When Does an AI Licensing Market Count?

A market can tell us that something has value. It cannot, by itself, tell us who is legally entitled to collect that value.

That distinction is becoming increasingly important in the copyright fights surrounding artificial intelligence.

Publishers, authors, technology companies and licensing intermediaries are beginning to construct markets for access to copyrighted material used in AI development. Once those markets exist, their existence can itself become evidence in litigation: if licenses are available, unauthorized use can be described as depriving copyright owners of licensing revenue.

But that reasoning contains a potential circle.

If the assertion of a licensing right helps create a market, and the existence of that market is then used to prove that the right should have existed, the market begins supplying evidence for its own legal foundation.

The question is therefore not simply whether an AI-training licensing market exists.

It is whether this is a market the law counts.

This question became concrete after Ross. The earlier essay found that ruling narrower than the reaction to it. But two parts of that episode matter here. The court found harm to a “rapidly developing” market for licensing headnotes as AI-training data, though for a non-generative competitor. And Maria Pallante, head of the Association of American Publishers, framed the ruling in terms of “a robust licensing market.”

One is a holding and the other is an advocate’s framing. Both treat a licensing market as something that counts. Whether it should is the question.

The Entitlement Question Comes First

Two questions that are often blended together need to remain separate.

The first is an entitlement question: does copyright law give the owner the right to control or demand payment for this particular use?

The second is an evidence question: if such a market exists, what does its existence tell us about economic harm?

They are related, but they are not interchangeable.

Many activities create value for others without giving every contributing party a legal right to charge for that contribution.

Copyright therefore cannot resolve the training question merely by observing that AI companies are willing to pay some copyright holders.

The law has to decide what kind of use it is evaluating before the price of permission can settle anything.

That matters particularly for generative AI, because public debate often moves among several distinct acts under the shorthand “AI training.”

Acquiring a book unlawfully is one question. Making computational copies from a lawfully acquired book is another. Using those copies to extract statistical relationships is another. Producing passages that reproduce protected expression is another. And producing a competing work that substitutes economically for the original may be another still.

A licensing regime that treats all five as the same act risks solving the wrong problem.

The Circular Market

The difficult case is a market that emerges while the legal entitlement itself remains contested.

Courts have long placed limits on which licensing markets count. In Campbell, the Supreme Court rejected a protectable derivative market for criticism. Texaco, drawing on that line of reasoning, limited potential licensing revenues to markets that are “traditional, reasonable, or likely to be developed.”

The 2025 AI-training cases brought that older problem into a new technological setting. In Kadrey, Judge Chhabria, ruling for Meta on the record before him, rejected “the theoretical market for licensing the use” as circular while leaving market dilution open as a different route to harm. Bartz, separately, held training on legally purchased books to be fair use.

Suppose publishers begin offering licenses specifically for AI training. Some developers accept them. Prices emerge. Contracts accumulate.

It may then be tempting to say that an unlicensed developer has damaged an established market by failing to purchase the same license.

Perhaps. But the inference cannot be automatic.

Otherwise copyright owners could create a market for permission to perform a disputed use and then point to the resulting market as proof that the disputed use requires permission.

The problem is not that new licensing markets are illegitimate. Copyright markets constantly evolve with technology. The problem is circularity.

The existence of transactions can demonstrate demand. It does not independently resolve whether the seller possesses the legal entitlement being sold.

That leaves courts with a harder factual question: when is an emerging licensing market evidence of genuine displacement? And when is it largely the product of the legal claim whose validity is still being decided?

There is unlikely to be a single bright line.

Timing matters. A market that developed independently before litigation carries different evidentiary weight from one assembled primarily in anticipation of it.

Volume matters. A handful of strategic agreements says something different from routine transactions across an industry.

Pricing matters. Negotiated prices reached by parties with meaningful alternatives tell us more than prices formed under substantial uncertainty about legal exposure.

And fit matters most of all. A market for reproducing or redistributing expressive works is not necessarily evidence of a market for extracting computational information from lawfully acquired copies.

The existence of a licensing market is therefore not enough. The relevant question is whether it represents the use actually under review.

Training Is Not Output

This distinction becomes clearer if training and outputs are evaluated separately.

One possible regime would treat computational analysis of lawfully acquired works as presumptively permissible, while policing acquisition and outputs separately. Its attraction is administrability. No one has to obtain individualized permission across millions of works.

Its weakness is that “lawfully acquired” carries the whole rule. If training needs no permission, control moves upstream to access: subscriptions, paywalls, terms of service. A rule meant to prevent a veto over analysis could shift substantial control to whoever governs access. Harm may also move downstream, to markets where outputs compete with human work. More on that below.

Unlawful acquisition remains unlawful.

Nor would the rule immunize outputs. If a model reproduces protected expression, substitutes for the original work, or competes in a market in a way copyright law recognizes as actionable, those consequences can be evaluated directly.

This avoids making the training copy carry every economic concern associated with the finished system. It does not make the line between training and output clean. A model that memorizes and regurgitates closes the gap. The separation is a way to ask questions in order, not a claim that they never touch.

It also separates two ideas that are too easily treated as identical: the right to refuse a use, and the right to receive compensation from an economic system built partly from copyrighted works.

A society could conclude that transformative computation over lawfully acquired material should not require millions of individualized permissions while also concluding that creators deserve some share of the economic value produced by large-scale training.

That would move the argument away from a binary choice between unrestricted training and universal licensing vetoes.

The harder question would then become institutional rather than rhetorical: if compensation is warranted, how should it be distributed without turning compensation itself into a new mechanism of exclusion?

Compensation Without Veto

Separating the right to refuse from the right to be paid clears a path. It doesn’t show what the path is built from.

None of the mechanisms below has been built for AI training. What follows is evidence from analogues: music licensing, radio, Canadian education. Each analogue shows where compensation can stop being compensation and become a gate.

Licensing Without a Collective

Direct licensing leaves the veto in place. An owner sets a price, and the developer pays or doesn’t train. It finds value honestly, but it can’t establish that the owner was entitled to the price.

It also scales badly, because one holdout can block use of that work, and millions of individualized negotiations can make comprehensive licensing impractical.

Opt-out and extended collective licensing keep a veto for everyone who exercises it, which brings the holdout problem back for exactly those owners.

The Collective

The closest analogue is the Mechanical Licensing Collective. Since 2021, digital services have licensed songs under a statutory blanket license. A nonprofit collects the royalties, matches usage to owners, and distributes.

In June, the Copyright Office reported roughly $3.9 billion distributed and about 92 percent of royalty value matched.

That is value, not works.

When matching fails, the statute redistributes unmatched royalties, after a holding period, to owners in proportion to market share. The Office’s data suggest that low-earning works are matched less often than top earners. That is an inference about where the money lands, not a demonstrated outcome.

Ten of the board’s fourteen seats go to publisher representatives. A system that removes the veto over use can concentrate substantial influence over distribution in the largest rights holders.

Canada shows the limit from another direction. Access Copyright reports that its education revenue fell by more than ninety percent after 2012. One reading is that a collective collects only as far as the underlying right reaches.

The Levy

A statutory levy is the strongest version of “without a veto,” because the law removes the refusal.

The Copyright Office treats compulsory licensing as exceptional for that reason. It calls it a derogation of the author’s right to control the use, and warns that it is hard to undo.

A flat levy would price out small and open-source developers. A tiered levy, scaled to revenue or compute, would reduce that barrier.

But tiers invite gaming: companies restructure to sit under thresholds, and the cheapest compute gets routed around the rule.

The Division Problem

Music distributes by usage because a stream is an event with a record.

Training has no generally accepted usage record equivalent to a stream.

Where usage isn’t counted, performing-rights organizations distribute on surveys and proxies, which makes them a useful analogue for the proxy problem AI training would face.

The one statute that hard-codes a creator share is the digital-radio split of 50 percent to the copyright owner, 45 percent to the featured artist, and 5 percent to non-featured performers.

It works only because streams produce usage reports. Training has neither.

So any formula runs on a proxy: catalog size, past earnings, or substitution risk.

Catalog size pays incumbents.

Substitution risk pays in proportion to how well the technology displaces you, and the harm it targets — market dilution — has not been established by any court as of early October.

Whoever picks the proxy decides who counts.

That is the entitlement question again.

The Registry

Transparency is usually called a precondition for fair division. You can’t distribute without knowing what was trained on.

But a disclosure record is also a list that can be checked, enforced, and eventually asserted against.

One design answer is a use limitation. If disclosure exists to divide money, the law can say so: records collected for distribution may not be used to prove infringement.

That doesn’t remove the registry. It narrows what the registry is allowed to become.

It could fail.

Discovery rules, parallel litigation, and plaintiffs’ ingenuity all pull against it. It also has a cost: shielding the registry from evidence decides the entitlement question by procedure.

But it is a mechanism, and a mechanism can be argued with.

Where This Leaves Us

Collection and division are two steps, and each can reinstall the control it was meant to replace.

The question that opened this essay returns one level up.

A market can tell us what something is worth.

It can’t tell us who is owed it.

If compensation is warranted, can it be designed to remain a payment rather than harden into a claim that excludes the next entrant, blocks the next use, or gates the next generation’s access to the culture the technology drew on?

Sources

• Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153 (3d Cir. Sept. 29, 2026).
• Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569 (1994).
• American Geophysical Union v. Texaco Inc., 60 F.3d 913 (2d Cir. 1994).
• Kadrey v. Meta Platforms, Inc., 788 F. Supp. 3d 1026 (N.D. Cal. 2025).
• Bartz v. Anthropic PBC, 787 F. Supp. 3d 1007 (N.D. Cal. 2025).
• Association of American Publishers, statement by Maria Pallante (“a robust licensing market”), via Publishers Weekly (Oct. 2, 2026).
• U.S. Copyright Office, Copyright and Artificial Intelligence, Part 3: Generative AI Training (pre-publication version).
• U.S. Copyright Office, Periodic Review of the Designations of the Mechanical Licensing Collective and Digital Licensee Coordinator, 91 Fed. Reg. 33208 (June 3, 2026).
• Letter from the U.S. Copyright Office to Representatives Jordan, Issa, and Fitzgerald on performing rights organizations (Nov. 20, 2025).
• 17 U.S.C. §§ 114(g)(2), 115(d)(3)(J).
• Access Copyright’s reported education-revenue decline: Quill & Quire and The Globe and Mail (2023).

Disclosure

Anthropic, the company that makes Claude, which assisted in editing this essay, was the defendant in Bartz v. Anthropic.

Dale Joseph is the author of Thought Partners: Preserving Cognitive Sovereignty in the Age of AI and founder of the Emergence Institute. He worked for years as a consultant helping install hospital networks before turning to writing and systems thinking. He lives in Boynton Beach, Florida.