The $1.5 Billion Library Card

Authored by Jay Rogers via American Greatness,

AI companies won the argument about learning from books. They lost the argument about stealing them. Both sides of this fight have a reason to blur the two.

On July 20, a federal judge in San Francisco gave final approval to the largest publicly reported copyright recovery in American history. Anthropic will pay $1.5 billion to the authors and publishers of 482,460 books, roughly $3,100 a title before fees. A hardcover runs about $30. The other $3,070 is the surcharge for pirating it instead of buying it.

Anthropic won on training. Sourcing sank it. In June 2025, U.S. District Judge William Alsup held that training on lawfully acquired books was fair use, calling the purpose “exceedingly transformative.” The company downloaded millions of titles from Library Genesis and Pirate Library Mirror and kept them in a permanent central library. Pirating copies a company could have bought, Alsup wrote, is “inherently, irredeemably infringing.”

Here’s my thesis. We need a fair use doctrine that protects analysis and refuses to launder piracy. Punish analysis, and we strangle the most useful technology since the transistor. Excuse theft, and we teach a trillion-dollar industry that acquisition costs are optional. The Copyright Clause exists to “promote the Progress of Science and useful Arts” by securing authors’ rights “for limited Times.” Progress is the purpose. Ownership is the means.

I’ve spent more than 30 years in institutional investing and served as an expert witness in securities and fiduciary disputes. I read chains of title for a living. When a firm can’t document how it acquired an asset, nobody in my business calls that a philosophical debate. We call it a diligence failure, and we price it. The AI industry spent three years arguing epistemology when its problem was inventory control.

Three Questions, Not One

Fair use under Section 107 is an affirmative defense. You admit the copying and then justify it. That’s why a plagiarism detector that ingested student papers won, and why Google won after scanning millions of books for a searchable index. A lawful use still doesn’t cleanse an unlawful acquisition. A stolen car will get you to work. It’s still stolen.

Three questions keep getting jammed together. How was the data obtained? What did the training do? And what comes out the other end?

Sort the cases that way and the record gets messy. Two days after Alsup ruled, U.S. District Judge Vince Chhabria ruled for Meta, though Meta also pulled books from shadow libraries. He reasoned that because the ultimate use was transformative, so was the downloading. He warned that cheap machine-made substitutes could dilute authors’ markets, then ruled against them anyway because they “made the wrong arguments.” Two judges, one courthouse, one week, two answers on piracy.

On Sept. 29, the 3rd U.S. Circuit Court of Appeals affirmed a ruling against a startup that copied Westlaw headnotes to build a competing research tool. It’s the first appellate decision on fair use in AI training. Don’t over-read it. The court unsealed its opinion soon after, calling the use “minimally transformative, at best,” and stressing that the system wasn’t generative and the defendant was a direct competitor. The pattern holds anyway: where output substitutes for the original, fair use runs thin. The Copyright Office’s May 2025 report said much the same.

Dirty Hands All Around

Both teams behave badly, in mirror image. In public, AI firms invoke fair use like converts who just discovered the Constitution. In private, once the download logs surface, they write checks. Anthropic paid to retire the piracy claim, with rightsholders claiming 91.3 percent of eligible books. Smart litigation. But the public argument and the private settlement can’t both be the whole story.

Publishers are no purer. Dow Jones and the New York Post sued Perplexity in 2024 for “massive freeriding,” while parent News Corp licensed content to OpenAI and Meta, with Meta’s reported at up to $50 million a year. CEO Robert Thomson calls it a “woo and a sue strategy.” CNN’s May complaint against Perplexity describes licensing talks that went nowhere. When identical conduct earns “partnership” from one chair and “theft” from the other, you’re hearing an opening bid.

Who the Rules Favor

A conservative should get nervous here. Mandate fully licensed training sets and you’ve built a fixed cost that favors incumbents. The checks go to big publishers, not the freelance historian or the 11-person regional paper. Sarbanes-Oxley was sold as investor protection, and the SEC’s own study found smaller companies paid more for compliance relative to their size.

The opposite rule is worse. Let anything go, and the sloppiest operator gets the cheapest data while the honest firm eats costs its rival skipped. Bad practice drives out good, the oldest story in finance.

The best objection to my view is plumbing. Court filings showed Meta pausing its book-licensing push in 2023 after “very slow uptake,” with fiction publishers admitting they didn’t hold the rights. Demand universal licensing with no licensing infrastructure and you’ve written a ban in a costume. Music solved the same mess a century ago with ASCAP, BMI, and statutory rates. Fragmentation is a case for building the pipes, and nothing in it excuses taking the water.

A Rule That Fits on an Index Card

Start with no piracy. Copies taken from a shadow library should be ordinary infringement with no fair use defense. Labs should document provenance the way a bank documents collateral.

Regurgitation is infringement. Verbatim output of protected expression should be actionable no matter how clean the training, with a safe harbor for firms that build and audit output filters. Alsup noted Anthropic already filtered outputs. That’s also how you answer Chhabria’s dilution worry: police substitution where it happens.

Lawfully acquired data gets default fair use. If you bought it, licensed it, or took it from the open web without breaching a contract, train on it. For material nobody can license title by title; create a collective licensing body with a published rate, an opt-out registry, and a small-claims track.

Congress should write this, not the courts. The Justice Department weighed in for OpenAI, and that New York case remains pending. Foreign courts are already splitting: Munich found OpenAI’s models infringed song lyrics, while London found Stability AI’s model didn’t store Getty’s images.

The Cheapest Thing They Could Have Done

Anthropic knew the honest route. It bought millions of used print books, cut off the bindings, and scanned them, and Alsup called that fair use too. At $30 a copy, half a million books would have cost about $15 million, a rounding error on a frontier training run. It pirated them instead and paid 100 times the sticker price.

Seinfeld’s library cop, Lt. Bookman, hounded Jerry for two decades over one overdue copy of Tropic of Cancer. The authors’ lawyers moved faster and sent a bigger bill.

I want the models: the protein structures, the earlier cancer detection, and the tutor at 3 a.m. But a country that lets the largest firms on earth skip the checkout line because the product is impressive has picked one set of owners over another and called it progress. Analysis is free. Expression is owned. Theft is theft, even when the thief is building the future.