📝 AI clauses in a publishing contract: what to include and what to refuse

Publishing agreement is almost never read twice, and that is exactly why the line about "all rights, known now and those that may appear in the future" quietly takes from the author what he never even thought to negotiate. When it comes to placing articles, a platform for guest publications helps, where the terms are transparent from the first screen, but a book contract works differently: every clause about artificial intelligence appears there only because someone wrote it in. Below is a breakdown of which formulations to keep on hand during negotiations and why an empty space in a contract today costs more than a bad advance.

How to check a contract for hidden AI risks in one evening

The first pass through a contract does not require a lawyer. You need a list of formulations and the habit of noting not what is written, but what the text is silent about.

💡 Quick overview:

• Step 1: find the words about "any technologies" and "formats that may appear in the future" in the rights transfer section, and make sure model training is explicitly excluded from it.
• Step 2: check the subsidiary rights section to see whether model training, generating responses based on the book's text, and dialogue with a character are each named separately.
• Step 3: make sure the audio version, translation, and cover require the author's written consent if they are produced by a machine.
• Step 4: check whether there is a prohibition on publisher employees uploading the manuscript to public chatbots.
• Step 5: set a maximum share of the author's own AI use and the disclosure procedure.
• Step 6: add a rights reversion clause if the training license never produced payments within the agreed period.

Six points cover roughly nine tenths of the typical claims authors make against publishers after the fact. The rest concerns truly exotic scenarios like synthetic interviews with the "author's voice," and those are worth discussing separately.

Why an empty space in a contract is more dangerous than a bad rate

The classic logic of copyright is simple: what is not transferred remains with the rights holder. In practice, publishing templates have been written for decades to capture as many future formats as possible, and the publisher's lawyers will honestly say that training a neural network formally falls under "reproduction in electronic form." A dispute in such a situation is resolved not by the letter of the law, but by whoever wrote in the clarification first.

The Authors Guild in the US took exactly this path and proposed ready-made formulations that an author or agent can bring to the negotiating table. The updated package was published on April 29, 2026 and is built around one basic idea: the publisher receives no rights to model training or to generating derivative results until those rights are transferred by a separate clause. Everything else is built on top of that foundation.

The logic resembles how audiobook rights were once won back. While the format was considered marginal, it was given away as a package along with "other electronic rights," and when the market grew, it turned out that recovering what had been assigned was nearly impossible. The only difference is that the market for training licenses is growing faster than contract templates can be updated.

Four doors through which AI enters a book without asking

An author usually keeps one threat in mind: his text will be fed to a model. There are more actual entry points, and three of them look so mundane that they raise no alarm at the signing stage.

First door: the audio version. Synthesized narration is many times cheaper than studio recording, and it is profitable for the publisher to make it the default for the backlist. Second: translation. A machine draft with light editing has already become the norm for small language markets, while the author's name remains on the cover. Third: the cover itself, which is increasingly assembled by an image generator. Fourth, the least noticeable: an editor uploads the manuscript to a public chatbot to write the blurb faster, and the text goes to a third-party server even before release.

The Guild's set of recommendations closes each door with a separate clause. In the list of model clauses, there are provisions for audiobooks, translation, and artwork, each built on the author's written consent rather than after-the-fact notification. Separately, it prohibits uploading the manuscript to commercial AI services without permission and prohibits substantive machine editing of the text.

The difference between "we will notify" and "we will obtain consent" seems like a matter of wording until a conflict arises. Notification does not give the right to stop the process; consent does. The British Authors' Association told a parliamentary committee about cases where authors were signed up for AI licenses without a clear explanation of their rights, and payments for a book turned out to be symbolic.

What the right to train a model costs

When an author is willing to license the use of the text, the question shifts to revenue shares. The logic here differs from the usual 50/50 split of subsidiary rights, and the reason is clear: when training a model, the publisher does almost nothing except transfer the file. The Guild offers revenue-sharing benchmarks where the author's share for training rights ranges from 80 to 90 percent.

The table is useful not in itself, but as an argument. When a publisher proposes splitting revenue 50/50 "as usual," the author gains a public industry benchmark to lean on rather than a vague sense of unfairness. A second argument from the same series: the right to audit payments and the right to reclaim the license if the deal does not close within the agreed timeframe.

It is worth keeping the other side in mind as well. The market for training licenses is still narrow, there are few large deals, and inflated income expectations lead authors to hold rights unused for years. A reasonable middle-ground position: do not give rights away for free, but do not treat them as a guaranteed revenue source either.

What an author must disclose about their own use of AI

The clauses work both ways. The publisher needs assurance that the "author's original text" was actually written by a human, otherwise the entire chain of warranties collapses, including rights registration. The model clauses require disclosure of generative tool use and limit the share of generated text, with the threshold set at five percent of the manuscript or as an insignificant volume without an exact figure.

It is important to distinguish between tools here. Spell checking, source research, interview transcription, and brainstorming title options do not turn a book into a machine-generated work. The issue is fragments of finished prose that entered the file from a generator. The practical advice is simple: keep a separate file with notes on where and how AI was used. Such a log takes five minutes a week and resolves half the questions when submitting the manuscript.

The legal backdrop makes this discipline more than a formality. The US Copyright Office has consistently held that purely machine-generated output is not eligible for protection, and protection arises only where there is human contribution in selection, arrangement, and transformation. Current materials on the topic are collected in the office's dedicated section, and they directly affect which parts of a book can even be considered protected.

A real case that changed the negotiating position

Abstract arguments about rights turned into numbers in the fall of 2025. Anthropic agreed to pay one and a half billion dollars to settle a class action brought by authors whose books ended up in training datasets sourced from pirate libraries, which worked out to about three thousand dollars per book plus interest. The judge preliminarily approved the settlement on September 25, 2025, as reported by the business channel CNBC, and it became the largest publicly known copyright recovery.

The practical takeaway for negotiations is unexpected. The point is not the size of the payment, but that the market now has a price benchmark where the parties previously argued in the dark. It is now harder for a publisher to convince an author that training rights are worth nothing, and harder for an author to demand sky-high rates, because the order of magnitude has become public.

The mood in the author community confirms this. A Guild survey showed that nine out of ten writers consider compensation mandatory when books are used to train generative systems, and about two thirds of respondents supported the idea of collective licensing. The Australian Society of Authors, in a 2025 study, recorded an increase in the share of those who have already lost work because of generative tools, from four percent to six percent in a year.

What to do if the publisher refuses to change the template

A refusal to amend a template is rarely a matter of principle. More often, the editor simply lacks authority, and getting sign-off from the legal department drags on for weeks. A few techniques speed up the process.

Bring a complete clause, not a request to "come up with something." A ready-made paragraph gets approved faster than an open-ended question. Separate your demands by priority: a basic ban on training without an explicit transfer of rights matters more than perfect percentages for character dialogue. Offer a review period if the publisher is not ready to lock in terms forever; an annual review of the AI section works for both sides. And keep a reversion right: a license that nobody has used for two years should return to the author automatically.

It is also useful to understand the other side's motivation. The publisher also has no interest in seeing the book end up in a training dataset for free, because that devalues its own catalog. An analysis of how to read these clauses from a working author's perspective is laid out in detail in an industry blog, and it shows that the zone of shared interests is wider than it seems at the start of a dispute.

A separate question concerns old contracts. Clauses do not work retroactively, but training rights generally remain with the author if they were not explicitly transferred. That means that when renewing a contract or signing one for a reissue, there is a legitimate reason to reopen the rights section.

⁉️🤔 Frequent questions

Is it mandatory to prohibit any use of AI by the publisher?

No, and a total ban often hurts the author themselves. The sensible line runs between internal operational tasks and creating derivative versions of the book. Catalog search, ad audience targeting, and internal analytics usually stay with the publisher, while audio narration, translation, design, and model training move into the category of things that require written consent.

What should I do if the book has already been published under an old contract?

Check whether training rights were transferred explicitly. If the text contains only a general clause about future technologies, the author retains a strong negotiating position. A reason to reopen the section appears when the term is extended, when a reissue is being prepared, or when the book comes out in a new format.

How realistic is getting an 80 to 90 percent share for training?

The benchmark is industry practice, not law, and major publishers negotiate. Experience shows that a conversation about specific numbers goes more smoothly if the author cites a public benchmark and is simultaneously ready to discuss other terms, such as the license duration or the right to audit payments.

Does using AI for editing count as a disclosure violation?

Grammar checking, transcribing audio recordings, and finding synonyms usually do not fall under disclosure. The issue is ready-made passages of prose generated by a machine and included in the manuscript. Keep a working log of tool use, and answering an editor's question will take a minute rather than an evening.

Do I need a lawyer to amend these clauses?

For a first pass, a checklist and ready-made wording are enough. A lawyer is needed when the publisher proposes a counter-draft with unfamiliar definitions, and also for package deals covering several books at once. A literary agent often handles this task as part of standard representation.

Summary

AI clauses have stopped being exotic and have become as essential a contract section as territory or the term of rights transfer. A simple rule applies: what is not explicitly transferred remains with the author, and the goal of negotiation is to get that rule written down in words rather than left implied.

Take the six checks from the beginning of this article and go through your current contract today. Mark the places where the text is silent, write out the wording you need, and send it to the editor in a single email. If you do not have a contract in hand right now, save the list for your next deal, and also take a look at our other breakdowns of publishing practice; they will help you build your negotiating position in advance.