Copyright and AI law

US fair use and AI training after Thomson Reuters v Ross

Pink NYC lettering overlaps a blue illuminated skyscraper outline in a dark interior.
NYC in neon.Photograph by Robert Harper. Robert Harper Photograph Library, Archive.London. From the archive, shown for illustration.

The first US appeal court to rule on fair use and AI training sided with the copyright owner, then said in a footnote that its reasoning did not reach generative AI. Both halves matter.

On 29 September 2026 the US Court of Appeals for the Third Circuit became the first appellate court in America to rule on whether copying protected material to train an AI system is fair use. Its answer, on the facts before it, was no. The same opinion took care to say that those facts did not include a generative model.

This article is general information, not legal advice. Checked on 18 November 2026.

What happened in Thomson Reuters v Ross?

Ross Intelligence built a legal research tool. To train it, Ross used material derived from the headnotes that Thomson Reuters' editors write for Westlaw, the short summaries of legal points that sit above court opinions. Thomson Reuters sued for copyright infringement in Delaware.

On 11 February 2025 the district court, in a revised opinion by Judge Stephanos Bibas sitting by designation, held that the headnotes were protected and that Ross's use was not fair use. The question went to the Third Circuit as an interlocutory appeal. A panel of Judges Restrepo, Montgomery-Reeves and Bove heard argument on 11 June 2026, and Judge Montgomery-Reeves wrote the opinion affirming.

According to Haynes Boone's account, the panel held that the 2,243 headnotes at issue are original, because editors make creative choices about which points of law to draw out and how to phrase them. The opinion was first issued under seal, with the parties given time to propose redactions, as IPWatchdog reported.

How did the court apply the four fair-use factors?

American fair use, under section 107 of the Copyright Act, is weighed across four statutory factors. The Third Circuit came out this way, as summarised by Haynes Boone and Ballard Spahr:

Factor Finding Reasoning
1. Purpose and character of the use Favoured Thomson Reuters Commercial, and at most minimally transformative: both products help users find relevant law
2. Nature of the copyrighted work Slightly favoured Ross Headnotes are more factual than creative
3. Amount used Favoured Thomson Reuters Ross took more than it needed
4. Effect on the market Favoured Thomson Reuters Harm to Westlaw, and to a developing market for licensing headnotes as AI training data

Two passages have been widely quoted. On convenience as a defence, the panel wrote: "Unlike necessity, ease is not a justification for copying." And it said that incorporating AI into a product does not by itself excuse infringement.

Why does the licensing-market point matter?

The fourth factor asks what the use does to the market for the original, including markets the owner might reasonably develop. Thomson Reuters already used its headnotes to train its own AI products. The court treated the possibility of licensing them to others for the same purpose as a real, "rapidly developing" market, and Ross's unlicensed use as harm to it.

For anyone who owns content, that is the part of the opinion with the longest reach. A recurring argument for AI developers has been that a market for training licences is circular: it exists only if the use is not fair, so it cannot be counted in deciding whether the use is fair. The Third Circuit counted it. Whether other circuits follow, and whether the reasoning carries over to generative models, is undecided.

Does the ruling cover generative AI?

By its own terms, no. The court wrote that "ROSS's AI platform cannot generate original expression", as quoted by Ballard Spahr. It returned existing passages from judicial opinions. On that basis, according to Haynes Boone's account, a footnote put to one side the district rulings on generative models and the US government's statement of interest in the OpenAI litigation.

So the opinion establishes that using copyright material to train a tool that competes directly with the source can fail fair use at appellate level. It does not establish what happens when the trained system produces new text or images.

Ross has asked the Third Circuit to revisit the decision, according to MLex. We have not seen a ruling on that request.

How do Bartz and Kadrey fit in?

Two decisions from the Northern District of California in June 2025 remain the main authorities on generative training, and both concern books and language models.

Bartz v Anthropic (23 June 2025). Training on books the company had lawfully bought was held to be fair use. Keeping a central library of pirated books was not. The case then settled for $1.5bn, about $3,000 per work. Judge Araceli Martínez-Olguín granted final approval on 20 July 2026 and cut class counsel's fee award to roughly $101.6m. We could not verify whether any appeal from the approval is pending.

Kadrey v Meta (25 June 2025). Summary judgment went to Meta on training, on what the judge described as a thin record about market harm. He added that a better-evidenced argument about market dilution, meaning a flood of AI-generated works competing with the originals, could defeat fair use in another case. The litigation continues on claims about distributing books through BitTorrent.

Read together with Ross, the decided cases agree on less than headlines suggest. All three treat the source of the copies and the evidence of market harm as central. They differ on how much weight transformation carries and on what kind of market harm counts.

The US Copyright Office released Part 3 of its report on copyright and AI, covering generative training, as a pre-publication version on 9 May 2025. It says fair use is fact-specific, that some training uses will qualify and others will not, and that pirated sources and market dilution weigh against. The administration moved to remove the Register of Copyrights the day after its release. Litigation over her position followed and, according to Lowenstein Sandler, court orders have so far allowed her to remain in office while it continues. The Office's website still lists Part 3 as a pre-publication version, with a final version to follow. The report is not binding on courts either way.

By early September 2026 the United States had filed a statement of interest in the OpenAI litigation arguing that training on text is highly transformative fair use. That is a different emphasis from the Office's report, and it is advocacy, not a ruling.

What about photographs and image models?

Here the honest answer is that nothing is decided. As of 18 November 2026:

  • Getty Images v Stability AI (N.D. Cal.) is at the pre-trial stage. A 23 April 2026 order let trade mark and related claims proceed; the copyright claim was not challenged at that stage.
  • Andersen v Stability AI, Midjourney and others, the artists' class action filed in January 2023, has not reached trial, and its schedule was extended twice in 2026. At least one widely indexed post claims a jury trial took place in September 2026. Docket-based sources contradict that.
  • Disney and Universal v Midjourney, consolidated with Warner Bros. Discovery's suit, is in discovery.
  • Evox Productions, an automotive photo library, filed three suits in July 2026 against Stability AI and others, Google, and Midjourney. They are at an early stage.

Each is tracked, with dates, in our AI copyright cases tracker. The English side of the Getty dispute, which turned on different law altogether, is in our Getty v Stability AI timeline.

The difference from Europe is structural. The EU and UK have written exceptions with defined conditions; the US has an open-ended defence decided case by case. Our comparison of text and data mining exceptions sets the three side by side.

What rights holders and buyers can take from this

Three cautious observations.

  1. Provenance of the copies matters in every decided case. Lawfully acquired material fared better than pirated material in Bartz. Knowing where a file came from is the beginning of any analysis, and tools such as our image hash and fingerprint generator help keep a record of exactly which files changed hands.
  2. Evidence of a licensing market helps rights holders. The Third Circuit relied on one. Each licence that is actually signed adds to that evidence.
  3. Uncertainty has a cost for both sides. A buyer who licenses is not betting on how a future court weighs four factors.

The archive's position

The Robert Harper Photograph Library is a single-photographer archive of 95,487 unique photographs on the inventory's count, owned by 825 Ventures Ltd, the owner and rights holder. It includes controlled studio product photography, such as the fragrance collections, and more than half of the archive is still on film and has yet to be scanned, and only a small selection has been published online by Archive.London.

None of that depends on how American courts resolve fair use. Licences are negotiated directly with the rights holder, subject to rights review, as described on our rights and stewardship page. Teams weighing licensed image data can start at AI training data licensing.

Frequently asked questions

Is AI training fair use in the United States?

It depends on the facts, and the courts have not agreed. Two California district judges found fair use for training language models on books in June 2025. The Third Circuit rejected fair use for a non-generative legal search tool on 29 September 2026. No court has ruled on an image model. Neither 'training is fair use' nor 'training is not fair use' is a safe general statement.

What did the Third Circuit decide in Thomson Reuters v Ross?

It affirmed partial summary judgment for Thomson Reuters. The panel held that Westlaw's editorial headnotes are original enough for copyright, and that Ross's use of them to build a competing legal research tool was not fair use. The first, third and fourth factors favoured Thomson Reuters; the second slightly favoured Ross.

Does Thomson Reuters v Ross apply to generative AI such as image generators?

Not directly. The court said Ross's platform, unlike the models in Bartz and the OpenAI litigation, cannot generate original expression, and it set those cases aside on that basis. Its reasoning about competing products and licensing markets may still be cited in generative cases, but the question was expressly left open.

Does the Bartz v Anthropic settlement set a price for training data?

No. The $1.5bn settlement, about $3,000 per work, resolved claims over pirated copies of books where the defendant faced statutory damages. It is not a licence fee, it was not set by a market, and it says nothing about photographs. Treat it as a litigation outcome, not a rate card.

Sources

This article is general information, not legal advice. It reflects the position as we understood it on the date shown and may be out of date. Published by Archive.London. Corrections are welcome under our editorial policy.

Licensing the archive for AI and data projects

95,487 photographs from one photographer and one rights holder. Read the dataset card, search the 503-folder inventory, discuss a sample set under NDA, or arrange a first look as scanning proceeds.

See AI and data licensingWrite to the archive