Can AI Companies Train on My Published Books?
By Alexandru FilipUpdated Facts verified
Legally it is unsettled and being fought in court right now. Practically, a great deal of published fiction and non-fiction has already been used. No general opt-out exists for books already published. What you can still control is work that has not been published yet, and whether you have the standing to join an action if one applies to you — which in the US means having registered your copyright.
The longer version
Three separate questions get collapsed into one. Pulling them apart is the only way to get a useful answer.
1. Is training on copyrighted books legal? Contested. The central argument is whether training is transformative fair use. In 2025 two US district courts, in cases brought by authors against Anthropic and Meta, found training on books to be fair use while treating how the books were obtained separately, and copying from pirated libraries far less sympathetically. The Anthropic case then settled for $1.5 billion over the pirated copies, with final court approval on 20 July 2026; it does not decide whether training itself is lawful. The first appeals-court ruling came in September 2026, when the Third Circuit held that training a legal-research tool on Thomson Reuters' headnotes was not fair use. That tool did not generate text, and no appeals court has yet ruled on generative AI trained on books. This area is moving quickly and any summary, including this one, dates fast.
2. Did it already happen? For a large share of published books, yes. Datasets assembled from pirated book collections circulated widely for years and were used in training runs before the legal questions were seriously tested. A 2025 University of Cambridge study found 59% of surveyed published novelists believed their work had already been used without permission.
3. Can you stop it going forward? Partially, and unevenly.
What actually gives you leverage
- Register your copyright. In the US you cannot sue for infringement of a US work until it is registered, and statutory damages require registration before the infringement or within three months of publication. This is also what determines whether you are eligible for class settlements. It is the single most concrete step available. If you have filed a settlement claim, expect emails impersonating the administrator — see how settlement-notice scams work
- Join an authors' organization. The Authors Guild and equivalents run collective actions and negotiate licensing frameworks. Individual authors have very little leverage; organizations have some
- Read your publishing contracts. New contracts increasingly contain AI clauses. Some grant AI training rights outright. This is negotiable and worth negotiating
- Control what has not been published. Drafts, unreleased series, notes. Nothing has been decided about these yet, and where they are stored is entirely your choice. See How to Keep Your Manuscript Private in the Age of AI
- Set robots.txt on your own website. This governs your site, not your books — but if you post excerpts, sample chapters or a serial, it is the difference between offering them to crawlers and not
What does not work
- Adding a "no AI training" notice to your copyright page. It has no established legal effect on training in the US, though it costs nothing and states your position. In the EU an express reservation of rights matters more, but whether a copyright-page notice qualifies has not been settled in court
- Asking for removal from a trained model. Models are not databases; there is no established mechanism to extract one author's contribution
- Opting out retroactively. The datasets already exist and have already been used
The honest position
Authors have been asked to accept a situation almost none of them consented to. A December 2023 Authors Guild survey of its members found 96% believed consent should be required and authors paid; the Cambridge study found 93% would probably or definitely opt out if an opt-out existed. Almost none were offered one.
The realistic response is a split: pursue the collective and legal routes for published work, because that is where any remedy will come from — and treat unpublished work as the part you can still decide about.
Common exceptions
- Licensing deals now exist. Several publishers have signed agreements permitting training in exchange for payment. If you are traditionally published, whether your books are covered depends on your contract.
- Public domain works can be trained on freely.
- Jurisdiction matters. The EU's approach, including text-and-data-mining exceptions and reservation of rights, differs from the US fair-use analysis.
- Self-published authors hold their own rights, which means no publisher can license their books away — and also that no publisher is litigating on their behalf.
Related questions
Sources
- Authors Guild, AI survey of members, December 2023.
- University of Cambridge, Minderoo Centre for Technology and Democracy, survey of published novelists, 2025.
- US Copyright Office, Circular 1 and 17 U.S.C. §§411–412 on registration and remedies.
- Bartz v. Anthropic (N.D. Cal.): fair-use ruling, June 2025; final approval of the settlement, 20 July 2026. Kadrey v. Meta (N.D. Cal.), June 2025.
- Thomson Reuters v. Ross Intelligence (3d Cir.), decided 29 September 2026.
This is general information, not legal advice, and the litigation landscape changes frequently. Verified 17 November 2026 — check current status before relying on it.
Related AuthorAZ resources
Check this against the source
Publishing Database
Rights & Privacy