Anthropic's $1.5 Billion Copyright Settlement Wasn't the End of the Bill. It Was the Price Tag Everyone Else Rejected.

Here is the problem with putting a price on something you were hoping to make disappear: everyone can now see the price, and some of them will decide it's an insult.

When Anthropic settled the Bartz copyright case for roughly $1.5 billion late last year — about $3,000 per book across some 482,000 works — the coverage treated it as closure. AI had its copyright reckoning, paid the bill, moved on. That reading lasted right up until six authors, led by two-time Pulitzer winner John Carreyrou, looked at the $3,000 figure, opted out of the settlement, and filed to argue in open court that it represented about two percent of what the law actually allows. They're seeking statutory damages of $150,000 per work — not from Anthropic alone, but from six labs at once: Anthropic, OpenAI, Google, Meta, xAI, and Perplexity. Stack the defendants and the theoretical number is $900,000 per book.

The settlement didn't end the question. It printed a number for everyone else to reject.

First, the discipline, because the coverage has been wrong in both directions

The press made two opposite mistakes here and I don't want to make a third. The first was declaring the matter settled. The second, in reaction, is the emerging "$50 billion, seventy lawsuits, extinction event" framing. Both are wrong, and getting this right requires being boring about the math for a moment.

$150,000 per work is a ceiling, not an expected recovery. Statutory damages run a court-set range — $750 to $150,000 per work — and the top of that band only applies to infringement a court finds willful. Judges rarely stack maximums. The six-author suit, in raw dollars, is trivial to a company Anthropic's size; you'll see a "$75 million against Anthropic" figure floating around, but that's a secondary-source estimate, not a number in the actual filing. And the complaint itself was filed in December 2025 — what's new is the salience, as the Bartz settlement nears final approval and the opt-out track becomes real. So: not an apocalypse, and not a fresh bombshell. If that were the whole story, I'd have skipped it.

It isn't the whole story, because two things underneath it are genuinely load-bearing, and almost nobody is saying them clearly.

Load-bearing thing one: you're liable for the download, not the model

The intuition everyone has — "if the pirated books didn't end up in the shipped product, where's the harm?" — is exactly backwards under the ruling this all rests on.

Judge Alsup's holding in Bartz was that pirating the acquisition is, in substance, inherently and irredeemably infringing even if the pirated copies are immediately used for a transformative use. Read that carefully. The liability attaches to the act of downloading the LibGen or Z-Library corpus in the first place. It does not depend on whether that corpus ever reached a released model. This is why Anthropic paid $1.5 billion despite certifying that the pirated books were not used in any commercial version of Claude. The certification was probably true and completely beside the point.

That's the fact that makes this field-wide instead of company-specific. "We didn't train production on it" is not a defense. Every lab that torrented a shadow library to build its training pile carries the same acquisition-stage exposure, regardless of what it did with the files afterward. The opt-out authors' complaint names the same pirate sources — LibGen, Z-Library, OceanofPDF — for all six defendants for exactly this reason.

Load-bearing thing two: a settlement caps liability. Opt-outs don't.

Here's the part that should worry a CFO more than any single dollar figure. A class settlement's whole purpose is to draw a line: pay this, and the claims are closed and bounded. Its value is the cap.

Opt-outs blow the cap off. A willful-statutory suit that walks away from the settlement can't be extinguished by that settlement, and every author who opts out is proof that the mechanism labs used to "resolve" this doesn't actually resolve the tail. So the risk sitting in front of these companies isn't $1.5 billion, or $75 million, or any specific number. It's that the liability is open-ended — un-capped and un-settleable in the way a class deal was supposed to guarantee — and it stays open-ended across roughly seventy pending suits while six of the defendants march toward public offerings priced as if the question were closed.

That's the actual story, and it's a finance story more than a legal one. Put $1.5 billion against Anthropic's ~$965 billion valuation and it's 0.16% — a rounding error. The threat to an S-1 isn't the magnitude; it's the character. The reportable question, the one worth watching as these prospectuses land, is whether the training-data risk disclosures move from generic copyright-boilerplate to specific, quantified, open-ended language — because the honest version of that disclosure is uncomfortable, and the Bartz settlement is what made the generic version untenable.

The precedent everyone's scared of threatens a different company than they think

There's a Third Circuit ruling pending — Thomson Reuters v. ROSS — that the scare coverage treats as the domino that detonates all seventy suits. It won't, and the reason is instructive.

ROSS is the worst possible fact pattern for AI defendants: a company that trained on a rival's own editorial product specifically to build a substitute for it. A ruling against ROSS most likely lands on narrow market-substitution grounds — which is a real problem for vertical AI, the legal and medical and financial tools trained on licensed data they then compete with. It is not a clean hit on general-purpose book training, which turns on the piracy-of-acquisition theory above, and where Bartz's separate finding — that legally acquired books can be fair use — was never even appealed.

So there are two different exposures here, with two different sets of defendants, and the coverage keeps welding them into one. Piracy-of-acquisition is the book authors' theory; output-market-substitution is the Third Circuit's. Conflating them is how you get to a $50 billion headline that doesn't survive contact with the docket.

What I think

The dollar damages stay bounded — low single-digit billions across the field, survivable individually, and the scariest appeal lands narrower than the panic suggests. I'd bet on that. What I don't think goes away is the overhang: the settlement converted training-data liability from a vague risk into a standing, priced, un-closeable line item, and it did so at the precise moment the labs need investors to believe the risk is behind them. Opt-outs are the mechanism that keeps it open, and Carreyrou just demonstrated that the mechanism works.

What would change my read: a willful-infringement verdict at or near the ceiling on the piracy theory, setting a per-work number far above $3,000, followed by a wave of opt-outs and new suits extending the same acquisition-stage logic to code, news, and academic papers. That's the bearish path, and it's plausible, not fringe. Watch whether Bartz's final approval holds and how many more authors decide $3,000 was an insult. The settlement gave every one of them a number to reject. Some will.