The Ross appeal and AI training fair use
The Third Circuit heard Thomson Reuters v. Ross on 11 June 2026 and has not ruled. What the first US appellate word on AI training can and cannot settle.
The first American appellate court to take up whether training an AI system on someone else's copyrighted material is fair use has heard the argument and has not ruled. The Third Circuit heard Thomson Reuters v. Ross on 11 June 2026. As of today there is no decision.
That gap is worth understanding rather than waiting out, because a lot of writing about AI and copyright is currently borrowing certainty from a case that has not produced any. Below is what the appeal is actually about, what a ruling could decide, and the much longer list of things it will not touch.
What the court below held
In February 2025 the District of Delaware held that Ross's copying was not fair use. Two features of that record did the work.
The first is that Ross's product was not generative. It was a legal research tool. Nothing in the case turns on a diffusion model, a text-to-video model, or a chat assistant producing new material.
The second is that the product was a direct market substitute for the thing it had copied from. Ross was building a competitor to the research service whose material it used. That is about as unfavourable a fair use posture as a defendant can have, and it is not the posture most generative model providers are in.
Those two facts are the reason careful readers treated the district ruling as narrow when it landed, and they are the reason the appeal is more interesting for its reasoning than for its result.
The other district rulings, which point the other way
Three district-level decisions exist on training. None of them binds any other court.
| Case | Court and date | On training | Where it stands |
|---|---|---|---|
| Thomson Reuters v. Ross | D. Del., 11 Feb 2025 | Not fair use — non-generative tool, direct market substitute | Third Circuit argument 11 Jun 2026, decision pending |
| Bartz v. Anthropic | N.D. Cal., 23 Jun 2025 | Training on lawfully acquired books was fair use and "spectacularly" transformative; retaining a pirated library was not | $1.5 billion class settlement, final approval 20 Jul 2026 |
| Kadrey v. Meta | N.D. Cal., 25 Jun 2025 | Summary judgment for Meta on training, expressly on a thin record | Interlocutory appeal denied Jul 2026; torrenting and distribution phase unresolved, hearing set 25 Feb 2027 |
Two things in that table get misread constantly.
The Anthropic settlement is not a ruling. Settlements create no precedent. And the district decision that preceded it went for the fair use argument on training and against the defendant on how the library was obtained. The largest sum paid out so far in this whole area attached to acquisition, not to training.
Kadrey was decided on a thin record. The court said so, and signalled that a market-dilution theory could succeed if a plaintiff actually evidenced it. A win on an underbuilt record is not the same as a rule.
What a Third Circuit ruling could settle
A published opinion would be binding precedent in the Third Circuit and persuasive everywhere else. Realistically it could resolve:
- How a circuit weighs the fourth fair use factor when the copier's product competes with the market for the copied work. This is the axis the district court leaned on hardest.
- Whether copying for a machine-learning purpose counts as transformative where the end product is not itself expressive output. Intermediate-copying doctrine has decades of case law behind it in software; nobody has yet applied it to a training corpus at circuit level.
- Whether the reasoning below survives review at all. An affirmance on narrow grounds and a reversal would send very different signals, and so would an affirmance on broad grounds.
That is a real contribution to a question that currently has three district answers and no appellate one.
What it cannot settle
Longer list, and the more useful one.
- It will not decide generative training. Ross's tool was non-generative. A panel can dispose of this appeal without ever addressing a model that produces images, video, or prose.
- It will not travel cleanly to creative models. "Substitutes for the market of the thing it copied" is a tidy fit for a legal research competitor. It is a much harder argument against a general-purpose video model, which is why plaintiffs elsewhere are building dilution theories instead.
- It will not bind the Ninth Circuit, where most of the significant AI training docket sits. Bartz and Kadrey are both Northern District of California.
- It will not touch acquisition. The pirated-library problem that drove the Anthropic outcome is a different question on a different record.
- It will not say anything about outputs. Training and output are separate legal questions, and for anyone actually generating content the output question is the one that matters.
There is also a regulator in the mix pointing a fourth direction. The Copyright Office's Part 3 report, issued 9 May 2025, rejects the claim that training is inherently transformative. It is a pre-publication draft, no final version has been issued, and it is not law — but it is on the record and it does not agree with the two California district rulings.
What a working creator should do with this
The honest answer is: not much, directly. Training legality is a question about what your model provider did before you ever opened the app. It is not a question about your video.
Your exposure sits somewhere else, and it is worth separating the two clearly:
- Provider-side risk is the training question. It resolves in courts you are not a party to, on timelines you do not control. The practical hedge is picking tools whose data story you can actually inspect, which is why training-data provenance is worth treating as a buying criterion rather than a footnote.
- User-side risk is the output question. Whether the thing you generated reproduces protected expression — a character, a logo, a recognisable performance — is decided on your file, not on the corpus. Our copyright and safety guide for creators covers the practical version.
Three habits that hold up regardless of which way the Third Circuit goes:
- Read the indemnity, not the marketing. Copyright indemnities from major providers are generally enterprise or API tier, conditioned on using the safety filters. Consumer plans usually carry none. Assume none unless you can point to the clause.
- Do not confuse output ownership with output protection. Terms that assign a provider's interest in the output to you cannot create copyright where the human-authorship test fails. Those are separate systems.
- Keep the paperwork a client will ask for. Model used, date, prompt, whose likeness appears, what was licensed. When a dispute happens, the studio with records settles it in an afternoon. See usage rights for the vocabulary and legal and licensing for AI content in business for the contract-side version.
If a specific model's data history matters to your work, the model catalog is the place to start comparing what each provider actually discloses.
FAQ
When will the Third Circuit decide?
Nobody knows. Argument was heard 11 June 2026. Federal appellate courts do not publish decision calendars, and there is no deadline. It could land tomorrow or well into next year. Do not build a launch plan around the date.
Does the Anthropic settlement mean training on books is illegal?
No. A settlement is a private agreement and sets no precedent. The court's actual ruling in that case held that training on lawfully acquired books was fair use. The problem was the pirated library, which is a distinct issue from training.
If Ross loses the appeal, am I liable for what I generated?
No. An adverse ruling would be a holding about Ross's copying of a specific set of material for a specific non-generative product. It would not create a cause of action against a downstream user of an unrelated model. Downstream exposure comes from outputs that reproduce protected expression, and that risk exists today independent of this appeal.
Should I switch models based on how this comes out?
Only if your contracts require it. A more durable filter is whether a provider publishes something meaningful about its training data and whether your plan carries an indemnity you could actually invoke. Both of those are knowable now and neither depends on an opinion that has not been written.