AI licensing vs online courses: which earns publishers more?

In 2024, Taylor & Francis brought in more than $75 million from AI data-access deals. Its parent company, Informa, classes that money as non-recurring. In 2025 the figure was lower. In the first half of 2026 Taylor & Francis signed no new deals of that kind, and Informa now sets its growth targets with those contracts excluded (Informa FY2025 results; Informa H1 2026 results).

That is the clearest public data on a question many rights, licensing and digital-learning teams are now arguing about. You can license your catalogue to AI companies, or you can turn the teachable parts of it into online courses. Which one earns more?

Before going further, a disclosure: Starweaver builds courses from publishers' catalogues, so we are not neutral. We have therefore stuck to published figures. We also say plainly where licensing is the better choice, and where nobody has the data to settle the question.

The short answer

Neither model wins every time. Which one earns more depends on what your catalogue contains and who holds the rights to it.

• AI licensing pays quickly and costs almost nothing to produce. But the money tends to arrive in lumps, and training deals in particular are often one-off. The disclosed deals go mainly to large STM, research and professional catalogues.

• Online courses cost money upfront and need a route to learners. In return they can earn repeatedly from the same IP, and you keep the relationship with the learner. They suit content that teaches a skill or a method, not content that mainly informs.

The two are not mutually exclusive. Many publishers will end up doing both: licensing reference and archive material, and building courses from the titles that teach.

What AI licensing actually pays publishers

Very few publishers disclose what they earn from AI licensing. Here is what is on the public record as of November 2026.

Wiley reported AI licensing revenue of $23 million in fiscal 2024 and $40 million in fiscal 2025. In fiscal 2026 it reported $49 million of AI revenue, up 23%, with lifetime AI revenue above $110 million (Wiley FY2025 results; Wiley FY2026 results). That is real money, and it is still about 3% of Wiley's $1.68 billion in revenue. The detail matters too. In fiscal 2026 Wiley's Learning segment fell 7%, partly because of "lower AI licensing revenue in Academic and Professional", while AI licensing on the research side grew. Even at Wiley's scale, licensing income from books moved down year on year.

Taylor & Francis fits the pattern in the opening: more than $75 million in 2024, less in 2025 and no new non-recurring contracts in the first half of 2026. Informa says the content still "is continuing to attract licensing interest". What it has stopped doing is planning on that money.

HarperCollins gives the clearest per-title price in trade publishing. Its 2024 deal paid $5,000 per title for selected nonfiction backlist books. That was split 50/50, so $2,500 to the author and $2,500 to the publisher, for a three-year term. Authors had to opt in (Quill & Quire; Josh Bernoff, a participating author). According to that author's account, the AI company can keep using models it has already trained after the term ends.

Cambridge University Press asks authors to opt in and pays them 20% of net AI licensing revenue. Its training licences typically run two to five years (Authors Alliance interview with CUP).

You may also see a figure of about $3,000 per book quoted. That comes from the $1.5 billion Bartz v. Anthropic settlement, which received final approval in July 2026 (Authors Guild). It is a settlement payment for pirated copies, not a market licence price, and should not go into a revenue model as one.

Why the licence type changes the economics

There are two kinds of AI licence, and they earn money in different ways.

A training licence lets a developer put your content into a model's weights. You are paid once, or once per term. After that the model has learned from the content. Developers renew only when they want new data for a new model.

A retrieval or RAG licence lets an AI product look up your content when someone asks a question, and can be charged per use. In a March 2026 piece based on 14 off-the-record conversations with publishing leaders, The Scholarly Kitchen described this split as "weights" versus "context". It found demand very uneven by discipline, strong in areas like computer science and clinical medicine, and reported that one publisher changed its pricing model four times in three months (The Scholarly Kitchen).

Retrieval licensing is where licensing could become recurring income. No standard terms or public rate cards exist yet. Ithaka S+R, which tracks these deals, notes that a standard model has "yet to emerge" (Ithaka S+R).

What online courses pay, and what they cost

The course side has a data problem of its own. Large publishers do not report course revenue as a separate line, so there is no public figure for "course revenue per title" to put next to Wiley's AI line. What is public shows where the demand sits and how the money is divided.

Employers are buying more outside content. US companies spent $102.8 billion on training in 2025, up 4.9%. Spending on outside products and services rose 29% to $16 billion (Training magazine 2025 Training Industry Report). Publishers' courses compete for that $16 billion.

Publishers' enterprise learning units are growing faster than the rest of the business. Pearson's Enterprise Learning & Skills division grew adjusted operating profit 40% on an underlying basis in 2025, from £20 million to £29 million. The group as a whole grew 6% (Pearson 2025 results). The starting base was small, but the direction is clear.

Third-party platforms keep most of the money, and pay out less than they used to. Udemy pays instructors 37% of marketplace sales. Their share of the subscription revenue pool was cut from 25% to 15% between 2024 and January 2026 (Udemy). Coursera has paid creators based on learner engagement since 2025, and in 2026 it began charging creators a platform fee on eligible sales (Coursera 10-K). Coursera and Udemy completed their merger on 11 May 2026 (Coursera), so there is now one fewer large marketplace to negotiate with. A publisher that relies entirely on someone else's platform is renting the shelf it sells from.

Courses cost real money to produce. No recent industry-wide benchmark exists. One 2024 practitioner estimate puts custom e-learning at $12,000 to $50,000 per finished hour, depending on how interactive it is. The same estimate says existing source material cuts design time by roughly half (LearningSim, 2024). Treat those numbers as a range, not a quote. The point holds anyway: a licence deal has almost no production cost, and a course does.

AI licensing vs online courses: side by side

How to tell which earns more for your catalogue

It helps to work out what you actually give up by choosing one over the other.

Take the HarperCollins-style price. Licensing a title for training pays the publisher about $2,500 over three years, if the author opts in. To beat that, a course built from the same title needs to earn its production cost plus $2,500 over the same three years. If a partner pays for production in return for a revenue share, the course only needs to earn more than $2,500 for you.

Even that overstates what you give up. A training licence and a course aren't automatically exclusive, so you could license the title and still build the course. Check whether your licence terms limit derivative products, and decide whether you are comfortable with an AI model that has learned from the book competing for the same learners.

Five questions decide which way your catalogue leans.

1. Does the title teach a skill or a method, or does it mainly inform? Professional and technical books with procedures, frameworks and worked examples turn into courses naturally. Reference works, research monographs and narrative nonfiction usually don't, and their value to AI developers is as data.

2. Who holds the AI rights? In trade publishing they often sit with authors, which is why HarperCollins and CUP use opt-in. Opt-in rates cap how much of a catalogue you can actually license. Course rights are usually easier to clear.

3. How large and specialised is the corpus? AI demand is concentrated in big STM and professional collections. A mid-sized list in a field with little demand may get no enquiries at all.

4. Can you reach learners? A course earns more if you can sell it to employers through an enterprise arm, your own channel or a mix of platforms, rather than only through one marketplace's subscription pool.

5. How long will the content stay current, and when do rights revert? A course needs refreshing, and a title whose rights revert soon gives you less time to earn back the cost of building it.

When AI licensing is the better choice

Licensing is the right call when the content is reference or archival with little teaching value, when the corpus is large and specialised enough to attract retrieval deals, when the rights are cleared and the authors are on board, and when you need cash now without spending on production. It is also sensible for titles near the end of their rights term, where there isn't enough time left for a course to pay back.

When courses are the better choice

Courses tend to win when the title teaches something people are paid to do: project management, data analysis, compliance, clinical procedures, finance. They also win when the publisher has, or can reach, enterprise buyers. This is the market where outside training spend grew 29% in a year, and where Pearson's enterprise unit grew profit fastest. A course also keeps the brand in front of the learner, whereas a licence passes the content to an AI product that does not name the publisher. And because each title has its own course, you can measure what each one earns, which an aggregate AI line on an annual report does not tell you.

The hybrid most publishers will end up with

Most catalogues contain both kinds of content. A practical approach is to divide the list:

• License the archive and reference material, and push for retrieval terms over one-off training fees where demand exists.

• Build courses from the teachable frontlist and the backlist titles that still have strong demand. (For more on turning authoritative content into learning products, see Artifactization.)

• Pilot before committing. Pick five to ten titles, build and distribute them, and compare what they earn over 12 to 24 months with the licence offers you have on the table.

Where Starweaver fits

Starweaver works on the course side of this decision. Our Lift, Polish & Shift programme takes the teaching content out of a publisher's books and turns it into interactive courses. A standard business or technical title takes five to seven weeks per production team. There are three commercial models:

• Fixed fee: you own the course IP outright.

• Advance plus royalties.

• Royalties only: Starweaver pays for production and shares revenue from the first dollar.

The royalties-only option speaks directly to the upfront-cost question above. We have rebuilt 10 Wiley For Dummies titles as structured courses. We currently produce more than 60 courses a month (Starweaver). Our Coursera partner page shows an average rating of 4.7 (Coursera).

If you are weighing a licence offer against building courses, start by finding out which of your titles would make good courses. Book a catalog audit and we'll go through your list with you.

Frequently asked questions

How much do AI companies pay publishers per book?

The clearest public figure is HarperCollins' 2024 deal: $5,000 per nonfiction title for a three-year training licence, split equally between author and publisher. The roughly $3,000-per-book figure from Bartz v. Anthropic is a settlement payment, not a licence price. Most publishers do not disclose their terms.

Is AI licensing revenue recurring?

Training licences mostly are not. Informa classes Taylor & Francis's AI data deals as non-recurring, and Wiley's book-related AI licensing revenue fell in fiscal 2026. Retrieval (RAG) licences can be charged per use and could recur, but they have no standard pricing yet.

What is the difference between a training licence and a RAG licence?

A training licence lets a developer build your content into a model, usually for a fee per term. A RAG licence lets an AI product look up your content when it answers a question, which can be charged per use and continue for as long as the product uses your content.

Do publishers need author permission to license books for AI?

It depends on the contract and the publisher's policy. HarperCollins and Cambridge University Press ask authors to opt in. Wiley has said its authors cannot opt out and are paid according to their existing contract terms. In trade publishing, AI rights often sit with the author.

Does licensing a book to AI stop you from turning it into a course?

Not automatically. The two use different rights and serve different buyers, so check the licence terms for any limits on derivative products. The commercial risk is that an AI tool trained on the book ends up answering the same questions your course would teach.

How much does it cost to turn a book into an online course?

There is no current industry benchmark. One 2024 practitioner estimate puts custom e-learning at $12,000 to $50,000 per finished hour, with existing source material cutting design time by about half. Some production partners, Starweaver included, offer royalty-only models that remove the upfront cost.

Which publishers are best placed to earn from courses?

Publishers whose content teaches a skill: professional, technical, business and education publishers, and the enterprise-learning divisions of larger groups, such as Pearson's Enterprise Learning & Skills. Narrative trade lists and highly specialised research monographs are usually better suited to licensing.

References

1. Informa plc, 2025 Full Year Results statement, 12 March 2026.

2. Informa plc, 2026 Half Year Results, 30 July 2026.

3. John Wiley & Sons, Fiscal 2025 results, 17 June 2025; Fiscal 2026 results, 16 June 2026.

4. Books+Publishing, "Wiley expects to make US$44 million from AI partnership; authors unable to opt out", September 2024.

5. Quill & Quire, "HarperCollins signs AI training deal", November 2024; Josh Bernoff, "Why I took the $2,500 check to license my book for AI", 15 April 2025.

6. Authors Alliance, "AI licensing: an interview with Ben Denne of Cambridge University Press", 17 March 2025.

7. Authors Guild, "Court grants final approval of Anthropic copyright settlement", 21 July 2026.

8. Todd Toler and Angela Cochran, "What publishing leaders say about AI when they're not on panels", The Scholarly Kitchen, 11 March 2026.

9. Ithaka S+R, Generative AI Licensing Agreement Tracker.

10. Training magazine, 2025 Training Industry Report, November 2025.

11. Pearson plc, 2025 Preliminary Results, 27 February 2026.

12. Udemy, instructor subscription terms update; Coursera, Form 10-K for fiscal 2025; Coursera, completion of combination with Udemy, 11 May 2026.

13. LearningSim, "2024 Cost Estimates for Learning Design".

14. Starweaver, Lift, Polish & Shift; AI augmentation; Coursera partner page.