AI and books: training data, authorship and disclosure

Generative AI touches books at every stage: the text it learned from, the text it can produce, the voices it can synthesise and the reviews it can fake. What has been decided so far, and what is still open.

Last reviewed October 2026 · A fast-moving subject; check the dates on anything cited here

Few subjects in publishing move as quickly as this one, and few attract as much confident speculation. This article sticks to what is on the record: court rulings, official guidance and published platform rules, with dates. Where something is unsettled we say so. It is general information, not legal advice.

Question 1: Is training AI on books legal?

Large language models are trained on very large collections of text, and books are valuable training material because they are long, edited and carefully written. Many authors and publishers object to their books being used without permission or payment, and since 2023 they have brought a series of lawsuits against AI developers, mainly in the United States.

The core legal question in the US is whether training is a fair use. Developers argue that training is a transformative use, like the book scanning in the Google Books case, because the model learns statistical patterns rather than storing and republishing books. Authors argue that the copying is wholesale, commercial and done to build products that compete with them, and that a market for licensing books as training data exists and is being bypassed.

The first rulings

In June 2025 two federal judges in California issued the first substantial rulings on the question, in cases brought by authors.

• In Bartz v. Anthropic, the judge held that training a model on books was a fair use, and that scanning lawfully purchased print books for that purpose was too, but that downloading millions of books from pirate sites to build a permanent library was not excused. The case later settled.
• In Kadrey v. Meta, a different judge also ruled for the developer on the record before him, but stressed that the decision rested on the authors' failure to present evidence of market harm, and suggested that AI-generated books competing with the originals could weigh heavily against fair use in a future case.

These are district court rulings, not appellate precedent, and they disagree with each other on important points of reasoning. Many other cases are pending, including suits by news organisations and by publishers, and courts outside the United States apply different rules. The question is far from settled.

The US Copyright Office's view

The US Copyright Office has published a multi-part report on copyright and artificial intelligence. The part on training, released in pre-publication form in May 2025, concluded that some training uses are likely to be fair and others not, depending on the facts, and gave weight to the effect on markets for the original works, including licensing markets. The Office's reports are not binding on courts, but they are likely to be cited in the litigation.

Licensing

Alongside the lawsuits, a licensing market has started to form. Some publishers have agreed to license books to AI developers, and some have asked authors to opt in. Authors' organisations have argued that such licensing should require each author's consent and a fair share of the payment, and contracts for new books increasingly address AI training explicitly. If you are an author, check what your contract says, and see how trade publishing works for the questions to ask.

Question 2: Can an AI-generated book be copyrighted?

In the United States, the answer for material generated entirely by a machine is no. The Copyright Office's position is that copyright protects only works of human authorship. In March 2025 the US Court of Appeals for the D.C. Circuit agreed, in Thaler v. Perlmutter, holding that the Copyright Act requires a work to be authored in the first instance by a human being; the case concerned an image listed as created autonomously by an AI system.

That does not mean any use of AI makes a work uncopyrightable. The Office's guidance, set out in the part of its report on copyrightability published in January 2025, is that human contributions remain protected: a person's own writing, their creative selection and arrangement of material, and their modifications to AI output can all be protected, while the purely machine-generated parts are not. Applicants registering a work are expected to disclose AI-generated material that is more than minimal. Prompts alone, in the Office's view, generally do not give enough control over the output to make the prompter its author.

Other countries take different approaches, and a few have provisions for computer-generated works. For an author, the practical point is that a book's protection is only as strong as the human authorship in it.

Question 3: Do readers have to be told?

There is no general legal requirement in most countries to label a book as AI-generated, but platforms have set their own rules. Amazon's Kindle Direct Publishing, for example, since September 2023 has asked publishers to say whether a book contains AI-generated text, images or translations when they publish or update it. It distinguishes this from AI-assisted content, such as using a tool to edit or brainstorm, which does not need to be disclosed. Other retailers and distributors have policies of their own, and some limit how many titles an account can publish in a period, partly in response to floods of low-quality generated books.

In the European Union, the AI Act introduces transparency obligations for some kinds of AI-generated content, which are being phased in. How they apply to books in practice is still being worked out.

For readers, the most visible problem has been low-quality generated books, such as imitation guides, travel books and summaries, sold under misleading titles or alongside real books on the same subject, sometimes under real authors' names. The best defence is the usual one: check the author, the publisher and the reviews, and be wary of books with no history behind them. See book reviews online.

Question 4: AI narration and translation

Synthetic narration has made audiobooks cheaper to produce, and several platforms now offer it. It extends audio to books that would never pay for a human narrator, but it raises concerns for narrators about their work and about the use of their voices to train synthetic ones. Retailers generally label synthetic narration.

Machine translation raises similar questions. It could make many more books available in more languages, and literary translators argue that translation is creative work which machines do badly for literature, and which depends on human judgement about voice, culture and nuance. Publishers' practice varies, and some translators' contracts now address AI use.

Question 5: What does AI mean for reading?

AI tools can summarise books, answer questions about them and generate study guides. For some uses, such as navigating a long technical manual, that is genuinely helpful. But a summary is not the book, in the same way that a review is not the book. The research on reading and comprehension suggests that understanding comes from working through a text, not just from having its conclusions. Whether AI tools end up helping people read more deeply or giving them a reason not to read at all is one of the open questions of the next decade.

What is settled, and what is not

Sources

• US Copyright Office, Copyright and Artificial Intelligence (reports on copyrightability, January 2025, and on generative AI training, pre-publication version, May 2025).
• US Court of Appeals for the D.C. Circuit, Thaler v. Perlmutter, No. 23-5233 (18 March 2025); summary at IPWatchdog.
• White & Case, Two California district judges rule that using books to train AI is fair use, June 2025, and IPWatchdog, three key decisions on AI training and copyrighted content from 2025.
• Amazon KDP, content guidelines (AI-generated content disclosure).
• The Bookseller, Amazon revises KDP guidelines to compel disclosure of AI content, September 2023.