Selling Books to AI Companies: Prices, Deals and Who Gets Paid

You can sell books to AI companies, but they buy by the title and from whoever holds the rights, which for most books is a publisher licensing its backlist, not an author selling one book. The one per-title price a major deal has made public is HarperCollins' licence to Microsoft: $5,000 a title for a three-year licence, split evenly with each author who opted in. Across the book sales and payouts on record the median is $2,750 per book or title, against $343 per book or title in what authors have been offered as their share. The largest licensors report it as revenue of its own: John Wiley & Sons reports $49.1 million of AI licence revenue in the year to April 2026 and more than $110 million since it began. And for books copied without a licence, the court-approved Anthropic settlement pays about $3,000 a work.

This guide is for anyone who holds the rights to books: a trade, academic or professional publisher with a backlist, a university press, a small or independent press, a literary estate, an author whose rights have reverted, and a self-published author who never granted them away. It covers who buys books for AI training and why, every publisher deal with a public figure, what a title earns on each side of the market, how much of it reaches the author, who holds the AI rights in a publishing contract, the court cases that set the price of an unlicensed book, the licence terms that matter, how to prepare a catalogue, and the routes to market. The figures come from the price board, the deals register and the text and archives page, each linked to its source.

AIData.Broker also acts as the agent of holders who license or sell their data, so we are a party in this market. Check any figure you plan to rely on against its source.

This guide is research from public sources, not legal, tax or investment advice.

Contents

1. Who buys books for AI training, and why
2. The publisher deals on record
3. Academic and professional publishers: Wiley, Taylor and Francis and the rest
4. What one book earns
5. How much reaches the author
6. Who holds the AI rights: publisher or author
7. The court cases behind the price
8. The terms that matter in a book licence
9. How to prepare a catalogue for sale
10. Routes to market
11. How a sale through us works
12. Questions people ask
13. What this means for you

Who buys books for AI training, and why

Books are the longest, most carefully edited text a language model can read. A book holds an argument across a whole volume, written and corrected by professionals, on subjects the open web covers thinly. That is why the labs reached for them early, and why so many of them now face lawsuits over how they got them: the publishers' and authors' cases in section 7 are, almost all, about copies taken from pirate libraries rather than bought or licensed.

That history is what turned books into something you can sell. In 2025 a federal judge held that training a model on lawfully acquired books was a fair use, but that copies taken from pirate sites were "inherently, irredeemably infringing" even if they were used for the same training and then discarded. The case settled for $1.5 billion. A lab that wants books now has three lawful sources: the public domain, copies it buys, and licences from the rights holders. A licence is the only one that also settles what the lab may do with the model's output, for how long and with whose consent, which is what the deals below buy.

Who pays

The buyers on record are a short list of the largest AI and technology companies, usually unnamed in the publisher's own filing and identified later by the press.

• Microsoft, the buyer reported behind HarperCollins' licence ( as identified by Bloomberg) and the named partner in Taylor & Francis' first deal ( Informa's own announcement).
• Unnamed "large tech companies", the counterparties of both of Wiley's first content rights projects ( the first and the second), of Arcadia Publishing's sublicences, and of most academic presses' agreements.
• Life-science and healthcare AI companies for specialist content: Wiley names IQVIA and OpenEvidence among its AI customers.
• Platforms and agencies that license onward: Created by Humans licenses authors' books to AI developers, and in Korea a platform called Mentat licenses the books of about 95 publishers to AI companies.

What they want

Buyers do not want copies, they want distinct titles. A US distributor buying Korean used books in bulk in 2026, suspected of supplying AI training, told a newspaper it buys one copy of each title, that "the number of titles matters more than the total number of books", and that it wants academic and specialist nonfiction. The deals say the same thing in their own way: HarperCollins licensed select nonfiction backlist titles ( the deal record), and the academic publishers, whose catalogues are dense with specialist knowledge, have made the most money.

The competition is free text. Institutional Books, built from Harvard Library volumes that Google scanned, released 983,004 public-domain volumes as an open dataset in 2025. Old, out-of-copyright books are therefore no longer scarce. What is scarce is recent, in-copyright, edited text on specialist subjects, with rights that can be shown, which is exactly what a publisher's backlist holds.

Supply is large but fragmented. More than four million books received ISBNs in the US in 2025 alone, most of them self-published. A buyer cannot negotiate with each author, which is why the market runs through publishers, platforms and agents that gather titles into one licence.

The publisher deals on record

Every publisher licence with a public figure or a public structure is in the table below, from the deals register and the price board. Most publishers disclose only what the deal added to revenue, not what each title earned, and several disclose nothing at all.

Three things stand out. First, only one trade publisher has priced a title in public, and its price came with an opt-in and a cap on what the model may reproduce. Second, the academic publishers carry the money, because their catalogues are large, specialist and owned under contracts that already reach digital uses. Third, the per-title figures that reach the public come from authors, not publishers: a cheque photographed on a blog, a royalty email posted online, a letter quoted to a newsletter. They show what reached the author, and only by inference what the publisher was paid.

Academic and professional publishers: Wiley, Taylor and Francis and the rest

The academic houses are the best-documented licensors in this market because two of them are listed companies that must report what AI licensing earns. Their filings show how the money behaves over time, which matters more to a holder than any single deal.

Wiley

Wiley reports AI licensing as a line of revenue of its own. Its first project, worth $23 million, lifted its quarter in spring 2024; a second, worth $21 million, followed within months. By the year to April 2026 its AI licence revenue had reached $49.1 million, up from $40 million, and Wiley described its recurring revenue as "rapidly scaling".

Inside that total the mix moved. Wiley splits AI licence revenue between Learning (books, textbooks and professional titles) and Research (journals and research content). In the year to April 2026, Learning earned $16.0 million, against about $29 million the year before, while Research earned $33.1 million, against about $11 million. In the quarter to July 2026, Learning's AI licence revenue was zero, against $13.1 million a year earlier, and Research's was $13.7 million.

Read plainly: the large one-off licences of book backlists came first and have not repeated quarter by quarter, while licences of research content, which keeps growing, have become the steady part of the business. A backlist sold once to a lab that has already trained on it is unlikely to sell to that lab again at the same price.

Wiley's filing also shows a route for smaller publishers. Its Research Solutions business sells AI licences that include content Wiley itself licensed from other publishers, paying them a royalty: about 55 percent of that business's quarterly revenue came from contracts that include such AI licence revenue. A small press does not have to reach a lab on its own; a larger publisher that already sells to labs can carry its titles, for a share.

Taylor and Francis

Informa (Taylor & Francis) announced its first AI partnership with Microsoft in May 2024: an initial data access fee of about $10 million, plus recurring payments over three years, for non-exclusive access to its learning content. A second partnership followed two months later, and by year end Informa reported more than $75 million of AI data access revenue for 2024.

Informa itself calls that revenue non-recurring. In the first half of 2025 it added about £15 million more, again "incremental, non-recurring", from a new contract with a different buyer. The pattern matches Wiley's: large payments for access to an archive, each one a new buyer rather than a renewal.

Its authors saw little of it: one academic author told Bloomberg they earned $97 for their book, a report rather than a figure from the publisher.

The university presses

The university presses have taken two opposite approaches. Cambridge University Press asked first: it wrote to the authors of about 45,000 books to opt in, with a royalty of 20% of licence income. Johns Hopkins University Press included its backlist by default, let authors opt out, and told them to expect "modest" returns of under $100 per title per licence. Oxford University Press confirmed partnerships but no terms. A smaller academic press that asked its authors by tick box paid one of them £500 for a monograph, after keeping half the fee.

What one book earns

A book is priced by the title, almost never by the word or the page. The table sets the public prices for one title side by side, from the rates page, which computes each median from every record of its kind.

The middle half of the public sale prices runs $881 to $3,082 per book or title. The spread is wide because the prices measure different things: a licence fee to a publisher, an author's share of one, a court's award for unlicensed copying, and the asking price of a holder who has not yet found a buyer.

Licences, settlements and asking prices

The HarperCollins fee is the clearest benchmark because it is a price for one title, for a defined use and term, paid by a named buyer. One author published the cheque: $2,500, his half, less his agent's share, for a three-year licence of a nonfiction book.

The settlement figure is a floor of a different kind. It is what a lab agreed to pay for books it had taken from pirate libraries, about four times the $750 statutory minimum for ordinary infringement and far below the statutory ceiling. It is not a licence: the rights holders gave no consent and set no terms, and the money arrives in four instalments over two years. After fees and costs, the court's September 2026 status report allotted about $2,203.56 to each claimed work from the funds then held, before the final payment.

Asking prices are not sale prices. One Indian publisher lists 100,000+ rights-cleared e-books for $53,514,000 and 18,000+ for $9,473,700 on a data catalogue, non-exclusive and perpetual; no sale at those prices is on record. The low median of catalogue listings reflects bulk collections priced per title in cents and dollars, not what a buyer pays for a curated specialist list.

How a price is anchored

Where a court or a settlement had to set a per-title price, it anchored on the book's own list price. In Korea, a contractor that copied e-books into a national language corpus without licences settled with publishers at 70% of the e-book list price times three copies, and in July 2026 a Seoul appeal court set damages at 70% of list price times five copies. It is a usable reference for a negotiation: it ties the price to what the market already pays for the book.

The other anchor is length and quality. The valuation prices edited, published writing at $5,000 to $50,000 per million words, against $1,000 to $5,000 per million words for general text, and the guide to AI training data prices shows where text sits against video, audio and code.

How much reaches the author

For most books the licence fee goes to the publisher, which then pays the author a share. That share is the most disputed number in this market, and it varies more than the fee itself.

What authors actually received tells the same story in cash: $205, $340 and $355 a title from Arcadia's one-time sublicences, under $100 expected at Johns Hopkins, a reported $97 at Taylor & Francis, and $2,500 from HarperCollins.

Opt-in pays more than opt-out

The sharpest pattern in those numbers is how the author was asked. Where authors had to opt in, HarperCollins paid them $2,500 a title; where they were included by default and could only opt out, their share was a fraction of that at Arcadia, Johns Hopkins and Taylor & Francis ( the comparison). A publisher that needs consent must offer enough to get it, and a buyer that wants consent shown pays for it.

The Authors Guild considers 75 to 85 percent to the author a fair split for most deals, depending on the publisher's role, and objects to the 25 percent some academic presses offer. That is an advocacy position, not a market rate, but it is the number authors and agents now bring to the table.

Who holds the AI rights: publisher or author

Before anyone can sell a book for training, someone has to hold the right to grant that use. For a self-published author who never signed a publishing contract, that is the author. For everyone else, it depends on the contract, and most publishing contracts were written before AI training existed.

The Authors Guild's position is that a trade publishing contract grants the publisher the right to publish the book in print, electronic and often audio formats, and that AI training is not among them: in its words, "AI rights remain with the author", and a publisher that wants to license them must seek the author's permission by separate agreement. Some publishers have argued that broad "other electronic" rights clauses already cover it. No court has settled the question.

The publishers' own conduct shows how they read the risk. HarperCollins, Cambridge and Bloomsbury asked authors to opt in; Black Inc asked its authors to sign a contract variation; Johns Hopkins and Arcadia included titles by default and allowed opt-outs. A buyer doing its diligence will want to know which of these happened for every title, because a licence is only as good as the right behind it.

What to check, title by title

1. The grant clause. Which rights the contract grants, in which formats and territories, and whether it reserves "all rights not expressly granted" to the author.
2. Reversion. Whether rights have reverted to the author because the book went out of print. A reverted title belongs to the author, whatever the publisher still lists.
3. Third-party content. Quoted passages, images, maps, forewords and translations may belong to others. Buyers expect third-party content to be identified and removed or licensed.
4. Consent records. Who opted in, when and on what terms. Licence practice calls for a consent ledger per person or per work and a process for later opt-outs, rather than a blanket promise.
5. Earlier licences. Whether the title is already in another AI licence, an aggregator's database or a platform's programme, and whether any of them is exclusive.

The guide to data licensing agreements explains the warranties a buyer will ask the holder to give on each of these, and the guide to training data provenance explains why buyers now pay for a record of where each work came from.

The court cases behind the price

The lawsuits over books are why a licensed title is worth something. They do not say training on books is unlawful; they say acquiring them the wrong way is costly. The cases register follows each one.

Bartz v. Anthropic: training was fair use, piracy was not

Authors sued Anthropic for training on books, some bought in print and scanned, many downloaded from pirate libraries. In June 2025 Judge Alsup ruled on fair use. Training on lawfully acquired books was "quintessentially transformative", and so was converting purchased print copies to digital. But keeping a library of pirated copies was not itself a fair use, and the piracy went to trial. The case then settled for $1.5 billion over 482,460 works, with claims filed for 440,490 of them, and the court gave final approval on 20 July 2026. Bloomsbury alone had 14,087 titles on the list.

The same order held something less welcome to rights holders: the judge assumed a market for licensing books for training could develop, but said it is "not one the Copyright Act entitles Authors to exploit". So the ruling protects the price of a lawful copy, not the price of a training licence. What makes a licence worth paying for is everything else it settles: consent, terms on output, and the certainty a buyer needs before it puts a model on the market.

Kadrey v. Meta and the other suits

In a parallel case, Judge Chhabria gave Meta summary judgment on fair use for training on thirteen authors' books, while stressing that the ruling affected only those plaintiffs; the authors were later allowed to add a claim over torrenting. In the Authors Guild's suit against OpenAI, filed in 2023, OpenAI now argues that a lost training licence fee is not harm the law recognises; no ruling on fair use has been made.

Publishers have started to sue on their own account. Hachette, Cengage and Elsevier, with the author Scott Turow, filed a class action against Google over books allegedly copied to train Gemini. Chicken Soup for the Soul sued eight AI developers at once. Entrepreneur Media sued Meta over pirated books and magazines, and in October 2026 the court ordered Meta to name who approved the downloads.

What the Copyright Office says

The US Copyright Office's report on generative AI training, still a pre-publication version, takes the view that where licensing markets are available, unlicensed uses will be disfavored under the fourth fair-use factor, and that direct licensing is most promising for "corporate entities with catalogs of high-quality and easily identifiable content". A publisher's backlist is the clearest example of such a catalogue. The Office's reading binds no court, but it is the argument rights holders make, and every licence signed strengthens it. The guide to fair use and AI training sets out every ruling so far.

The terms that matter in a book licence

A book licence is short on price and long on conditions. The conditions decide what the price buys, and whether the same titles can be sold again.

• Term. Disclosed book licences run three years at HarperCollins, and most of Cambridge's run 2 to 5 years of training. Settle separately what survives the end: the models already trained, and the deletion of the copy. The guide to licence term length has the options.
• Output limits. HarperCollins' deal reportedly capped what the model may reproduce at 200 consecutive words or 5% of a book. A limit on verbatim output, with filters the buyer must run, is the standard ask for text that would harm the market for the book if a model repeated it.
• Price structure. A fixed fee, a fee per title, or a minimum plus usage. A per-title price makes consent and payouts traceable to each author, and a minimum plus refresh fees keeps a floor under the deal.
• Exclusivity. The book licences whose terms are public are non-exclusive ( Taylor & Francis' licence to Microsoft, for one). An exclusive grant fetches about 2.5 to 5 times a non-exclusive one, and gives up every later sale of the same titles; the guide to exclusive and non-exclusive licences explains when it is worth it.
• Use. Training a model, retrieval that quotes the book in answers, or both. They are different uses with different risks to the book's sales, and a licence should name each it grants.
• Payment timing. A licence usually pays on signature or delivery; the settlement pays over two years. The guide to when buyers pay covers payment and refund clauses.

How to prepare a catalogue for sale

A buyer prices a catalogue by what it can verify quickly. The publishers that sold first had their titles in clean digital files, their contracts in a database and their metadata in order. A smaller holder can get to the same place in a few steps, without sending a single file to anyone first.

1. Count distinct titles, by subject

Buyers count titles, not copies, and they value specialist subjects. List every title you hold the rights to, with its subject classification, language, year and word count. The count of distinct titles in each specialist field is the first thing a buyer asks for.

2. Settle the rights for every title

Go through the checks in section 6, title by title: the grant clause, reversion, third-party content, consent and earlier licences. Put each title in one of three groups: rights clear, rights clear once the author opts in, rights unclear. Only the first two go into an offer. The rights page sets out the questions a buyer's lawyers ask.

3. Ask the authors, and record their answers

Where the AI rights sit with the authors, or where the contract is unclear, ask them in writing, on stated terms, and record each answer. The evidence in section 5 says an opt-in on fair terms gets more titles cleared and holds up better in a buyer's diligence than an opt-out.

4. Prepare the files

Clean digital text is what a buyer trains on: EPUB or well-made PDF, with the text layer intact, front and back matter marked, and third-party material removed or flagged. Scanned print without a text layer needs OCR first, and the buyer will discount it.

5. A sample and a datasheet

Write a short description of the catalogue (titles by subject, word counts, years, rights status, file formats) and pick a small sample of titles whose rights are clear. A datasheet does this in the form buyers expect. The guide to preparing data for sale covers what buyers refuse.

Routes to market

There are four ways to get a catalogue in front of AI buyers. Which fits depends on its size and how specialist it is.

Created by Humans' co-founder expects authors' deals to pay hundreds to thousands per deal, in proportion to each book's sales; its terms say payments may be reduced by processing fees and its own fee. In Korea, Mentat runs the same model for publishers: a searchable repertoire of about 34,000 titles and 8 billion tokens, priced to buyers per token. The buyer board lists the buyers that publish their terms, and the page on going direct or through us sets out the trade-off between selling yourself and using an agent.

How a sale through us works

AIData.Broker represents holders selling or licensing their data to AI companies, a publisher's or estate's catalogue among them. We work for the holder only: we value the catalogue against public prices, settle its rights with the holder, and put it before every buyer that fits at once, so the holder sees competing offers rather than the first one. The holder itself licenses or sells to the buyer, and we do not take ownership of the titles.

1. A valuation first. The valuation asks a few questions about the catalogue (how many titles, what subjects, who holds the rights) and returns a range from the public prices on the board, with no upload.
2. The rights file and the datasheet. We go through the rights questions with the holder, title by title, record the answers, and write the datasheet that describes the catalogue; the holder approves it before any buyer sees it.
3. The round. A blind profile of the catalogue goes to every buyer that fits, without the holder's name, less any buyer the holder leaves out. Buyers who want to go further sign an undertaking and see the datasheet and, where the holder agrees, a sample.
4. Offers and the licence. Buyers make offers, which we set side by side; the offer check reads any written offer against market terms. The holder chooses, and signs the licence with the buyer directly.

The steps and documents are on how it works. Our fee is a success fee on what the holder is actually paid, band by band: 12% of the first $250,000, 10% up to $1 million, 8% up to $3 million and 6% above $3 million, and nothing if nothing sells; on a $300,000 sale it comes to $35,000. The mandate is not exclusive: the fee is due only on a sale to a buyer we introduced, during the mandate and, if that buyer is on the list we send in writing when it ends, within 12 months after it. We run sales only for holders based in the United States, the United Kingdom and Canada.

Which catalogues fit. The smallest sale we take on is $25,000. At the sales median of $2,750 per book or title, a single book does not reach it, and an individual author is usually better served by a rights platform or by the publisher's own programme. The catalogues where a round makes the most difference are a publisher's or estate's backlist of specialist titles with settled rights: professional, technical, medical, legal, scientific and reference publishing, and well-documented nonfiction, where buyers' needs and prices differ widely.

Questions people ask

Can I sell my book to an AI company? Yes, if you hold the AI rights to it, which for a self-published book you usually do and for a traditionally published one depends on your contract. AI companies rarely buy single books from authors; they license catalogues from publishers and platforms. An author with rights can join a rights platform such as Created by Humans, which pays authors 80% of proceeds, or opt in when a publisher offers a licence.

How much do AI companies pay per book? HarperCollins' licence to Microsoft paid $5,000 a title for three years, and across the book sales and payouts on record the median is $2,750 per book or title. Authors' shares are much lower where the publisher included books by default: $205 to $355 a title at Arcadia and under $100 at Johns Hopkins University Press.

How much did HarperCollins pay authors for AI training? $2,500 a title, half of the $5,000 per-title fee, for a three-year licence of select nonfiction backlist, paid only to authors who opted in. One author published his cheque for $2,500, from which his agent took the usual share.

How much does Wiley make from AI licensing? $49.1 million in the year to April 2026, and more than $110 million since its first deal in 2024. The money has moved from books to research content: AI licence revenue from its Learning books segment was zero in the quarter to July 2026, while its Research segment earned $13.7 million.

Did Taylor & Francis sell authors' books to Microsoft? It licensed access to its learning content to Microsoft in 2024, for an initial fee of about $10 million plus recurring payments over three years, and its parent Informa reported more than $75 million from AI partnerships that year. One of its authors reportedly earned $97 for their book.

Do publishers need my permission to license my book for AI? The Authors Guild says yes for a standard trade contract, because AI training is not among the rights granted, and that the publisher must ask by separate agreement. Some publishers read broad electronic-rights clauses differently, and no court has decided it. In practice HarperCollins, Cambridge and Bloomsbury asked authors to opt in, while others included books by default and allowed opt-outs.

How much does the Anthropic settlement pay per book? About $3,000 a work before fees, from a $1.5 billion fund, split by default 50-50 between author and publisher for trade and university press books. It covers only books on the court's works list, which Anthropic had downloaded from pirate libraries, and is paid in instalments.

Can self-published authors license their books for AI training? Yes, and they usually hold every right needed to do so. The practical difficulty is reach: buyers license by the catalogue, so a single author's titles sell through a rights platform or an aggregator that combines many authors, rather than directly to a lab.

What this means for you

Books are one of the few kinds of training data where the rights holder's position has improved since 2024: courts have made pirated copies expensive, the largest labs have paid for licensed ones, and every deal adds a price the next holder can point to. The holders who earn the most have specialist titles, settled rights and clean files, and they choose their terms rather than accepting the first offer.

1. Know the numbers before anyone quotes one. The rates page and the price board hold every public price per title, and the valuation puts a range on a catalogue from them. The text and archives page and what labs buy show where demand is now.
2. Settle who holds the rights. Read the grant clauses, find the reverted titles, and ask the authors where the contract is unclear, on terms that will get a yes. The rights page and the guide to fair use and AI training explain why buyers pay for certainty.
3. Prepare the catalogue. Distinct titles by subject, clean text files, third-party content removed, and a datasheet.
4. Choose the route that fits the size. One author: a rights platform or the publisher's programme. A small press: a larger publisher's sublicensing, or a round if the list is specialist. A publisher or estate with a substantial backlist: competing offers, on a non-exclusive licence with limits on output and a defined term.
5. Read every offer against the market. The offer check compares a written offer with market terms, the guide to data licensing agreements covers the clauses, and the deals register holds every disclosed licence, including the news publishers' deals. The guide to selling company data to AI labs covers the wider market.

AIData.Broker acts as the holder's agent: the holder itself licenses or sells its titles to the buyer, and we do not take ownership of them. The holder pays us a fee fixed in the mandate and paid only if a sale closes, and we run sales only for holders based in the United States, the United Kingdom and Canada.