The going rate for a book in an AI training deal is $5,000 — and your publisher may split it with you

What happened

Publishers Weekly's "Publishing's AI Reckoning," picked up and circulated on Sept. 21, names Google as having signed AI partnerships with Bloomsbury, De Gruyter Brill and Johns Hopkins University Press. Wiley has done multiple deals since 2024. Taylor & Francis licensed to Microsoft for a reported ~$10 million in the first year.

None of the Google deals come with a public dollar figure, a term length, or an answer on whether authors get a cut. We asked; we don't have it. Treat the deal list as confirmed and the terms as unknown.

The news peg here is not the contracts. It's that book publishing flipped from litigation posture to procurement posture, and the only disclosed per-unit price in the entire market is still two years old.

That price: the HarperCollins-Microsoft deal. $5,000 per participating title, three-year term, split 50/50 with the author, opt-in per title. The Authors Guild said publicly at the time that this was far too little for a training license.

What this is not

It is not a market rate. It is one number, from one house, from one deal, that everyone keeps citing because nothing else is public. Whether $5,000 is still what a title fetches in late 2026 — or $500, or $50,000 — nobody outside those rooms knows.

It is also not the same thing as the Anthropic settlement. Anthropic's $1.5 billion class settlement covers past conduct. It buys no forward license. A payout for books already scraped and a contract to use books going forward are different transactions at different prices, and conflating them makes the licensing market look richer than it is.

And "AI licensing" is at least three products. Training. Retrieval. Transformation. A startup, Bookpact.ai, is selling per-title opt-in where a publisher picks which one is permitted. Those should not carry the same price, and if your contract doesn't separate them, you sold all three for the price of one.

Why this matters if you've never written a book

Because you probably have a corpus. A podcast back catalog. Six years of newsletter archives. A video library with transcripts. Scholarly and trade publishers are selling first because their content is structured, rights-cleared at the publisher level, and boring in the way training data likes. Your archive is the same shape.

The number being set now for a book is the anchor that gets pointed at when someone offers you a lump sum for your archive. $5,000 for three years of a full-length book, half of it to the creator, is the ceiling that the Authors Guild called too low — and it's the comp that will show up in your deal memo.

The second thing to notice: most pre-2023 book contracts are silent on AI training. That silence is where the fight is. If your publisher, distributor, or network holds broad digital rights, they may be able to license your work without asking you.

What to do about it

1. Pull your own back contracts and search for "machine," "data," "derivative," "electronic" and "all media now known or hereafter devised." That last phrase is the one that sweeps.
2. Before signing any AI addendum, make the buyer name the use: training, retrieval, or transformation. Price each separately. Never sign one line covering all three.
3. Ask whether the fee is per title or per corpus. Per-corpus pricing is how a 40-episode catalog becomes one $5,000 check.
4. Get the term in writing, plus what happens to a model already trained when the term ends. In practice the license expires and the training doesn't.
5. Check for opt-in. HarperCollins was opt-in per title. If your deal is opt-out by default, that is the single most consequential sentence in it.
6. Confirm whether audio, translation and derivative formats are swept in. Those are separate rights you may have already sold elsewhere.
7. Don't sign anything against the $5,000 comp without asking what deals have closed in the last six months. It's the only public number, not the current one.