. If a fair-use ruling actually goes to appeal somewhere, the template is worth much less.

Anthropic paid a record sum to preserve a ruling it liked and close a case it was going to lose the narrower half of. That is not a landmark on AI training. It is a landmark on how AI training cases end.

Footnotes