Is Training AI Models Fair Use? What the Third Circuit's Thomson Reuters v Ross Ruling Decides

Is training AI models fair use? For the first time a US federal appeals court has given an answer, and in the case before it the answer was no. On Tuesday 29 September 2026 the Philadelphia-based Court of Appeals for the Third Circuit upheld Thomson Reuters' win against ROSS Intelligence, rejecting ROSS's argument that copying material derived from Westlaw headnotes to train an AI legal search engine was fair use. The catch is that the court's reasoning is sealed while the parties propose redactions, and the tool in question was not generative AI. This analysis explains what the appeal decided, what the trial court held on each fair use factor, why the non-generative label matters, what happens next and how the ruling sits beside Bartz v Anthropic, Kadrey v Meta, Getty Images v Stability AI and the EU's AI Act.

Is training AI models fair use? The short answer after Thomson Reuters v Ross

The short answer is that it depends on the use, and a commercial AI tool trained on a rival's editorial content to build a competing product has now lost at appellate level. Fair use under section 107 of the US Copyright Act is judged case by case against four factors: the purpose and character of the use, the nature of the work, the amount taken, and the effect on the market for the work. No statute says that AI training is, or is not, fair use.

What the Third Circuit has done is affirm a specific result. Judge Stephanos Bibas, a Third Circuit judge sitting by designation in the District of Delaware, held in February 2025 that Westlaw's headnotes were original enough to be protected and that ROSS's use of them was not fair use because ROSS was building a competing legal research tool. The appeals court agreed with that outcome. Until the opinion is released, nobody outside the court and the parties can say which parts of Judge Bibas's reasoning the panel adopted, narrowed or replaced.

Two limits matter for anyone applying the ruling to their own business. First, Judge Bibas expressly confined his analysis to non-generative AI, writing that "only non-generative AI is before me today". Second, the facts were unusually bad for the defendant: ROSS had asked for a licence, been refused because it was a competitor, and then obtained training material built from Westlaw content through a third party. So the honest answer to "is training AI models fair use?" after this ruling is: not when the training copies serve the same purpose as the original and feed a direct market substitute, at least in the Third Circuit, and at least for a tool that retrieves rather than generates.

What the Third Circuit decided on 29 September 2026

The Third Circuit affirmed. A three-judge panel of Judges L. Felipe Restrepo, Tamika R. Montgomery-Reeves and Emil J. Bove heard the case, and the opinion was written by Judge Montgomery-Reeves, according to Copyright Lately. What the public received on the day was a one-page judgment whose operative word is "AFFIRMED". Reports did not describe any dissent, but with the opinion sealed it is not yet public whether any judge wrote separately.

The sealing is procedural rather than unusual. The court's order, quoted by LawSites, says the opinion "has been temporarily filed as a sealed document" and directs the parties to propose redactions, with reasons for each, within 10 days. The underlying record contains confidential material: Judge Bibas himself noted that the actual questions and headnotes in the case were sealed, which is why his opinion illustrated the comparison with an invented example. Once the panel rules on the proposed redactions, the opinion will be published, in full or in part.

Thomson Reuters, the parent of Reuters News, said through a spokesperson that it was pleased with the ruling and "firmly believes that respecting copyright is essential for fostering innovation while protecting the intellectual property that fuels fiduciary-grade AI solutions". Reuters reported that spokespeople and an attorney for ROSS did not immediately respond to requests for comment. ROSS shut down its platform in 2021, citing the cost of the litigation. For advisers tracking US technology disputes, the Corporate INTL directory of US advisers covers the relevant practice areas.

Date	Step in Thomson Reuters v ROSS Intelligence
2020	Thomson Reuters and West Publishing sue ROSS in the District of Delaware
2021	ROSS shuts down its platform, citing litigation costs
2023	Judge Bibas largely denies summary judgment and sends fair use towards a jury
August 2024	Trial date continued; parties invited to renew summary judgment briefing
11 February 2025	Revised opinion: headnotes original, 2,243 infringed, fair use rejected
4 April 2025	Interlocutory appeal certified and the case stayed
17 June 2025	Third Circuit grants permission to appeal
11 June 2026	Oral argument before Judges Restrepo, Montgomery-Reeves and Bove
29 September 2026	Judgment: affirmed; opinion sealed pending proposed redactions

How the dispute began: Westlaw headnotes, Bulk Memos and a refused licence

The dispute began with a refused licence. According to Judge Bibas's opinion, ROSS set out to build an AI legal search engine and needed a database of legal questions and answers to train it. It asked to license Westlaw content, and Thomson Reuters refused because ROSS was a competitor.

ROSS then made a deal with LegalEase to obtain training data in the form of "Bulk Memos", lawyers' compilations of legal questions with good and bad answers. LegalEase gave its lawyers a guide explaining how to create the questions using Westlaw headnotes, while telling them not simply to copy and paste the headnotes. LegalEase sold ROSS roughly 25,000 Bulk Memos, which ROSS used to train its tool. At the summary judgment hearing, ROSS's counsel confirmed that ROSS had used 80% of the memos for initial training and 20% for later validation, which the judge treated as meaning ROSS had used practically all of them.

A headnote is a short statement of a point of law drawn from a judicial opinion, and Westlaw organises its content through the Key Number System, a numerical taxonomy of legal topics. Thomson Reuters accused ROSS of infringing 21,787 headnotes, the editorial decisions in 500 judicial opinions and the Key Number System itself. The February 2025 ruling dealt only with a batch of 2,830 headnotes, and found actual copying and substantial similarity for 2,243 of them, listed in a sealed appendix. The judge granted summary judgment only where copying was "so obvious that no reasonable jury could find otherwise", in practice where the headnote's language closely tracked the Bulk Memo question but not the underlying opinion.

The two certified questions and the route to the appeal

The appeal was interlocutory, meaning it came before any final judgment. Under 28 USC section 1292(b), a district judge may certify an order that involves "a controlling question of law as to which there is substantial ground for difference of opinion" where an immediate appeal may materially advance the end of the litigation, and the court of appeals then decides whether to take it. On 4 April 2025 Judge Bibas granted ROSS's request, certified two questions and stayed the case, postponing the trial that had been due the following month.

The Third Circuit granted permission to appeal on 17 June 2025. The appeal, No. 25-2153, concerned two issues, as summarised by Baker Botts:

1. Originality. Whether the West headnotes (and the Key Number System) are original works protected by copyright as a matter of law.
2. Fair use. Whether ROSS's use of the headnotes was fair use.

The parties framed those questions very differently in a pre-argument filing reported in June 2026. ROSS asked whether "a short quote or paraphrase of judicial holding" is copyrightable and whether fair use protects its "internal use" of headnotes in training memos. Thomson Reuters asked whether a legal research company that copied original editorial content from its competitor to build its own commercial product can escape liability. By affirming, the panel resolved the case in Thomson Reuters' favour on both certified issues; how broadly it did so will only be clear from the opinion.

Are Westlaw headnotes copyrightable?

Judge Bibas held that they are, and the Third Circuit affirmed. His starting point was that the text of judicial opinions is not copyrightable, and that the originality threshold set by the Supreme Court in Feist is "extremely low", requiring only "some minimal degree of creativity". The key question, he wrote, is whether a work is original, not how much effort went into it.

He found originality at two levels. The headnotes as a whole are a compilation whose selection and arrangement "easily clears that low bar". More strikingly, he held that each headnote is itself an individual copyrightable work, using a sculptor analogy: a block of marble, like a judicial opinion, is not copyrightable, but a sculptor who chooses what to cut away creates a protected work. On that view, "even a headnote taken verbatim from an opinion is a carefully chosen fraction of the whole", and all headnotes, even those quoting opinions word for word, have original value.

He also held that the Key Number System was original, because there are many logical ways to organise legal topics and it is enough that Thomson Reuters chose one.

That holding was the most criticised part of the 2025 ruling. Commentators, including Copyright Lately, questioned the sculptor analogy and the brief treatment of the merger doctrine, which withholds protection where an idea can be expressed in only a few ways. Judge Bibas had dismissed merger on the basis that "there are many ways to express points of law from judicial opinions". Copyright Lately also noted that Judge Montgomery-Reeves observed at argument that the ruling focused on headnotes that were not direct quotations, which may point to a narrower basis for affirmance. Whether the panel embraced the sculptor reasoning, or affirmed on narrower ground, is one of the questions the unsealed opinion will answer.

The four fair use factors as the trial court weighed them

The trial court split the four factors two-two and found that the two that matter most both favoured Thomson Reuters. Judge Bibas treated fair use as a question for the judge rather than a jury because, once actual copying was established, the remaining issues were about applying law to undisputed facts. Fair use is an affirmative defence, so ROSS bore the burden of proof.

Factor	Trial court finding	Key reasoning
1. Purpose and character	Thomson Reuters	Use was commercial and not transformative: ROSS used headnotes to build a competing legal research tool, with no further purpose or different character
2. Nature of the work	ROSS	Headnotes have creative elements but are far from the most creative works; factor rarely decisive
3. Amount and substantiality	ROSS	ROSS's output to users did not include headnotes, so little was made available to the public as a substitute
4. Market effect	Thomson Reuters	ROSS meant to create a market substitute for Westlaw, and harm to a potential market for AI training data was enough
Overall	Thomson Reuters	Factors one and four outweigh two and three; factor four is the most important

Factor one: why the use was not transformative

Applying the Supreme Court's 2023 decision in Warhol, Judge Bibas asked whether the original and the secondary use shared the same or highly similar purposes. ROSS's tool returned relevant judicial opinions in response to a legal question, a process that "resembles how Westlaw uses headnotes and key numbers to return a list of cases". He accepted that the intermediate nature of the copying, turning headnotes into numerical data about the relationships among legal words, made the factor "much trickier". But he held the intermediate copying cases, such as Sega, Sony v Connectix and Google v Oracle, were about computer code and depended on copying being necessary to reach unprotected functional elements. Neither was true here. His conclusion, quoted again by Reuters this week: "Ross took the headnotes to make it easier to develop a competing legal research tool. So Ross's use is not transformative."

Factor four: the market for AI training data

On market effect, the judge identified the original market as legal research platforms and "at least one potential derivative market" as "data to train legal AIs". It did not matter whether Thomson Reuters had itself used the data to train AI; "the effect on a potential market for AI training data is enough". He also rejected the public interest argument, noting that judicial opinions are freely available and that "the public has no right to Thomson Reuters's parsing of the law". Data access disputes are increasingly shaped by this kind of licensing-market reasoning, as the EU's order on Google search data sharing under the DMA shows from a regulatory angle.

Why "non-generative" matters

The non-generative point is the single most important limit on what the Thomson Reuters v Ross ruling decides. Judge Bibas said it was "undisputed that Ross's AI is not generative AI (AI that writes new content itself)"; when a user entered a legal question, ROSS returned existing judicial opinions. Reuters' report of the appeal makes the same distinction, noting that unlike the dozens of other AI training suits filed by authors, news outlets and music labels, this case "does not involve generative AI".

That matters because the transformative-use argument in generative AI cases rests on a different claim. Developers of large language models argue that training turns works into statistical parameters used to produce new text, images or code, a purpose different from reading or displaying the originals. In ROSS, the output competed directly with the source: a legal research engine trained on a legal research engine's editorial content. Commentary published this month by the American Bar Association's Business Law Today, before the appeal was decided, described the case as a narrower precedent than headlines suggested, pointing to direct competition with the source, paywalled content and the non-generative tool as distinguishing features.

The ruling is still likely to be cited in generative cases, in the view of commentators, for two propositions if the panel endorses them: that intermediate copying is not automatically transformative outside the software context, and that a potential licensing market for training data can count under factor four. Copyright Lately has pointed out that generative AI defendants rest their own defences on a similar idea, that training copies are only a step along the way, which is why they will read the panel's treatment of intermediate copying closely. Judge Bibas also noted that a reasonable jury might find bad faith in ROSS's conduct after the refused licence, but said it would not have changed his result. Legal teams using AI tools face a separate set of risks around accuracy, discussed in our analysis of AI hallucinations in court filings and the court warning after the Pinsent Masons AI error.

What each side argued in the Third Circuit

ROSS argued that its copying was an internal, intermediate step that produced a new kind of tool; Thomson Reuters argued it was straightforward substitution. At the 11 June 2026 argument, LawSites reported that ROSS's counsel, Mark S. Davies of White & Case, said ROSS had transformed headnotes into question-and-answer pairs, had used only 0.08% of Westlaw's 28 million headnotes and had built distinct deep-learning technology. Dale M. Cendali of Kirkland & Ellis, for Thomson Reuters, called it "a classic case of substitution" and identified three harms: competitive substitution, loss of exclusive use for Thomson Reuters' own AI training, and damage to potential licensing markets.

The bench's questions foreshadowed the outcome. Judge Bove pressed ROSS on how its engine differed from what happens when he logs into Westlaw in the morning, and, as Copyright Lately recounted, asked what was so "earth-shattering" about it. Judge Restrepo focused on market harm, and Judge Montgomery-Reeves explored whether West's editorial condensation of judicial language showed sufficient originality. Davies told the panel that headnotes simply track public domain opinions: "It's not poetry."

Outside parties lined up on both sides. Amici supporting ROSS, filed in October 2025, included the Electronic Frontier Foundation with library and access-to-law groups, the Authors Alliance, and law professors including Edward Lee, Matthew Sag, Pamela Samuelson, Christopher Sprigman and Rebecca Tushnet, who called the fair use analysis "unsound". MediaPost reported that Disney, other studios and publishers backed Thomson Reuters, warning that a ROSS win would chill incentives to invest in content. In April 2026 ROSS also drew the panel's attention to a Third Circuit fair use decision on public access to building codes, ASTM v UpCodes, according to a filing summarised at the time.

What happens next: unsealing, rehearing and the Supreme Court

The immediate next step is the redaction process, followed by publication of the opinion. After that, the usual routes for further review open, and the case can eventually return to Delaware for the issues that were never decided.

• Rehearing. Under Federal Rule of Appellate Procedure 40, a petition for panel rehearing or rehearing en banc must ordinarily be filed within 14 days after judgment is entered, unless the court shortens or extends the time. Whether the sealing affects the timetable in practice is a matter for the court.
• Supreme Court. Under Supreme Court Rule 13, a petition for certiorari is due within 90 days after entry of the judgment, with time running from the denial of any timely rehearing petition. Reports of the ruling did not indicate whether either side would seek further review.
• Back in Delaware. The case was stayed for the appeal. Judge Bibas left several matters for trial, including which headnotes remain covered by valid, unexpired copyright registrations, the separate batch of 5,367 headnotes, the Key Number System and the 500 annotated opinions, and damages. No damages figure has been reported.

Whether ROSS, which no longer operates a platform, has the appetite to seek en banc or Supreme Court review is unknown. Observers have noted that the practical stakes now lie less with the parties than with the many generative AI cases whose judges will read the panel's reasoning closely. Post-verdict and appellate strategy in technology cases is a specialist field, as our explainer on whether a jury verdict can be reversed in the Apple Taptic Engine case sets out, and dispute resolution and litigation advisers are listed in our directory.

How the ruling fits with other AI copyright decisions

Thomson Reuters v Ross is the only appellate ruling so far, but it lands among trial-level decisions that point in different directions depending on the facts. Read together, they suggest courts are separating the act of training from how the training data was obtained and what the trained system outputs.

Case	Court and date	Outcome on training	Key point
Thomson Reuters v ROSS	D. Del., Feb 2025; Third Circuit affirmed 29 Sept 2026	Not fair use	Non-generative tool; direct competitor; potential training data market
Bartz v Anthropic	N.D. Cal., June 2025	Training on books fair use; pirated central library not	$1.5bn class settlement, final approval July 2026
Kadrey v Meta	N.D. Cal., 25 June 2025	Summary judgment for Meta on fair use	Plaintiffs failed to prove market harm; ruling expressly narrow
Getty Images v Stability AI	High Court of England and Wales, 4 Nov 2025	Secondary infringement claim failed	Model weights not an "infringing copy"; permission to appeal granted
GEMA v OpenAI	Munich Regional Court I, 11 Nov 2025	Infringement found	Memorised song lyrics in model and outputs; TDM exception did not apply

In Bartz v Anthropic, Judge William Alsup ruled in June 2025 that training AI on books was fair use but that Anthropic's storing of more than seven million pirated books in a "central library" was not. The parties then settled for $1.5 billion, which Reuters called the largest known US copyright settlement, and Judge Araceli Martínez-Olguín granted final approval in July 2026. In Kadrey v Meta, Judge Vince Chhabria granted Meta summary judgment on 25 June 2025 because the plaintiff authors had not built a record of market harm, while cautioning that the ruling did not stand for the proposition that Meta's use was lawful.

Outside the US, fair use does not exist in the same form, and courts ask different questions. In Getty Images v Stability AI [2025] EWHC 2863 (Ch), Mrs Justice Joanna Smith held on 4 November 2025 that Stable Diffusion was not an infringing copy because it did not contain the works it was trained on; Getty had abandoned its primary training claim and won only a limited trade mark point on watermarks. The judge later granted Getty permission to appeal the secondary infringement question. In Germany, Munich Regional Court I held in GEMA v OpenAI that memorising song lyrics in model parameters and reproducing them in outputs infringed copyright. Patent and copyright strategy increasingly spans forums, as our guide to UPC patent litigation shows, and AI developers face other claims too, as in British Columbia's lawsuit against OpenAI.

Europe and the UK: a rules-based answer

In the EU the question is answered less by case-by-case fair use and more by statute. Article 4 of the 2019 Copyright in the Digital Single Market Directive permits text and data mining for any purpose unless rightsholders have expressly reserved their rights, and the AI Act now requires providers of general-purpose AI models to respect those reservations.

Article 53(1)(c) of the AI Act obliges providers to "put in place a policy to comply with Union law on copyright and related rights", in particular to identify and comply with a reservation of rights expressed under Article 4(3) of the Directive. Article 53(1)(d) requires a "sufficiently detailed summary" of the content used for training, following a template from the AI Office. The European Commission's General-Purpose AI Code of Practice, published on 10 July 2025, includes a copyright chapter; the obligations have applied since 2 August 2025, with Commission enforcement from August 2026. Our EU AI Act timeline sets out the wider sequence.

The UK has not settled its position. In March 2026 the government published its report on copyright and AI and, as Osborne Clarke reported, a broad text and data mining exception with a rightsholder opt-out is "no longer the government's preferred way forward"; it will gather more evidence and monitor litigation instead. That leaves the Getty appeal as the main live test of UK law, and UK advisers are watching both tracks.

What AI developers and content owners should take from it

The practical message of the ruling is about facts, not slogans. It does not decide that AI training is unlawful, but it shows how a court will treat training where the output competes with the source and a licence was refused.

• Map the use against the source's purpose. A tool that does what the source product does, for the same customers, is exposed on factor one whatever the technical method.
• Document provenance. Across Ross and Bartz, how data was obtained mattered: a third-party workaround after a refused licence in one, pirated libraries in the other. Business Law Today's commentary described lawful acquisition as the line that matters most.
• Take licensing markets seriously. If a court accepts that a potential market for training data counts, rightsholders that already license data are better placed to show harm.
• Plan for divergence. A model trained lawfully under US fair use may still face EU opt-out obligations under the AI Act and different questions in the UK and Germany.
• Watch the unsealed opinion. The breadth of the panel's reasoning on originality, intermediate copying and factor four will determine how far the ruling travels.

For law firms and in-house teams deploying AI, liability questions extend beyond training data, as our analysis of who is liable when law firms use AI and the Garfield AI approval explain. Advisers in technology, media and telecoms can help map the exposure.

When to get specialist advice

Businesses that train, license or deploy AI models on third-party content, and publishers whose content may be in training sets, should take advice before the unsealed opinion lands, because it may shift negotiating positions in licensing talks and pending cases. Corporate INTL lists intellectual property advisers worldwide, including in New York, and claimants weighing the cost of a copyright action may want to read our guide to litigation funding in the UK. The answer to "is training AI models fair use?" will keep depending on the facts, and Thomson Reuters v Ross has now shown one set of facts where it is not.

Frequently asked questions

Is training AI models fair use?

It depends on the facts. In Thomson Reuters v ROSS, the Third Circuit affirmed that training a non-generative legal search tool on material built from a competitor's Westlaw headnotes was not fair use. Two California trial courts found training on books could be fair use in Bartz v Anthropic and Kadrey v Meta, on different records.

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

On 29 September 2026 a panel of Judges Restrepo, Montgomery-Reeves and Bove affirmed Judge Bibas's February 2025 ruling that Westlaw headnotes are protected by copyright and that ROSS's use of them to train its AI tool was not fair use. The opinion, by Judge Montgomery-Reeves, is sealed pending redactions.

Why is the Thomson Reuters v Ross opinion sealed?

The court filed the opinion under seal temporarily and directed the parties to propose redactions, with reasons, within 10 days. The record contains confidential material, including the actual headnotes and training questions. Once the panel decides on the redactions, the opinion will be published in full or with parts blacked out.

Does the ruling apply to generative AI such as chatbots?

Not directly. The trial judge stressed that ROSS's tool was not generative AI and that only non-generative AI was before him; ROSS returned existing judicial opinions rather than writing new content. Generative AI defendants argue their use is different, but the appellate reasoning, once public, is likely to be cited in their cases.

Are Westlaw headnotes protected by copyright?

Yes, on the rulings so far. Judge Bibas held that the headnotes, individually and as a compilation, and the Key Number System clear the low originality threshold, even where a headnote quotes an opinion verbatim. The Third Circuit affirmed, although its exact reasoning on originality is not yet public.

Can AI get sued for copyright?

An AI system itself is not sued; the companies that build, train or operate it are. Authors, news organisations, music rightsholders and image libraries have brought dozens of claims against AI developers in the US, and cases such as Getty Images v Stability AI and GEMA v OpenAI have been decided in the UK and Germany.

What happens next in Thomson Reuters v Ross?

After the opinion is unsealed, ROSS could seek rehearing, ordinarily within 14 days of judgment, or petition the Supreme Court within 90 days. Otherwise the case returns to Delaware, where issues such as copyright validity for particular headnotes, other batches of headnotes and damages remain for trial.

How much did Anthropic pay to settle the authors' AI case?

Anthropic agreed to pay $1.5 billion to settle Bartz v Anthropic, described by Reuters as the largest known US copyright settlement, and the court granted final approval in July 2026. The settlement followed Judge Alsup's 2025 ruling that training on books was fair use but storing pirated copies was not.

How does the EU treat AI training on copyrighted works?

The EU allows text and data mining unless rightsholders reserve their rights. Under Article 53 of the AI Act, providers of general-purpose AI models must have a copyright compliance policy that respects those reservations and must publish a sufficiently detailed summary of their training content. These obligations have applied since 2 August 2025.

Sources

1. US District Court for the District of Delaware: Thomson Reuters v ROSS Intelligence, Memorandum Opinion (11 February 2025)
2. Claims Journal (Reuters): Appeals court upholds Thomson Reuters' landmark win in AI training lawsuit (30 September 2026)
3. Copyright Lately: Signed, Sealed, Affirmed: Third Circuit backs Thomson Reuters over ROSS (29 September 2026)
4. LawSites: 3rd Circuit issues opinion in Thomson Reuters v ROSS case, but for now it is sealed (29 September 2026)
5. ChatGPT Is Eating the World: Third Circuit affirms summary judgment rejection of fair use defence (29 September 2026)
6. MediaPost: Appeals court sides with Thomson Reuters (30 September 2026)
7. LawSites: At 3rd Circuit, judges press ROSS and Thomson Reuters on fair use, AI training and market harm (June 2026)
8. Baker Botts: Third Circuit hears oral argument in Ross v Reuters AI training case (July 2026)
9. ChatGPT Is Eating the World: Parties submit dueling questions posed by appeal (2 June 2026)
10. ChatGPT Is Eating the World: ROSS informs Third Circuit of recent fair use decision (9 April 2026)
11. ChatGPT Is Eating the World: Third Circuit grants ROSS interlocutory appeal (17 June 2025)
12. MLex: US judge OKs Ross interlocutory appeal, stays case (4 April 2025)
13. IPWatchdog: Amici back AI company's Third Circuit appeal (1 October 2025)
14. Claims Journal (Reuters): Judge approves Anthropic's $1.5bn settlement of copyright lawsuit (21 July 2026)
15. ChatGPT Is Eating the World: Meta prevails on fair use in Kadrey v Meta (25 June 2025)
16. Courts and Tribunals Judiciary: Getty Images v Stability AI [2025] EWHC 2863 (Ch) (4 November 2025)
17. The IPKat: Permission to appeal granted in Getty Images v Stability (January 2026)
18. CMS: GEMA v OpenAI, Munich Regional Court I issues landmark copyright decision (November 2025)
19. EU AI Act: Article 53, obligations for providers of general-purpose AI models
20. European Commission: The General-Purpose AI Code of Practice
21. Osborne Clarke: UK government moves away from preferred position on AI and copyright (March 2026)
22. Business Law Today: Thomson Reuters v Ross, one year later (September 2026)
23. Legal Information Institute: 17 US Code section 107
24. Legal Information Institute: 28 US Code section 1292
25. Legal Information Institute: Federal Rule of Appellate Procedure 40
26. Legal Information Institute: Supreme Court Rule 13

About this article

This analysis was researched and written by The Corporate INTL Newsroom, which covers cross-border legal, regulatory and business developments for lawyers, professional advisers and financiers in over 150 jurisdictions. It has been checked against the District of Delaware's February 2025 opinion, the relevant statutes and court rules, and primary reporting of the Third Circuit's judgment. The Third Circuit's opinion was sealed at the time of writing, so this article does not describe the panel's reasoning. This article is general information, not legal advice; for advice on a specific matter, consult a qualified adviser. Last reviewed 30 September 2026. For more analysis like this, visit the Corporate INTL newsroom or subscribe to Corporate INTL.