
Picture a board meeting at a mid-sized publisher. An AI developer has made an offer for the backlist. The fee is real money, the deadline is close, and somebody asks a perfectly reasonable question: “What’s our policy on this?” At a good number of houses, the honest answer is that there isn’t one yet.
That house would be in plenty of company. When BISG and BookNet Canada surveyed 771 people across the North American book industry in the summer of 2026, 49% said their organization had an AI policy or guidelines, up from 31% the year before (BookNet Canada). That is real progress, but it still leaves roughly half without one. For organizations with 21 to 50 employees the figure was 29%. And where policies did change over the past year, only 23% of those changes touched contracts, rights and royalties. Just 4% of organizations use AI in their rights and licensing work at all.
The offers, meanwhile, keep arriving. The UK Publishers Association’s 2026 report expected the number of publishers active in AI licensing to “almost double” that year, and described retrieval licensing as “a meaningful and growing segment” (STM summary). So I suspect many boards will be asked to approve a deal before they have settled what they think about AI more broadly.
What follows is a short rubric a board can work through in a single meeting. It rests on four tests (rights, brand, proof of origin and recurring revenue), with a checklist of terms for counsel and a few alternatives worth weighing before anyone signs.
It helps to remember that a deal signed without a policy quietly becomes the policy. The terms you accept, and whatever you tell your authors, will shape every conversation that comes after. Before negotiating, I’d suggest the board agree on a single page that answers five questions.
First, what is licensable? That means which imprints, formats and rights-cleared titles are in scope, and which are left out (trade titles, for instance, where AI rights often sit with authors).
Second, how will authors be treated? You can ask them to opt in, give them a way to opt out, or rely on existing contracts, and you need to decide what share they receive. Cambridge University Press asks authors to opt in and pays them 20% of net AI licensing revenue (Authors Alliance). Wiley, by contrast, told authors they could not opt out (The Bookseller). Either approach can be defended, but each one shapes your relationship with authors for years.
Third, which uses will you license: training, retrieval (often called RAG), or both? Fourth, what is your floor, the terms below which you would rather walk away? And fifth, who has authority to sign, and who needs to be consulted first? Editorial, rights, legal and author relations will all have a view.
I’d score each of the four tests 0 if it isn’t met, 1 if it’s partly met and 2 if it’s fully met. Please treat the thresholds at the end as a starting point for discussion. They aren’t an industry standard, and every house will want to adjust them.
The first thing to check is whether your author contracts give you the right to license this use at all. Many contracts signed before 2023 say nothing about AI. The Authors Guild’s April 2026 model clauses treat AI training, retrieval, AI summaries and “chat with a book” apps as separate subsidiary rights, and none of them passes to the publisher by default (Authors Guild). Agents will increasingly bring that language to the table.
Check, too, whether third-party material you don’t control, such as images, epigraphs or contributed chapters, has been carved out, and whether the warranty you’re being asked to give matches what you can genuinely verify. If the draft asks you to warrant that every title is fully cleared, backed by an uncapped indemnity, I would treat that as a reason to pause.
Here the questions are about how your work appears once it’s inside someone else’s product. Will AI outputs attribute or link back to you when they draw on your content? Are there limits on how much text an output can reproduce? HarperCollins’ reported terms, for example, capped outputs at 200 consecutive words and 5% of a book per user session (Josh Bernoff).
You’ll also want to know whether the licensee can use your name or imprint in its marketing without approval, and whether its product could end up answering the very questions your reference works or courses are sold to answer. A deal with no attribution, no output limits and nothing to stop a competing product deserves a careful second look.
This test is about visibility. Will you receive usage reporting by title, or only a single payment? Do you have audit rights? When the term ends, does the licensee have to delete the dataset and confirm in writing that it has done so?
There’s one point boards are sometimes surprised by. Under HarperCollins’ reported terms and under Cambridge’s model, licensees may keep models they have already trained after the term ends. That may be perfectly acceptable to you, but it’s far better to know it before signing than to discover it later. If a deal offers no title-level reporting, no audit right and no obligation to delete, you’ll have very little to show your authors, or yourselves, about what happened to the work.
Finally, the money. Is the payment a one-off, a fee per term, or tied to usage? What happens at renewal, and is there any real reason for the licensee to renew once its model is trained? And how does the price compare with what’s known?
The public reference points are thin. HarperCollins’ deal paid $5,000 per title over three years, split equally with authors (Quill & Quire). You may also see a figure of roughly $3,000 per work from the Bartz v. Anthropic settlement, but that was a damages payment for pirated copies, not a licence price (Authors Guild).
What we can see so far suggests that training revenue tends to arrive in lumps. Informa reported more than $75 million of non-recurring AI data revenue at Taylor & Francis in 2024, less in 2025, and no new contracts of that kind in the first half of 2026. It now sets its growth targets with those contracts left out (Informa H1 2026). Wiley’s AI revenue grew to $49 million in fiscal 2026, though its AI licensing revenue from book content fell while research and corporate licensing grew (Wiley FY2026).
Retrieval licensing, which is paid per use, looks like the model most likely to keep paying. It’s still very new, though. The Scholarly Kitchen reported in March that one publisher had changed its retrieval pricing four times in three months (The Scholarly Kitchen). If the offer in front of you is a one-off training fee, with the licensee keeping the model indefinitely and nothing linked to usage, ask whether the price reflects everything you’re giving up.
Total (0 to 8) | A suggested response |
7 to 8 | Proceed, subject to counsel’s review |
4 to 6 | Take the weaker tests back to the licensee and negotiate before signing |
0 to 3 | Decline or defer for now, and look at the alternatives below |
One caveat matters more than the totals. A deal that scores well on revenue but zero on rights shouldn’t go ahead, whatever it adds up to. Rights is the one test I don’t think can be traded against the others.
Term | What to ask |
Licensed use | Is it training, fine-tuning, retrieval, or all three? Each is worth pricing separately |
Scope | Which titles, formats and editions? Is new content added automatically? |
Exclusivity | Non-exclusive by default. If exclusive, what is the premium, and for how long? |
Term and renewal | How long does it run, how does renewal work, and what triggers renegotiation? |
Model retention | Can the licensee keep or use models trained during the term after it ends? |
Data handling | How is the content stored and secured, and is deletion at the end certified? |
Sublicensing | Can the licensee pass the content, or models trained on it, to anyone else? |
Output controls | Are there limits on verbatim reproduction, and is attribution required? |
Reporting and audit | Will you get usage data by title, and can you audit it? How often? |
Price and adjustment | What is the fee basis? Is there a minimum guarantee, a most-favoured-nation clause or a usage escalator? |
Warranties and indemnities | What are you warranting about rights, how is liability capped, and who defends claims? |
Author treatment | How is consent handled, what share do authors receive, and how are payments reported to them? |
Termination | What happens on breach, and what becomes of the content and trained models afterwards? |
Governing law | Which jurisdiction applies, given that US, UK and EU law are moving in different directions? |
Leverage matters here. In June 2025, two US district court rulings, in Bartz v. Anthropic and Kadrey v. Meta, found that training AI models on books was fair use on the records in those cases (Authors Alliance). That weakens a publisher’s hand in a straightforward training deal. Then in September 2026 the Third Circuit upheld the ruling against ROSS in Thomson Reuters v. ROSS, a case where the AI product competed directly with the original (IPWatchdog).
In practice, publishers seem to have the most bargaining power where an AI product would compete with what they already sell, as with reference, legal, clinical and professional content. They have the least where a developer simply wants text to train a general model. It helps for a board to know which of those two positions it’s negotiating from.
Licensing is one option among several, and I’d encourage boards to set it alongside a few others before deciding.
The simplest is to wait. If an offer scores between 0 and 3, the cost of holding off is usually modest, because prices and terms are still settling. A second option is to license for retrieval rather than training. A usage-based licence keeps you involved for as long as the product relies on your content, and it’s more likely to bring in repeat income than a single training fee.
A third route is collective licensing. In the UK, the CLA, working with ALCS and PLS, has developed a collective generative-AI licence that covers both training and retrieval (CLA). It can make deals more practical for smaller publishers, though published rates and payouts are still hard to find.
A fourth is to build learning products from the titles that teach. Professional, technical and educational books can become courses sold to employers and learners, and the publisher keeps both the brand and the relationship with the learner. Courses do carry real costs in production, distribution and keeping material current, so I wouldn’t present them as a free alternative. The demand is real, though. US companies spent $16 billion on outside training products and services in 2025, a rise of 29% (Training magazine), and Pearson’s enterprise learning division grew profit 40% in a year when the group grew 6% (Pearson).
Finally, some publishers are building their own retrieval-based tools on their content rather than licensing it out. It’s the most expensive route and needs the most technical capacity, but it keeps the most control.
Most catalogues aren’t all one thing, so many publishers will probably combine these. They might license archival and reference material for retrieval, for instance, while building learning products from the titles that teach.
Before the vote, it may help for each board member to answer a single question for themselves: if this turns out to be the only AI deal we ever sign, are we comfortable with the precedent it sets for our authors, our brand and our revenue? If the answer is no, the most useful next step probably isn’t to sign. It’s to write the policy and score the deal again.
Four things: whether you hold the rights to grant the use, how your brand will be credited, whether you can see how your content is used, and whether the income will recur. Alongside those, ask counsel to review model retention, data deletion, audit rights, warranties and how authors are treated.
It helps a great deal. A deal signed without a policy tends to become the policy, because its terms set the precedent for later deals and for what you tell authors. In BISG and BookNet Canada’s 2026 survey, about half of book-industry organizations still had no AI policy or guidelines.
A training licence lets a developer build your content into a model, usually for a fee per term. A retrieval (RAG) licence lets an AI product look up your content when it answers a question, and it can be paid per use for as long as the product relies on your work.
Few deals are public. HarperCollins’ 2024 deal paid $5,000 per nonfiction title for a three-year training licence, split equally between author and publisher. Wiley reported $49 million of AI revenue in fiscal 2026. The roughly $3,000 per work in Bartz v. Anthropic was a settlement, not a licence price.
Training fees usually aren’t. Informa classes Taylor & Francis’s AI data deals as non-recurring and signed no new ones in the first half of 2026. Retrieval licences, paid per use, are more likely to recur, although pricing for them is still settling.
Often they can keep models already trained on it. Under HarperCollins’ reported terms and Cambridge University Press’s model, the dataset is deleted at the end of the term but trained models are kept. That is why model retention and certified deletion belong on every term sheet.
It depends on the contract and on the publisher’s own approach. Cambridge University Press asks authors to opt in and pays them 20% of net AI revenue, while Wiley told authors they could not opt out. The Authors Guild’s 2026 model clauses give publishers no AI rights by default.
Publishers can wait while terms settle, license for retrieval rather than training, join a collective licence such as the UK CLA’s, build learning products from titles that teach, or build their own AI tools on their content. Many will combine several of these.
BookNet Canada and BISG, AI use across the North American book industry 2026, 2026.
STM Association, UK Publishers Association publishes comprehensive account of AI content licensing, 17 March 2026.
Authors Alliance, AI licensing: an interview with Ben Denne of Cambridge University Press, 17 March 2025.
The Bookseller, Wiley set to earn $44m from AI rights deals, confirms no opt-out for authors, August 2024.
Authors Guild, Authors Guild releases new AI model clauses and issues updates, 29 April 2026.
Josh Bernoff, Why I took the $2,500 check to license my book for AI, 15 April 2025.
Quill & Quire, HarperCollins signs AI training deal, November 2024.
Authors Guild, Court grants final approval of Anthropic copyright settlement, 21 July 2026.
Informa plc, 2026 Half Year Results, 30 July 2026.
John Wiley & Sons, Fiscal 2026 fourth quarter and full year results, 16 June 2026.
Todd Toler and Angela Cochran, What publishing leaders say about AI when they’re not on panels, The Scholarly Kitchen, 11 March 2026.
Authors Alliance, Meta wins on fair use (for now), but court leaves door open for market dilution, 26 June 2025.
IPWatchdog, Third Circuit affirms revised fair use ruling against ROSS AI legal research platform, 30 September 2026.
Copyright Licensing Agency, Development of the CLA generative AI licence.
Training magazine, 2025 Training Industry Report, November 2025.
Pearson plc, 2025 Preliminary Results, 27 February 2026.
Medium, https://starweavergroup.medium.com/what-should-publishers-check-before-signing-ai-licensing-deals-a6b067bfe2ba,19 November 2026.
.png)
How publishers protect author consent, credit and IP when AI helps turn books into courses, plus what provenance tools like C2PA can and can’t prove.
.png)
Wiley, Taylor & Francis and HarperCollins figures compared with course economics, plus a 5-question framework for deciding what to do with your catalogue.