Market dilution: the theory Kadrey left open
Meta won summary judgment on an expressly thin record while the court signalled that market-dilution evidence could change the result. This is unsettled.
Meta won. That is the headline, and it is accurate, and it is close to the least informative thing you can say about Kadrey v. Meta.
Summary judgment went to Meta on the training claims on 25 June 2025. The court granted it expressly on a thin record, and in doing so pointed at a theory the plaintiffs had not properly evidenced: that a model trained on someone's work can harm the market for that work by flooding it with substitutes, even where no individual output copies anything.
That theory did not decide the case. It survived it. Fourteen months on, the interlocutory appeal has been denied, the torrenting and distribution phase is unresolved with a hearing set for 25 February 2027, and the dilution argument is still sitting there waiting for a plaintiff who builds the record for it.
Informational, not legal advice.
What was actually decided
Two rulings landed in the Northern District of California two days apart, and they get merged constantly.
On 23 June 2025, Bartz v. Anthropic held that training on lawfully acquired books was fair use and that retaining a pirated library was not. On 25 June 2025, Kadrey v. Meta granted summary judgment to Meta on the training claims.
The Kadrey opinion is not a broad endorsement. The court decided the question in front of it on the evidence in front of it, and was unusually direct that the evidence was insufficient to support the plaintiffs' market-harm case. That is a ruling about a record, not a rule about training. The docket is public if you want to follow the remaining phase.
Fair use turns on four factors, and the fourth is the effect of the use on the potential market for or value of the work. That factor is where dilution lives, and it is the factor plaintiffs in this field have consistently had the hardest time evidencing.
What market dilution actually means
The conventional market-harm argument is substitution: your copy displaces a sale of my work. That is what sank the defendant in Thomson Reuters v. Ross, decided in the District of Delaware on 11 February 2025, where a non-generative legal research tool was found to be a direct market substitute for the thing it was built from. Clean facts, clean theory.
Generative training does not fit that shape well. Nobody reads a language model instead of buying a specific novel. So the substitution argument tends to collapse, and the defence points at that collapse as proof of no harm.
Dilution is the alternative. Stated plainly: if a model is trained on ten thousand novels and can then generate an unlimited supply of competent novels in that general territory, the market for novels can be materially damaged even though no single output infringes any single input. The harm is to the market, not to a work. The mechanism is supply, not copying.
Three things make it awkward, and all three are why it remains untested:
- It is not the harm copyright usually recognises. Competition from new works is ordinarily lawful and desirable. Copyright does not protect against being out-competed; it protects against copying. Dilution asks the fourth factor to do something adjacent to what it has traditionally done.
- It cuts across the transformativeness finding. A use can be highly transformative under the first factor and still, on this theory, be devastating under the fourth. The Copyright Office's Part 3 report of 9 May 2025 rejected the claim that training is inherently transformative, though that report remains a pre-publication draft with no final version as of August 2026 and carries no force of law.
- It requires economics, not documents. Which is the actual obstacle.
Why it is hard to build
To win on dilution a plaintiff has to show, with evidence, that a specific model's outputs displaced demand in a market their works occupy. That means expert economic testimony, market definition, and a causal story that survives cross-examination.
The counter-arguments are obvious and strong. Output volume in a category grew for many reasons at once. The category was already changing. Buyers who take a generated substitute were never buying the original. Attribution to one defendant's model, rather than to the whole field, is genuinely difficult when a dozen models shipped in the same window.
None of that is unanswerable. It is expensive. Building a dilution record is a discovery-heavy, expert-heavy project, and the plaintiffs in Kadrey were litigating a case that had not been framed that way from the start. The court's signal amounts to: this was not the case for it, and someone else's might be.
Where the record might get built first
Three live matters look better suited to it than a books case did.
| Matter | Why the theory fits | Timing |
|---|---|---|
| Disney, Universal & Warner Bros. v. Midjourney (C.D. Cal.) | Consolidated 4 Nov 2025. Studios have the resources and the licensing-market evidence, and the case already targets outputs | Expert disclosures Oct 2026 |
| NYT v. OpenAI/Microsoft (S.D.N.Y.) | A defined market with measurable substitution, and a plaintiff with its own economics | Discovery; sanctions motion filed Jul 2026; no trial date |
| UMG v. Suno | Music has the clearest observable dilution mechanics of any category | Ongoing; Suno defending on fair use |
Music is worth dwelling on, because it is the one category where the market effect is already visible without an economist. Deezer tags AI-generated tracks and excludes them from editorial and algorithmic playlists, an economic penalty applied to content that is entirely lawful. Spotify took the opposite route, adopting DDEX AI-disclosure fields in credits while stating that disclosure does not affect royalties or recommendations. Two platforms, two theories of what synthetic supply does to a catalogue, and real data accumulating on both. What streaming AI-music labels actually change covers the operator-facing side of that split.
Whether any of it ends up in a fourth-factor expert report is a different question. But it is the first time the raw material for one has existed.
What this means if you buy model capacity
Not a reason to change what you ship. A reason to change what you assume.
Stop treating Kadrey as a green light. It is regularly cited in vendor materials as though training had been blessed. It was a summary judgment on an inadequate record with an explicit signal that a better record could produce a different outcome, and the same court's neighbouring case drew liability for how material was obtained. Neither is appellate authority. There is still no US appellate ruling on AI-training fair use, though the Third Circuit heard argument in Thomson Reuters v. Ross on 11 June 2026 and a decision is pending.
Diligence the sourcing, not just the training claim. The money that has actually moved in this field attached to acquisition. A provider that can describe how it obtained material is answering the question that has produced consequences. Licensed training data as a buying criterion is the version you can hand to procurement.
Keep the per-asset provenance record. The realistic downside scenario for a studio is not liability, it is disruption: a model restricted, withdrawn or re-licensed mid-campaign. Teams that know which shots came from which model version reshoot a list. Teams that do not reshoot everything.
Do not build a pipeline that only works on one model. A broad catalogue and a house standard for quality checks makes substitution a scheduling problem rather than a crisis. Quality control for AI-generated marketing assets is where that standard should live, and open-weight video on your own terms is the fallback worth having tested before you need it.
Watch the output side more than the input side. Nothing in the dilution debate touches your exposure for generating something that depicts a protected character, a real person or a brand asset. That risk was always yours and is unaffected by how any of these cases resolve. Legal and licensing for AI content in business covers where that line currently sits.
FAQ
Did the court in Kadrey endorse the dilution theory?
It did not adopt it as a holding. It granted summary judgment to Meta because the plaintiffs had not evidenced their market-harm case, while indicating that a properly evidenced dilution argument could change the analysis. That is a signal, not a ruling, and it binds nobody.
Is Kadrey over?
No. The training claims were resolved by that summary judgment and the interlocutory appeal was denied in July 2026, but the torrenting and distribution phase remains unresolved with a hearing set for 25 February 2027. The acquisition question, which is the one that produced the largest payment in the neighbouring Anthropic matter, has not been decided here.
If dilution wins somewhere, what happens to models already trained?
Unknowable, and anyone who tells you otherwise is guessing. The plausible outcomes range from damages against providers to licensing regimes negotiated in the shadow of a ruling, and none of them retroactively invalidates work you have already produced and delivered. The practical exposure for a studio is availability, not liability, which is why the provenance record and model breadth are the two things worth doing now.
Does this apply outside the United States?
Fair use is a US doctrine, and dilution as described here is an argument within it. Other jurisdictions have their own frameworks and their own gaps: UK training legality was never decided in the Getty proceedings, which ended on 4 November 2025 with the training and output claims dropped mid-trial for want of territorial evidence. The economic argument travels; the doctrine does not.