Publishing Was Already Broken. AI Is the Stress Test.

On July 14, Hachette, Cengage, Elsevier, and the novelist Scott Turow sued Google in the Southern District of New York. The complaint says Google trained its Gemini models on books it held for Google Books and the Play Store, stripped copyright information to hide the sources, and understood the exposure: an internal document cited in the filing warned of "$10Bs–$100Bs in potential fines."

Set the legal questions aside for a moment and look at the situation underneath them. Publishers handed those books to Google twenty years ago so readers could search inside them. No contract from that era said a word about training a language model, because no one had built one. Two decades later, there is still no working mechanism for a rights holder to say yes, no, or "yes, at this price" to a use nobody anticipated. That gap is a publishing problem, and AI has made it impossible to ignore.

Three more exhibits from this month.

The Anthropic settlement is moving toward payment, and the machinery creaks at every joint. Authors and publishers claimed 91.3 percent of the 482,460 eligible works, a rate class-action administrators rarely see; ten percent is typical. Authors want this money. The base payout runs about $2,931 per work, with the first checks expected in late fall. Here is the broken part: many authors will be paid only because their publisher filed the claim, and the split follows contracts that never contemplated AI, or a 50-50 default where the contract is silent. It took a federal class action and a $1.5 billion fund to produce the first dollar figure most authors have ever seen attached to the use of their own work in training.

The marketplaces are forming without books in them. Microsoft opened its Publisher Content Marketplace in February: usage-based payment, terms set by the publisher, an exchange rather than a bilateral deal. The pilot partners are Condé Nast, Hearst, Business Insider, and the Associated Press. Not one trade book publisher appears on the list. In the UK, more than 250 publishers joined the Publishers' Licensing Services collective AI licensing scheme this month, and Bloomsbury says it is exploring opt-in deals for its authors. The infrastructure for pricing content to machines is under construction, and the book business is mostly watching from the sidewalk.

The trust layer failed a live test. Granta severed its relationship with the Commonwealth Foundation over suspected AI-written stories on the Commonwealth Short Story Prize shortlist. The Foundation investigated and concluded no AI had been used. Granta ended the partnership anyway. When there is no accepted way to document how a work was made, institutions act on suspicion before the evidence arrives, and they retreat to the safest available position: refusal.

Four stories, one pattern. The contracts cannot answer the consent question. The payment machinery cannot route money to the person who wrote the book without a courtroom. The new markets are being built without the book business in the room. And disclosure norms are so unsettled that an investigation clearing a writer could not save the relationship.

None of this machinery broke last month. Trade contracts have been silent on unimagined uses for as long as there have been trade contracts, and royalty systems have always strained to track money at the level of a single title. AI companies put real money against every one of those weak joints at once, which is what a stress test does.

The fix is unglamorous: infrastructure. A record of who holds which rights for which uses, legible to a human author and to a machine counterparty at the same time. Terms an author can read on one page. Consent that can be given, priced, or refused before the training run, instead of litigated three years after it. Some thirty startups are working on pieces of this; Amlet, where I run US publishing, is one of them, and the problem is bigger than any of us. The measure of success is simple to state and hard to reach: the next Google Books moment should open as a negotiation, not a decade of discovery.

The system around the books is the broken part. For the first time, the money to fix it is on the table.