Publishers Are Licensing Your Backlist to AI Companies: What Say Do Authors Actually Have?

The answer depends almost entirely on one paragraph in a contract you may have signed a decade before generative AI existed. Not on industry norms, not on what other publishers are doing, and not on whether the arrangement seems fair. On the grant clause.

That is a less satisfying answer than authors want, but it is the one that determines outcomes, and it points at where the useful work is.

What the Market Has Established So Far

The reference transaction is HarperCollins. In November 2024 it became the first of the Big Five to license nonfiction backlist titles for AI training, in a deal reported to be with Microsoft. The terms set the first public price for book training data:

  • $5,000 per title over three years
  • Split 50-50 between publisher and author
  • Paid separately, not recouped against the advance
  • Opt-in — authors were asked title by title
  • Output guardrails: no more than 200 consecutive words and/or 5 percent of a book’s text
  • The licensee is required to avoid scraping piracy sites

Academic publishing moved earlier and larger. Wiley announced roughly $44 million across two AI licensing deals in June 2024, covering academic titles and the Dummies line — a catalogue heavily weighted toward work-for-hire arrangements where the publisher holds the copyright outright and no author permission is needed.

That contrast is the shape of the whole issue. Where the publisher owns the copyright, there is no author consent question. Where the author retains it, everything depends on what was licensed.

The Authors Guild’s Position, and Why It Is Not Merely Advocacy

The Authors Guild’s analysis of the HarperCollins deal makes a claim worth taking seriously as a contract-interpretation argument rather than a bargaining position.

Its core contention: authors own AI training rights, because standard publishing agreements reserve all non-book rights to the author. On that reading, a publisher needs separate authorization to license a title for training — which is precisely what HarperCollins sought, and the Guild credits it for doing so properly.

The Guild’s objection is to the split, not the structure: “AI training rights are not book or excerpt rights; it is the authors’ expression that produces value in AI licensing.” Its position is that authors should receive the large majority, with the publisher compensated for the actual labor involved — file preparation, metadata — at something closer to an agent’s fee than a rights-holder’s share.

The Guild’s model contract language addresses this directly through a no-AI-training clause: the publisher may not use or sublicense the work to train generative AI without the author’s express written consent. For any new agreement, that clause is the single highest-value AI-related ask.

Where Your Contract Actually Decides It

Four provisions do the work. Reading them in this order will tell you most of what you need to know.

The grant clause. Did you grant specific, enumerated rights — hardcover, paperback, ebook, audio, translation — or did you grant “all rights”? An enumerated grant is a strong position: rights not listed are reserved to you, and AI training appears in no pre-2023 enumeration. An all-rights grant is a weak one.

“Now known or hereafter devised.” This is the pivotal language in most disputes. Contracts routinely grant rights in “any and all media now known or hereafter devised,” and publishers can argue it reaches AI training as a later-invented use. The counterargument is that such clauses are generally read to cover new media for delivering the work to readers — a new format, a new device — not a fundamentally different use in which the work is consumed by a machine to build a product that competes with it. The same reasoning appeared in earlier disputes over whether print-era grants covered ebooks. No settled rule governs this yet, which means the answer today is negotiated rather than looked up.

The subsidiary rights schedule and its splits. If AI licensing gets characterized as a subsidiary right, the percentage in your existing sub-rights table may control the money — and those splits were set for permissions, serial, and anthology uses, with no thought of training corpora. Check whether your agreement has a catch-all sub-right category, and what it pays.

Reversion and out-of-print provisions. If rights in a backlist title have reverted, the publisher has nothing to license. Authors with genuinely dormant backlist should confirm reversion status before anything else, because reversion converts the entire question from a negotiation into a simple no. Older contracts often define out-of-print by criteria that print-on-demand and ebook availability quietly defeated, so this is worth reading closely rather than assuming.

The Structural Problem With Older Contracts

Agreements signed before roughly 2023 contain no AI clause at all. That silence is not neutral — it is the whole dispute. Both readings are available: the publisher points to broad future-media language; the author points to the reserved-rights principle and the absence of any specific grant.

Which reading prevails will depend on the specific wording, the governing law, and eventually on litigation that has not yet produced clear answers. In the meantime, silence in an old contract is being resolved in practice by whoever has more bargaining power, which is generally not the author.

This is separate from — and sits underneath — the unresolved question of whether AI-generated output attracts copyright at all, and separate again from the infringement litigation now running against developers who acquired books from piracy sites rather than licensing them. A licensed corpus is exactly what those cases suggest developers should have built. That is why the licensing market exists, and why it will keep growing.

Practical Position for an Author Approached Today

Read the grant clause before responding. Do not negotiate price before establishing whether the publisher needs your permission at all. Those are different conversations and the second one is worth more.

Treat consent as bargaining power over terms, not only a yes-or-no. The HarperCollins guardrails — output caps, a no-piracy-scraping obligation, payment outside the advance — are now a documented market baseline. An offer without them is below market.

Ask what the license permits, and for how long. Three years is not perpetual, and a fixed term with defined scope is materially different from a broad, evergreen grant at the same price.

Get the no-AI-training clause into every new agreement. For contracts not yet signed, this is the cheapest protection available, and it converts every future approach into a negotiation you control. The scope of what a license transfers is decided when the contract is drafted, not when the opportunity arrives.

The industry’s direction is not seriously in doubt — the licensing market is being built. The open question is whether authors participate in it as rights holders or as recipients of a share someone else determined. That gets settled clause by clause, in contracts, before the money is on the table.

This article is editorial and informational, not legal advice. Consult a licensed attorney about your specific situation.