Books reach AI training sets in three ways: licensed from the rights holder, bought in print and scanned, or copied from pirate libraries, and only the first pays the author or publisher. The Authors Guild states that HarperCollins's opt-in license pays USD 5,000 per title, split 50-50 between author and publisher (November 19, 2024), while the Bartz v. Anthropic settlement of USD 1.5 billion (final approval July 20, 2026) covers past use and is not a license price.
This page sets out, for books within AI training data licensing, the routes, licensors, terms, figures, training right and disputes. Not licensing is a valid outcome.
The DataPacts Editorial Team explains how we source, date and check every figure.
Why AI Developers Want Book Text
Books give AI developers long, edited, coherent text, and the dataset papers on record show book text as a named share of training mixtures.
Meta's LLaMA 1 paper (February 2023) lists books, from Project Gutenberg and Books3, at 4.5 % of the sampling mixture and 85 GB (Table 1). The Pile paper (December 2020) lists Books3 at 100.96 GiB raw and a 12.07 % weight, derived from the Bibliotik tracker. The RedPajama dataset card (checked November 9, 2026) calls its books configuration defunct due to reported copyright infringement for Books3.
Mercor's publisher listing asks for long-form text in hundreds of titles or hours (checked November 9, 2026). In the Bartz order (Dkt. 231, June 23, 2025), Anthropic hired in February 2024 a former head of partnerships for Google's book-scanning project, who emailed major publishers about licensing in spring 2024.
The parent page compares all data types AI companies buy.
Three Routes From a Book to a Training Set
A book enters a training set by license, by purchase and scanning of a print copy, or by copying from a shadow library, and US courts have treated the three differently.
The table compares the routes by position on the court and government record, as of November 10, 2026.
| Route | Position on record | Example on record | Source |
|---|---|---|---|
| Licensed from the rights holder | documented terms; no ruling requires a license | HarperCollins opt-in license; CLA Generative AI Solution | authorsguild.org; cla.co.uk |
| Bought in print and scanned | digitizing purchased print books is fair use (Dkt. 231, 2025-06-23) | purchased books in the Bartz record | archive.org (court PDF) |
| Copied from shadow libraries | pirated library copies not excused (Dkt. 231); knowing use weighs against fair use (Copyright Office, p. 52) | Books3 196,640 books; at least five million LibGen and two million PiLiMi copies; over seven million kept | archive.org; copyright.gov |
| Library and public-domain corpora | commercial training use on request (IDI) | Institutional Books 1.0, 983,004 volumes; Gutenberg in LLaMA 1 | arXiv; huggingface.co |
In Bartz v. Anthropic, the court held that training and digitizing purchased print books were fair use, but that pirated library copies were not excused. The record counts Books3 at 196,640 books (early 2021), at least five million LibGen copies (June 2021) and two million PiLiMi copies (July 2022), and over seven million pirated copies were kept. Pirated books and shadow libraries have their own page.
The Copyright Office's Part 3 report (pre-publication, May 9, 2025, p. 52) states that knowing use of pirated or illegally accessed works should weigh against fair use without being determinative. Institutional Books 1.0 holds 983,004 volumes (arXiv, June 10, 2025), and commercial use requires contacting the Institutional Data Initiative (checked November 9, 2026).
Who Licenses or Handles Book Rights for AI Developers
Book rights reach AI developers through five kinds of licensors and handlers: publishers licensing directly, collective licensing bodies, author platforms, library initiatives and company data programs.
The table lists each licensor or handler by role, alphabetical within a role, as of November 10, 2026.
| Licensor or handler | Role | What it licenses or handles | Licensee named? |
|---|---|---|---|
| HarperCollins Publishers L.L.C. | publisher, direct | opted-in nonfiction backlist | no |
| John Wiley & Sons, Inc. (Wiley) | publisher, direct | book and learning content; content licensed from other publishers | no |
| Taylor & Francis (Informa PLC) | publisher, direct | academic content and data access | no |
| The Copyright Licensing Agency Limited | collective, licenses on opted-in holders' behalf | books, journals and magazines via PLS and ALCS | n/a |
| Created by Humans | author platform, licenses on authors' behalf | rights-cleared titles | not stated |
| Institutional Data Initiative (Harvard Law School Library) | library initiative; see Institutional Books | Institutional Books 1.0 | n/a |
| Mercor (publisher or education listing) | company data program, acquirer | long-form text, courses or video | n/a |
The Copyright Licensing Agency and Created by Humans license on rights holders' behalf and do not buy books. The Authors Guild calls the HarperCollins buyer "an unknown AI company", Wiley says "large tech company", and Informa names no buyer.
Book License Terms on Record: Fees, Splits, Scope and Output Caps
Only one book license has published terms in any detail: HarperCollins's opt-in license for nonfiction backlist titles, as the Authors Guild described it on November 19, 2024.
The table compares the stated terms of five routes, each value the stating party's own, as of November 10, 2026.
| Route | Pay (as stated) | Scope and consent | Limits or exclusions | Source |
|---|---|---|---|---|
| CLA Generative AI Solution | not published, checked 2026-11-09 | training, fine-tuning, RAG; indemnity; alongside direct deals; authors opt in via ALCS, publishers via PLS | exclusivity, term, deletion not stated | cla.co.uk; pls.org.uk |
| Created by Humans, Author AI rights licensing | not published (dashboard), checked 2026-11-09 | training, reference, transformative rights; author or agent opts in per use | titles rights-cleared | createdbyhumans.ai |
| HarperCollins (per the Authors Guild, 2024-11-19) | USD 5,000 per title, split 50-50 | nonfiction backlist; separate opt-in addendum | outputs max 200 consecutive words and/or 5 %; no scraping of piracy sites; not set against advances | authorsguild.org |
| Institutional Books 1.0 (IDI) | not stated | commercial and AI training use by contacting IDI | noncommercial Early-Access terms otherwise | huggingface.co |
| Mercor publisher or education listing | Mercor states USD 35,000 to 370,000 per company, checked 2026-11-09 | rights the company holds; hundreds of titles or hours | not a fit: others' licensed content; open access | Mercor listing feed |
The Authors Guild states that each author opts in through a separate addendum, payments are not set against advances, outputs are capped at 200 consecutive words and/or 5 %, and the licensee may not scrape piracy sites. It calls the split "far too much to the publisher".
Other routes publish scope but not price: the Copyright Licensing Agency states indemnity protection and a design to work in harmony with direct deals. Exclusivity, term and deletion, blank in the table, are clauses set out in every clause of an AI data licensing agreement.
What a Book Earns: License Fees, Publisher Filings and a Class Settlement
No market rate exists for books: the figures on record are one per-title fee stated by an authors' body, one program's deal range, publishers' filed revenue from unnamed buyers, and a class settlement that is not a license price.
The table compares the figures by what each one is, with the stating party and date, as of November 10, 2026.
| Figure | Value | Stated by, date | What it is | Source |
|---|---|---|---|---|
| HarperCollins per-title fee | USD 5,000, split 50-50 | Authors Guild, 2024-11-19 | license fee | authorsguild.org |
| Mercor publisher listing | USD 35,000 to 370,000 | Mercor, checked 2026-11-09 | stated deal range per company | Mercor feed |
| Wiley first GenAI project | USD 23 million | Wiley, 2024-06-13 | one-time project, buyer unnamed | sec.gov |
| Wiley second project | USD 21 million; Q1 FY2025 USD 16 million (release) vs 17 million (presentation) | Wiley, 2024-09-05 | project value, two own figures | sec.gov |
| Wiley AI license revenue, six months to 2025-10-31 | USD 34.9 million, incl. USD 15.7 million from content licensed from other publishers | Wiley 10-Q | filed revenue | sec.gov |
| Wiley AI license revenue, Q1 FY2027 | USD 13.7 million vs 28.9 million | Wiley 10-Q | filed revenue | sec.gov |
| Informa AI-related revenue 2024 | USD 75m+ | Informa, FY2024 | "non-recurring" data access revenue | investegate.co.uk |
| Bartz gross per-work estimate | approximately USD 3,000 | Dkt. 680, 2026-07-20 | settlement of past claims | court PDF |
| Bartz initial net distribution | approximately USD 2,203.56 per Claimed Work | Class Counsel, Dkt. 688, 2026-09-02 | first distribution, not final | court PDF |
| US statutory damages | USD 750 to 30,000 per work; up to 150,000 if willful | 17 U.S.C. 504(c) | court remedy | govinfo.gov |
The per-title fee describes one license, and Mercor's range is for a whole company, not a title. Wiley's USD 23 million (June 13, 2024) and USD 21 million (September 5, 2024) are project values from unnamed buyers, and its filed revenue includes USD 15.7 million related to content licensed from other publishers (six months to October 31, 2025). For the quarter ended July 31, 2026 the 10-Q shows USD 13.7 million against 28.9 million a year earlier, and Informa calls its USD 75m+ for 2024 non-recurring. No per-title rate is derived from a project value.
The Bartz v. Anthropic settlement is a USD 1.5 billion non-reversionary fund with final approval on July 20, 2026 and 482,460 works on the Works List. It settles past claims only, and the initial net distribution is not final; first payments are expected on or before November 15, 2026 (settlement website, checked November 9, 2026).
Who Holds the AI Training Right in a Book
The publisher signs a book license, but whether it holds the AI training right depends on the publishing contract, and the Authors Guild holds that standard trade contracts grant no AI training rights.
It recommends 75 to 85 % of training license income for authors (December 12, 2024), which is advocacy, not a rule.
The Bartz settlement divided each per-work share among copyright owners by elected default splits for non-education works, or by contract or publishing agreement (Dkt. 680). Under 17 U.S.C. 201(b), the employer or other person for whom a work made for hire was prepared is considered its author and owns all rights unless both parties agree otherwise in a signed written instrument.
A rights audit before licensing reads the publishing contract and records who holds the right. This section is information, not legal advice.
Book Disputes: What Authors and Publishers Allege and Where Each Case Stands
Eleven US cases and case groups on record allege that AI developers trained on books without a license; one has settled, one training claim was decided for the developer on its record, and the rest are pending as of November 9, 2026.
The table compares the eleven cases by filing date, with court, docket and status as of November 9, 2026 (the xAI row groups three separate suits).
| Case | Court and docket | Books at issue (as alleged) | Status as of 2026-11-09 |
|---|---|---|---|
| In re Google Generative AI Copyright Litigation | N.D. Cal. 5:23-cv-03440 | books and images; publishers intervened | pending |
| Kadrey v. Meta | N.D. Cal. 3:23-cv-03417-VC | shadow-library books for Llama; 13 authors | training decided for Meta on its record (Dkt. 598); related trial 2027-05-03 |
| Authors Guild v. OpenAI | S.D.N.Y. 1:23-cv-08292-SHS-OTW (MDL No. 3143) | books for GPT training | summary judgment briefed; no ruling |
| Huckabee v. Bloomberg | S.D.N.Y. 1:23-cv-09152-MMG | Books3 for BloombergGPT | motion to dismiss denied 2025-11-24 |
| Nazemian v. NVIDIA | N.D. Cal. 4:24-cv-01454-JST | Books3 and alleged shadow libraries | fact discovery closes 2027-02-12 |
| In re Mosaic LLM Litigation | N.D. Cal. 3:24-cv-01451 | books datasets for MPT and DBRX | cross-motions for summary judgment pending |
| Bartz et al. v. Anthropic PBC | N.D. Cal. 3:24-cv-05417-AMO | books incl. Books3, LibGen, PiLiMi copies | settled; final approval 2026-07-20 (Dkt. 680) |
| Hendrix v. Apple | N.D. Cal. 4:25-cv-07558-YGR | books for Apple's models | consolidation stipulated 2025-11-14 |
| Tanzer v. Salesforce | N.D. Cal. 3:25-cv-08862-CRB | The Pile, RedPajama, Books3 | pending |
| Elsevier, Cengage, Hachette, Macmillan, McGraw Hill and Scott Turow v. Meta Platforms, Inc. et al. | N.D. Cal. 3:26-cv-11415-VC | willful infringement in training Llama | pending; no trial date |
| Cognella, Chicken Soup for the Soul and Cambronne v. SpaceX Super Intelligence Corp. et al. (authors and publishers v. xAI) | N.D. Cal. 3:26-cv-13189, -13188, -13187 | books for Grok, alleged pirate-library copies; filed 2026-11-04 | no ruling |
Kadrey v. Meta (Dkt. 598, June 25, 2025) granted summary judgment to Meta on training on that record, stating the plaintiffs are "not entitled to the market for licensing their works as AI training data". The court calls market dilution "potentially winning" but undeveloped, and its ruling "does not stand for the proposition that Meta's use ... is lawful".
In the OpenAI MDL, an order of October 6, 2026 (Dkt. 2090) struck an expert's supplemental report and reliance on the paper "Generative AI Floods and Dilutes the Market for Books", without opposition. In pending cases the table states allegations, not findings.
Weighing a Book License: Benefits, Risks, Downsides and Alternatives
A book license turns backlist titles into income with written limits on use, but author and publisher may disagree on who holds the right and how to split the fee, and project revenue has not recurred.
Benefits:
- Per-title income from backlist titles, as in the HarperCollins license.
- Written limits on use: an output cap and a no-piracy pledge (November 19, 2024).
- Indemnity and one negotiation through the Copyright Licensing Agency.
Risks:
- The publishing contract may not grant AI rights, so the publisher may lack the right to license.
- The fee split may be disputed between author and publisher.
- The buyer may stay unnamed, and the Kadrey court held the plaintiffs not entitled to the licensing market on that record.
Downsides:
- Revenue can fall away: Wiley's quarterly AI license revenue fell year on year, and Informa calls its 2024 data access revenue non-recurring.
- Fees on the collective and platform routes are not published (checked November 9, 2026).
Alternatives:
- Opt in title by title.
- Join a collective license through ALCS and PLS.
- Use an author platform that decides per use.
- Not license at all, as set out in alternatives to selling data, including not selling at all.
Neighboring Text Types and the Publisher's and Author's Next Steps
This page covers books as a data type; the publisher's and author's licensing process, other text and the case tracker each have their own page.
Neighboring pages:
- Book publishers and authors licensing books
- Academic and scientific publishers
- Collective licensing for AI training
- Libraries, archives and museums
- Text corpora
- Scientific data and papers
- Speech and audio data: Sieve lists audiobooks among wanted content (checked November 9, 2026)
- Books3
Next steps:
- AI copyright lawsuits over training data
- AI data licensing deals
- Copyright and AI training across jurisdictions
- When AI data buyers pay
Frequently Asked Questions
How much do AI companies pay per book?
No market rate exists; the one per-title fee on record is USD 5,000, split 50-50, which the Authors Guild states for HarperCollins's opt-in license (November 19, 2024). The Bartz figures are a settlement of past claims, not a price. Other types are compared in AI training data prices per work, hour, image.
Was my book used to train an AI model?
No public register lists which books trained which model. The Bartz Works List held 482,460 works; its claim deadline passed on March 30, 2026. EU AI Act training summaries disclose licensed data only as Yes or No by modality (Art. 53, Regulation (EU) 2024/1689, applying from August 2, 2025).
Can an author license a book for AI training without the publisher?
Only where the author holds the training right, which the publishing contract decides. The Authors Guild holds that standard trade contracts grant no AI training rights. The Copyright Licensing Agency takes author opt-ins via ALCS, and Created by Humans works with authors and agents (checked November 9, 2026).
Is training AI on books fair use?
Two US courts found training on books fair use on their own records, and the Bartz court did not excuse pirated library copies. Bartz Dkt. 231 (June 23, 2025) and Kadrey Dkt. 598 (June 25, 2025) decided that way, and Kadrey states its ruling does not make Meta's use lawful. The fair use and AI training in US law page sets out the four factors.
Literary agents and advisers can use the introducer form.