How do publishers protect author rights and proof of origin when AI helps create courses?

Consent, credit and a paper trail for every lesson

In 2025, BISG and BookNet Canada surveyed 559 people across the North American book industry. 86.4% named weak controls on copyrighted material as a concern about AI, and 73.9% named the lack of disclosure to consumers when AI is used. Only 2.8% were using AI in rights and licensing work (Publishing Perspectives; BookNet Canada).

Put those numbers together and you get the problem this article deals with. Publishers worry most about rights and disclosure, but the teams that handle rights are the least involved in AI work. That gap matters most when a book is turned into a course with AI involved: AI drafts scripts from the chapters, writes quiz questions, narrates lessons and sometimes translates them.

A disclosure first: Starweaver turns publishers’ books into courses, so this is our working world. We’ve kept to published law, contracts and standards, and we point out where the answer isn’t settled yet.

The short answer

Publishers protect author rights and proof of origin by treating them as two separate jobs that meet in one production workflow.

• Author rights is a contract and consent job. Check that the author agreement covers courses and AI-assisted derivatives. Get written consent where AI touches the author’s words, voice or likeness. Keep the author credited, and keep the manuscript out of other companies’ models.

• Proof of origin is a record-keeping and disclosure job. Record the source edition, every AI tool used and what it did, every human change and every sign-off. Use signing and fingerprinting tools on the media files, and disclose AI use where the law or the platform requires it.

No single technology proves where a course came from. A documented process does most of the work, and the technology only supports it.

Why courses raise rights questions that books didn’t

A course adapted from a book is a derivative work. The publisher needs the right to make it, usually under the subsidiary, electronic or derivative-rights clauses of the author contract. That part isn’t new.

AI adds steps, and each step touches a different right:

Part 1: Protecting author rights

Read the contract for AI before you read it for courses

Many existing contracts were signed before generative AI and say nothing about it. New ones are starting to. In April 2026 the Authors Guild published new model contract clauses that agents can bring to negotiations (Authors Guild). Under them:

• the publisher gets no AI rights by default, apart from a narrow carve-out for internal tools used for marketing and search;

• AI training, retrieval (RAG), AI summaries and “chat with a book” apps become separate subsidiary rights, each licensed on its own;

• the publisher needs the author’s approval before uploading the manuscript to a consumer chatbot, and must then stop third parties training on it;

• the publisher may not use AI to edit the manuscript substantively, though spelling and grammar tools are fine;

• AI narration and AI translation need the author’s prior written consent.

These are model clauses, not law, but they show what a well-advised author will ask for. Map your course workflow against them before production starts, not after an agent raises them.

Ask for consent, and make it real

Cambridge University Press set the precedent most often cited. It asked authors to opt in to AI licensing through a contract addendum, rather than treating silence as consent (Authors Alliance). Wiley took the opposite line on its AI licensing deals and told authors they could not opt out, and was criticised for it (The Bookseller).

Asking for permission doesn’t settle everything. In late 2024 VBK, the Dutch publisher owned by Simon & Schuster, planned an AI-assisted translation pilot for a few titles. It had asked the authors for permission and still faced objections from translators’ groups and PEN America (Publishers Weekly). What upset people was the use itself, not only how consent was handled.

For courses, the safest rule is simple. The more personal the AI use, the more explicit the consent should be. Summarising a chapter into a lesson sits at one end. Cloning the author’s voice sits at the other.

Treat voice and likeness as rights of their own

An AI avatar or a cloned voice of a real author is a different kind of right from the text. In the US, Tennessee’s ELVIS Act (2024) and California’s AB 2602 and AB 1836 (in force from January 2025) restrict unauthorised digital replicas. AB 2602 makes vague replica terms in contracts unenforceable (Davis+Gilbert). The federal NO FAKES Act was reintroduced in May 2026 and cleared the Senate Judiciary Committee in June, but it is not yet law (Holland & Knight).

In the EU, Article 50 of the AI Act has applied since 2 August 2026. It requires disclosure when AI-generated audio or video resembles a real person (CPD UK analysis). If you use a synthetic version of an author, get specific written consent covering what it may say, where it may appear and for how long. Then disclose it.

Keep the author’s name on the work, and the work intact

In the UK, authors have a moral right to be identified, though they must assert it, and a right to object to derogatory treatment of their work (CDPA s77; s80). A course that compresses a careful argument into a few bullet points can cross that line, especially if AI drafting has introduced errors. Show the author the outline and a sample lesson. Credit the source book and author on the course page and in the course metadata.

Know where the manuscript goes

This is the risk teams most often miss. Some AI tool and platform terms take broad rights in what you upload or what the tool produces. Coursera’s Enterprise Content Authoring Addendum, for its AI course-building tool, makes the organisation “solely responsible” for the accuracy of generated content. It also grants Coursera a perpetual, royalty-free, sublicensable licence to AI-conceived content, including to improve Coursera’s own products (Coursera). In 2024 Udemy gave instructors only a short window to opt out of generative-AI training on their content (404 Media).

Before a chapter goes into any tool, find out three things: whether the vendor trains on inputs, whether it claims rights in the outputs, and where the data is stored. Keep a list of approved tools. Your production partners should be able to answer the same three questions.

Part 2: Proving where the course came from

Why a paper trail has legal value

In January 2025 the US Copyright Office concluded that prompts alone do not make someone an author. Human selection, arrangement and modification of AI output can be protected, as can human-written material that remains visible in the result, such as the book’s own text (US Copyright Office). In March 2026 the Supreme Court declined to hear Thaler v. Perlmutter, leaving the human-authorship requirement in place (Mayer Brown via JD Supra).

For a course, this means the human work is what you can own: the instructional designer’s structure, the expert’s rewrites, the author’s approved text. If you can’t show that work, your claim to the course is weaker. The UK is still deciding. Its March 2026 copyright and AI report kept the “computer-generated works” provision under review, with a stated preference for removing it (UK government report).

What to record for every course

A practical production record has six parts:

1. Source: title, edition, ISBN, and the chapters and figures used.

2. Rights basis: the contract clause or consent addendum that permits the course and each AI use.

3. AI log: each tool and model, what it was asked to do, and which outputs were kept.

4. Human changes: who rewrote, restructured or fact-checked which parts.

5. Approvals: author or subject-expert sign-off, with dates.

6. Release record: final file versions, where they were distributed, and the disclosure text used.

None of this needs new software. It needs someone to own it, usually rights or editorial operations, and it needs to stay with the course files for as long as the course is sold.

What provenance technology can and can’t prove

Several tools get mentioned as “proof of origin”. Each proves something narrower than its name suggests.

• C2PA Content Credentials attach a signed record to images, video, audio and PDFs. They can prove who signed a file and that it hasn’t changed since. They cannot prove the content is accurate, that a human wrote it or who owns the copyright. They also depend on signing keys being secure: in 2025 Nikon suspended its C2PA feature after a flaw let altered images carry valid signatures (PetaPixel). We found no learning platform that reads or displays C2PA yet, so it is most useful on lesson videos, narration files and downloadable PDFs.

• Text watermarking, such as Google’s SynthID Text, is probabilistic. Google’s own documentation says it works less well on factual text and much less well after heavy rewriting or translation (Google). It can’t prove a human wrote something.

• ISCC content codes (ISO 24138) fingerprint a file from its content, so you can show that two files match or are near-copies. They don’t record ownership.

• The Authors Guild’s “Human Authored” certification covers books, not courses, and translations and audiobooks need separate certification (Authors Guild). It tells you about the source book, not about the course made from it.

Used together, these are useful supporting evidence. None replaces the production record. NIST’s overview of synthetic-content transparency covers provenance, watermarking, metadata and detection as separate techniques, each with its own limits (NIST AI 100-4).

Disclose AI use where it’s required, and decide where it’s wise

Platforms are setting their own rules. Udemy does not accept courses that are entirely AI-generated, and requires instructors to say in the course description whether they used AI (Udemy).

In the EU, Article 50 gives an exemption for AI-generated text that has had human review and has someone with editorial responsibility. But the exemption applies to text published to inform the public on matters of public interest. Whether commercial course content qualifies is unclear, so take legal advice rather than assume editorial review is enough. AI voices and avatars of real people need disclosure regardless.

Scholarly publishing is ahead here. COPE’s position is that AI tools cannot be authors and that their use must be disclosed (COPE). In 2025 STM published a classification of AI use in manuscript preparation that publishers can adapt into a consistent disclosure line for courses (NISO). Authors expect publishers to lead on this. Wiley’s 2025 survey of 2,430 researchers found 73% look to publishers for clear AI guidelines (Wiley).

A seven-step workflow for AI-assisted courses

1. Rights check. Confirm the contract covers courses and each planned AI use. Flag titles that need an addendum.

2. Consent. Send the author a short, specific request that lists each AI use. Voice, likeness and translation each get their own yes or no.

3. Approved tools. Use only tools whose terms don’t train on your inputs or claim your outputs. Record the version of each.

4. Production log. Keep the AI log and the human-change record as the work happens, not afterwards.

5. Review and sign-off. Get author or subject-expert approval of the outline, a sample lesson and the final course.

6. Package the evidence. Sign media files with Content Credentials where your tools support them. Fingerprint final files. Credit the author and source in the metadata.

7. Disclose and monitor. Add an AI-use statement that meets platform and legal requirements, and review it when the law or a platform’s rules change.

Mistakes we see most often

• Assuming “electronic rights” covers AI. Older clauses were written for ebooks. Courses, AI narration and translation may each need explicit permission.

• Uploading chapters to a consumer AI tool to “just draft something”. This is how a manuscript ends up under terms nobody on the rights team has read.

• Treating a watermark as proof. Watermarks and signatures support a record; they aren’t one.

• Writing the disclosure after launch. Agree the disclosure text with the author during consent, so nobody is surprised by it.

Where Starweaver fits

Our Lift, Polish & Shift programme turns publishers’ books into courses. Two parts of it bear on author rights:

• Two paths for authors. In the author-led path, the author teaches on camera. In the author-optional path, a Starweaver subject-matter expert teaches the material, faithful to the source and to your brand. The publisher decides which fits each title and each author’s wishes.

• Clear IP terms. Under a fixed-fee agreement, the publisher owns the course outright. Under the advance-plus-royalties and royalties-only models, the course is treated as a derivative work and its IP and revenue are shared as agreed.

Our production tooling, Boostr, uses review gates so that no content reaches the catalogue until every gate has been cleared.

Ask any course partner, including us, the same questions this article recommends: which AI tools touch your manuscript, whether those tools train on it, who reviews the output, and what record you get at the end. If you’re planning to turn part of your list into courses, book a catalog audit and we’ll go through rights, consent and production with you title by title.

Frequently asked questions

Do publishers need author permission to turn a book into a course using AI?

A course is a derivative work, so the publisher needs derivative or course rights under the author contract. Whether AI-assisted production needs separate consent depends on the contract. The Authors Guild’s 2026 model clauses give publishers no AI rights by default and require the author’s prior written consent for AI narration and translation.

Can a course made with AI help be copyrighted?

In the US, yes, for the human contribution. The Copyright Office protects human selection, arrangement and modification of AI output, and human-written material that remains in the result, but not material generated from prompts alone. Keeping a record of the human work strengthens the claim.

What is proof of origin for AI-assisted content?

Evidence of where content came from and how it was made: the source, the AI tools used, the human changes and the approvals. Tools such as C2PA Content Credentials and ISCC fingerprints can support that evidence, but none of them proves authorship on its own.

Does C2PA work for online courses?

Partly. C2PA can sign the media files a course is built from, such as video, audio, images and PDFs. We found no learning platform that reads or displays Content Credentials, and the standard has no format for a SCORM package or an LMS page as a whole.

Do courses have to disclose AI use?

Sometimes. Udemy requires instructors to disclose AI use in the course description. In the EU, Article 50 of the AI Act has required disclosure of AI-generated voices and avatars resembling real people since 2 August 2026. Whether its text exemption covers commercial courses is not settled.

Can a publisher use an AI clone of the author’s voice?

Only with the author’s specific written consent. In the US, laws such as Tennessee’s ELVIS Act and California’s AB 2602 restrict unauthorised digital replicas, and in the EU the use must be disclosed.

What should publishers ask an AI course production partner?

Which AI tools touch the manuscript, whether those tools train on your content or claim rights in the output, who reviews and approves each stage, and what production record you receive with the finished course.

This article is general information, not legal advice. Check specific contracts and jurisdictions with your counsel.

References

1. BookNet Canada and BISG, “Results from the AI use across the North American book industry survey”, 27 April 2026; Publishing Perspectives coverage.

2. Authors Guild, “Authors Guild releases new AI model clauses and issues updates”, 29 April 2026.

3. Authors Alliance, “AI licensing: an interview with Ben Denne of Cambridge University Press”, 17 March 2025.

4. Publishers Weekly, “Dutch publisher’s AI translation plan sparks industry concern”, November 2024; Davis+Gilbert, “California passes new legislation prohibiting unauthorized AI replicas”; Holland & Knight, “Senate Judiciary Committee advances legislation”, June 2026.

5. The Bookseller, “Wiley set to earn $44m from AI rights deals, confirms no opt-out for authors”, August 2024.

6. CPD UK, “The deadline that moved and the duty that did not: AI Act transparency obligations after the Digital Omnibus”, 2026.

7. UK Copyright, Designs and Patents Act 1988, s77 and s80.

8. Coursera, Enterprise Content Authoring Addendum, effective 19 March 2025.

9. 404 Media, “Massive e-learning platform Udemy gave teachers a gen AI opt-out window. It’s already over”, 2024.

10. US Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability, 29 January 2025.

11. Mayer Brown via JD Supra, “Supreme Court denies review in AI authorship case”, 10 March 2026.

12. UK Government, Report on Copyright and Artificial Intelligence, March 2026.

13. C2PA, Specifications; PetaPixel, “Nikon suspends C2PA functionality on the Z6 III”, 5 September 2025.

14. Google, SynthID documentation.

15. Authors Guild, “Human Authored” certification and FAQ.

16. Udemy, “Course Quality Checklist: Use of AI”.

17. COPE, “Authorship and AI tools”; NISO I/O, “STM releases recommendations for classification of AI use in manuscript preparation”, September 2025.

18. Wiley, “AI adoption jumps to 84% among researchers”, October 2025.

19. NIST, AI 100-4: Reducing Risks Posed by Synthetic Content, November 2024.

20. Starweaver, Lift, Polish & Shift; Boostr.

21. Medium, https://starweavergroup.medium.com/how-do-publishers-protect-author-rights-and-proof-of-origin-when-ai-helps-create-courses-1696b8cb2465

22. Substack, https://starweavergroupinc.substack.com/p/how-do-publishers-protect-author?r=9ec4vv&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true