What the Third Circuit decided, and on what facts

On 29 September 2026 the United States Court of Appeals for the Third Circuit decided Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153. The panel was Judges Restrepo, Montgomery-Reeves and Bove, and Judge Montgomery-Reeves wrote the opinion after argument on 11 June 2026. The appeal was interlocutory. The District of Delaware had certified two questions: whether Westlaw's headnotes and Key Number System are original as a matter of law, and whether ROSS's use of the headnotes was fair use. The court answered the first for Thomson Reuters as far as the headnotes go. It did not address the Key Number System, because ROSS never argued it. It answered the second against ROSS, and affirmed partial summary judgment that ROSS infringed 2,243 headnotes and that the copying was not fair use. There are no concurring or dissenting opinions.

The court wanted the case read as an ordinary one. Under ROSS's framing, it said, the case "appears to concern the future of AI legal technology. But appearances can be deceiving. In truth, this is no more than an ordinary copyright case". The holdings rest on established authority: Feist on originality, Andy Warhol Foundation v. Goldsmith on purpose, and Campbell, Harper & Row and the court's own Video Pipeline decision on the amount taken and on market harm. Nothing in the judgment depends on a special theory of artificial intelligence.

The judgment does not end the case. Judge Bibas's February 2025 opinion left several issues for trial, and they now return to Delaware. They include whether some headnote copyrights have expired or fall outside the registrations, the remaining headnotes, the Key Number System, the editorial content in about 500 opinions, and damages.

The reasoning depends on the facts, and the court stressed how those facts differ from the large generative-model cases. ROSS was founded by three University of Toronto computer science students after an IBM competition built around Watson. Its product answered plain-language legal questions by returning passages from about ten million uncopyrighted judicial opinions. The court notes that it was not generative: it "would not create any new expression".

To train it, ROSS commissioned about 25,000 memos from a company called LegalEase. Each memo posed a legal question and offered four to six opinion passages, labelled great, good, topical or irrelevant. The drafters, at LegalEase and a subcontractor, used thousands of Westlaw headnotes to frame the questions because they offered "an easy way" to do it. The "great" answer was most often the passage Westlaw linked to that headnote. For the appeal, ROSS accepted that the contractors' conduct was attributable to it and that they copied headnotes.

Judge Bibas's memorandum opinion of 11 February 2025 adds two background facts. ROSS had asked Thomson Reuters for a licence and was refused because it was a competitor. And ROSS's counsel confirmed that 80% of the memos were used for training and 20% for validation, so nearly all were used. The judge compared memo questions, headnotes and opinions one by one for a batch of 2,830 headnotes that ROSS's own expert had described as closely tracking the headnotes. He granted summary judgment only for the 2,243 where no reasonable jury could find otherwise. ROSS's commercial aim was open: it advertised against Westlaw at prices "in line with" Westlaw's, and some law firms switched.

Originality: why a headnote is protected

Under Feist, the originality bar is "extremely low". The court held that all 2,243 headnotes had "some creative spark". Editors chose which points of law mattered and how to express them under Westlaw's drafting rules, and each headnote had to stand alone while accurately reflecting the opinion. The court relied on Callaghan v. Myers (1888), which recognised a reporter's copyright in headnotes, and on Georgia v. Public.Resource.Org (2020), which allows protection for explanatory material written by someone with no authority to make law.

The court rejected three counter-arguments. On the claim that this gives Thomson Reuters a monopoly over the law, it said: "Headnotes are not law; judicial opinions are". The merger doctrine did not apply, because there are many ways to express a point of law; RELX, Lexis's parent company, explained as amicus that Lexis writes different headnotes for the same opinions. And Matthew Bender v. West (2d Cir. 1998) concerned parallel citations and similar additions dictated by industry convention. That court had itself called headnotes "independently composed".

Two footnotes set limits. Footnote 5 leaves open whether headnotes that quote an opinion verbatim are original; Judge Bibas thought they could be. Footnote 10 affirms that each headnote is a separate work, which becomes important under the third factor.

Factor one: same purpose, and the end of the intermediate-copying argument

Following Warhol, the court weighed the degree of difference in purpose against commerciality. ROSS's use was commercial: it set prices to win Westlaw's customers. On purpose, the court compared the two businesses. Thomson Reuters uses headnotes to help researchers find and understand opinions. ROSS used them to build a platform that helps users find responsive opinion passages. Training an AI program was "an intermediate step" that "arguably presents a slight degree of difference". But both used the headnotes "to create and optimize a legal-research platform that helps users find responsive legal material". The use was "minimally transformative, at best".

Authors Guild v. Google did not help ROSS. Google Books served a different function from reading and could lead users back to the books. ROSS, by its own admission, aimed to replace Westlaw.

The answer to the intermediate-copying argument is the part most likely to be cited in other cases. ROSS relied on Google v. Oracle (2021), Sega v. Accolade (9th Cir. 1992) and Sony v. Connectix (9th Cir. 2000). The court read all three as turning on necessity: copying code was necessary to reach unprotected functional elements and make software work with an existing system. Footnote 8 adds that the copied Java interface had value because programmers were used to it, while headnotes "do not derive value from others copying them". ROSS did not need the headnotes. It had the opinions and "could freely copy them". It chose the headnotes because they were "easy", and the court held: "Unlike necessity, ease is not a justification for copying."

Judge Bibas had reached the same result partly on a different ground: the intermediate-copying cases concerned computer code, which is functional, and not written text. The Third Circuit placed less weight on code against text and more on necessity. A rule based on subject matter would have been easy to distinguish in later cases. A rule based on necessity can apply to any training input.

Footnote 9 addresses good faith. The court noted the Supreme Court's doubts about its relevance. It then recorded undisputed evidence that ROSS "at times acted in bad faith": using an investor law firm's credentials despite notice that the terms of service forbade it, an employee posing as a solo practitioner to obtain an account, and another using a student account while concealing that he worked for a competitor. To the extent good faith matters, it weighs against ROSS. Judge Bibas had treated bad faith as irrelevant to the outcome.

Factors two and three: where the appeal court went further

The second factor, the nature of the work, favoured ROSS. The headnotes were published and "more factual than fictional". As usual, it carried little weight, and the court's conclusion describes it as weighing only "slightly" for fair use.

The third factor is where the Third Circuit departed from the district court. Judge Bibas decided it for ROSS. He relied on Authors Guild: what counts is how much is made available to the public as a potential substitute, and ROSS's users never saw a headnote. The Third Circuit instead asked whether "no more was taken than necessary" for the copier's purpose. The purpose was minimally transformative, and the opinions were freely available, so copying the headnotes "was also not necessary to train ROSS's AI". ROSS argued that it took only 0.08% of 28 million headnotes. The court replied that, under footnote 10, "for each headnote taken, ROSS copied an entire work".

One detail: the passage says ROSS copied "the entire text of the 25,000 Westlaw-written headnotes", while elsewhere 25,000 is the approximate number of memos and the judgment covers 2,243 headnotes. The figure appears to conflate the two. The holding does not depend on it.

The doctrinal point matters more. AI developers commonly argue that the third factor should be measured by what a model outputs, not by what goes into training. On these facts the Third Circuit measured it at the point of copying and against necessity. How far that carries into cases with a strongly transformative purpose is open. The court tied its reasoning to the absence of such a purpose, and Campbell holds that permissible copying varies with the purpose of the use.

Factor four: a market for training data that did not yet exist

On the original market, ROSS said no standalone market for headnotes exists. The court answered that the statute protects the value of the work, citing its own decision in Video Pipeline v. Buena Vista (2003) on film trailers. Thomson Reuters uses headnotes to sell Westlaw, and ROSS took their value and weakened them as a draw. The court also treated the legal-research platform market as relevant. "Unrestricted and widespread" copying to build substitute tools would harm Westlaw there, and ROSS offered no rebuttal.

The derivative-market holding is the one most likely to be cited in later cases. Thomson Reuters argued that ROSS harmed "the potential derivative market for licensing headnotes as AI training data". Applying Campbell and Castle Rock, the court held that a market is not illusory just because the author has shown little interest in exploiting it. It found that:

  • the evidence showed "the market for licensing headnotes as text to train AI is rapidly developing";
  • Thomson Reuters was using its headnotes as training data for its own AI search products, such as WestSearch Plus, which the opinion mentions in its account of the facts;
  • "That Thomson Reuters did not license its headnotes to others does not disprove that a market exists to do so."

By training on the headnotes without authorisation, ROSS "usurped Thomson Reuters's opportunity to enter that derivative market". Judge Bibas had put it more bluntly: "the effect on a potential market for AI training data is enough".

The public-benefit arguments failed. The opinions are free, and ROSS charged Westlaw-level prices, so better access to law was "unclear". There was no evidence that the ruling would halt AI development. On national security, the court noted that the Department of Justice had raised such concerns elsewhere but not here, and that AI gives no "carte blanche to violate copyright law". The overall balance: factor two favours fair use slightly, and factors one, three and four weigh against it.

Footnote 7: what the court said about generative AI, and what it did not

Footnote 7 refers to the Statement of Interest the Department of Justice filed on 1 September 2026 in In re: OpenAI, Inc. Copyright Infringement Litigation, No. 1:25-md-3143 (S.D.N.Y.). As the court summarises it, the Department relied on Bartz v. Anthropic to argue that training a large language model that can "generate original responses" is transformative, and that such training did not cause "substitutive competition". The Third Circuit said these concerns "do not apply here". ROSS's platform "cannot generate original expression", and ROSS trained it to build a commercial substitute for Westlaw. The court added that the Department "knows how to assert its interests" but "notably did not do so here".

The footnote does three things. It expressly sets generative AI aside. It accepts, implicitly, that whether a system can produce new expression is relevant to transformation. And it treats the government's silence in this case as meaningful. What it does not do is just as important. It does not say that generative training is fair use, it does not adopt the Department's reasoning, and it does not limit the court's general statements on necessity, the training-data market or the third factor. Those statements are not tied to any type of AI, and parties in generative cases will cite them.

What it means for Bartz, Kadrey and the OpenAI litigation

This is a precedential opinion of the Third Circuit, which covers Delaware, New Jersey, Pennsylvania and the US Virgin Islands. The best-known generative cases are in the Northern District of California (Ninth Circuit) and the Southern District of New York (Second Circuit), where it is persuasive rather than binding. It is nonetheless widely described as the first decision by a federal court of appeals on fair use and AI training. Two district-court rulings are its main points of comparison:

  • Bartz v. Anthropic (N.D. Cal., 23 June 2025, Judge Alsup) held that training on books was "exceedingly transformative" and fair use, and that digitising lawfully bought print copies was fair use. Building a permanent library from pirated copies was not. On a market for licensing books as training data, the court wrote that "such a market for that use is not one the Copyright Act entitles Authors to exploit". The case later settled for $1.5 billion; according to the Authors Guild and press reports, the court granted final approval of the settlement on 20 July 2026.
  • Kadrey v. Meta (N.D. Cal., 25 June 2025, Judge Chhabria) granted Meta partial summary judgment on thirteen authors' training claims. The judge held that they "are not entitled to the market for licensing their works as AI training data", because lost licence fees for a transformative use are not cognisable harm and counting them would make the analysis circular. He stressed that the ruling did not make Meta's use lawful, only that "these plaintiffs made the wrong arguments". He identified market dilution as the potentially winning theory.

The tension over the licensing market is real but narrower than it looks. Kadrey excluded licence-fee harm only for transformative uses. The Third Circuit counted the training-data market after it had found the use minimally transformative. On that reading, the factor-one result determines whether the market counts. The Third Circuit did not address circularity, however, and its language is general. Rights holders will rely on it, and defendants will argue that it applies only to minimally transformative uses.

The necessity reasoning will also be tested. Generative developers argue that only large and varied corpora teach the patterns of language, which is a functional need of the kind the intermediate-copying cases protected. ROSS does not decide that question, because ROSS had a free substitute: the opinions. Plaintiffs will point to licensed or public-domain alternatives. Defendants will say none would produce a comparable model. "Ease is not a justification for copying" also fits conduct where developers chose pirated sources over lawful acquisition, which Bartz treated separately and refused to excuse.

The Department of Justice's filing is now on the record in In re OpenAI, and the Third Circuit has already distinguished it. In Concord Music Group v. Anthropic (N.D. Cal.), case trackers report a summary-judgment hearing on 21 October 2026, though the date may change. Each of these courts will now have to explain why generative training is different, if it finds that it is.

The strongest objections to the ruling

Originality was decided wholesale. Westlaw's guidelines tell editors generally to "follow the court's language". The court decided originality for 2,243 headnotes together, based on a general description of the editorial process. The district court's exclusion of verbatim headnotes and Lexis's evidence of different drafting are a reasonable answer, but the verbatim question remains open.

The necessity test may prove too much. Almost any training input could be replaced, at greater cost, by material the developer creates itself. If ease never justifies copying, cost does not matter, and intermediate copying is protected only where there is technical lock-out. Judge Bibas said there was nothing Thomson Reuters created "that Ross could not have created for itself". Critics will see this as a rule that developers must build their own data whenever they can. Defenders will say that is exactly what copyright asks of a direct competitor.

The purposes were compared at a high level. At the level of "helps users find responsive legal material", Google Books also "helps users find" books. The real distinction is substitution. That is consistent with Warhol, but it means factors one and four did much of the same work.

Circularity. Recognising a training-licence market because the owner could license, and uses the data internally, comes close to the circularity Kadrey warned against. The implied answer is that the circularity concern protects transformative uses, and this use was not one. An express statement would have helped.

None of this makes the result wrong on these facts: a direct competitor, a refused licence and a substitute product. The objections concern how far the reasoning should be extended to other cases.

The European comparison: Article 4 DSM and the AI Act

The EU works with exceptions rather than a flexible defence. The DSM Directive (2019/790) provides a research TDM exception in Article 3 and a general one in Article 4. Article 4 covers reproductions and extractions of lawfully accessible works unless the rightsholder has expressly reserved the use "in an appropriate manner, such as machine-readable means" for online content. Article 2(2) defines TDM as "any automated analytical technique" for analysing text and data to generate information. Under Article 53(1)(c) and (d) of the AI Act, providers of general-purpose AI models must have a policy to comply with EU copyright law, including respecting Article 4(3) reservations, and must publish a summary of training content.

A ROSS-type case in the EU would probably also go against the developer, but by a different route:

  • Protection. In Infopaq (C-5/08), the Court of Justice accepted that even an eleven-word extract can be protected if it is the author's own intellectual creation. Editorial headnotes would very likely qualify, and the database could also attract the sui generis right under Directive 96/9/EC.
  • TDM would probably not help. Human memo-writers rewriting headnotes is not "automated analysis". Article 4 requires lawful access, which for a subscription database is governed by contract. And a commercial publisher would be expected to reserve its rights.
  • No balancing. Article 7(2) of the Directive applies the three-step test of Article 5(5) InfoSoc, but there is no general weighing in which public benefit could offset a "minimally transformative" purpose.

On 29 September 2026, the day of the Third Circuit's ruling, the Commission opened a targeted consultation on technology and copyright, closing on 3 November 2026. Its first topic is the use of protected content in AI.

Türkiye: no fair use, no TDM exception, and a bill that goes the other way

The ROSS facts have direct parallels in Türkiye. Commercial case-law databases add summaries, keywords and classifications to judgments, and legal-tech firms are building AI research tools on them. Turkish law would analyse such a case differently in four respects.

1. Judgments are free; the editorial layer is not. Article 31 of the Law on Intellectual and Artistic Works (FSEK, No. 5846) makes officially published laws, regulations, circulars and court decisions ("kazai kararlar") free to reproduce, adapt and use. Under Article 1/B(a), however, a work is protected if it bears its author's "hususiyet" (individual character). A summary that reflects real editorial choices can meet that test, and Article 6(1)(9) lists the explanation or abridgement of another work among adaptations. Turkish doctrine generally reads hususiyet as requiring the author's personal imprint, which is arguably stricter than Feist's "modicum of creativity". Short, formulaic summaries may therefore fail where Westlaw's headnotes succeeded. Each has to be assessed individually.

2. The database is protected twice. Article 6(1)(11) protects databases whose selection and arrangement follow a particular plan, but not the data inside them. Additional Article 8 (Ek Madde 8), added in 2004 and modelled on the EU Database Directive, gives a maker who has made a substantial investment in creating, verifying or presenting the contents a right to prohibit the permanent or temporary transfer of all or a substantial part of the contents to another medium. The right lasts fifteen years. It does not require individual summaries to be original, and it targets the kind of extraction that building a training set involves.

3. There is no general defence. FSEK has no fair-use clause. Its limitations, set out from Article 30 onwards, form a closed list, and none is drafted for automated analysis or training. Private use under Article 38 requires personal use without profit. Article 22 defines reproduction as copying "directly or indirectly, temporarily or permanently", so the intermediate copies made during training are reproductions, with no fair-use analysis in which to argue that they are only intermediate. A ROSS-type defendant would lose before reaching any factors, because Turkish law has none. Infringement can also attract criminal liability under Article 71.

4. Unfair competition and contract. Under Article 55(1)(c)(3) of the Turkish Commercial Code (TTK), taking over another's market-ready work product by technical reproduction, without an appropriate contribution of one's own, is unfair competition. This fits automated scraping better than the ROSS facts, where human contractors rewrote the material. Access in breach of subscription terms, like the conduct described in footnote 9, also creates contractual liability.

Data protection adds a separate issue. Judgments contain personal data. The ground in Article 5(2)(d) of the KVKK (No. 6698), data made public by the data subject, does not obviously apply where a court published the decision. A training set built from Turkish judgments therefore raises a question that a US fair-use analysis never reaches.

On policy, Türkiye has no TDM exception, and the proposal on file goes the other way. Bill No. 2/3634, a private member's bill dated 11 March 2026 by Halil Öztürk, MP for Kırıkkale, would add, among other provisions, Article 42/D to FSEK. It would make reproducing, processing, data mining or storing works to train, develop, fine-tune or test AI systems offered in Türkiye subject to a licence "for appropriate remuneration". The licence would be administered by a single joint licensing body with extended collective effect, from which rightsholders could opt out. The licensing duty would also extend to commercial use of outputs that substitute for protected works. The bill has been referred to the National Education, Culture, Youth and Sports Committee as lead committee. It is a proposal, and we have found no record of it being debated. The AI Action Plan 2026–2030 (Presidential Circular 2026/9, Official Gazette of 18 August 2026) does not address copyright in training data.

Korea: an open fair-use clause and a market-minded Supreme Court

Korea sits between the US and EU models. Article 35-5 of the Copyright Act, introduced in 2011 as Article 35-3 and renumbered in 2019, permits a use that does not conflict with the normal exploitation of a work and does not unreasonably prejudice the author's legitimate interests. It lists four factors that are close to those in section 107, including the effect on the "current or potential market or value". A Korean court can therefore carry out a ROSS-style analysis.

The Supreme Court's recent approach to the fourth factor is close to the Third Circuit's. In its judgment of 11 July 2024 (2021다272001), the Court held that the Korea Institute for Curriculum and Evaluation's unrestricted online posting of exam papers containing protected works was not fair use, despite its public and non-profit character. The fourth factor asks whether the use replaces demand in the current market or in an ordinary market "with a reasonable probability of being developed in the future". A market for licensing such works as study material already existed. A market for licensing their online transmission was one that could reasonably be expected to develop. Like the Third Circuit, the Court held that a market does not have to be mature to count.

Like FSEK, the Korean Act leaves court judgments unprotected (Article 7) and gives database producers a separate right over a substantial part of their database (Article 93). The Unfair Competition Prevention and Trade Secret Protection Act also contains a general clause against misappropriating another's work product, which a rightsholder could invoke alongside copyright. Korea has no TDM exception. Earlier proposals, notably the 2021 full revision bill that would have permitted copying for information analysis, lapsed when the 21st National Assembly's term ended in May 2024. On 26 February 2026 the Ministry of Culture, Sports and Tourism and the Korea Copyright Commission published a guide on fair use in generative-AI training, which states that it is not an official interpretation. According to published summaries, it does not exclude commercial training or web crawling from fair use as such, weighs the four factors together, and under the fourth factor asks whether the use substitutes for existing or reasonably foreseeable markets, including lost licensing opportunities. In March 2026, press reports said the government was considering wider reform, including Copyright Act amendments and an opt-out system in sectors with established licensing markets. As far as we can verify, nothing has been enacted.

What organisations should do

  • Separate raw material from editorial layers. Judgments and statutes are free in the US, Türkiye (FSEK Article 31) and Korea (Article 7). Publishers' summaries, keywords, classifications and annotations are not. Training pipelines should record which layer each item comes from.
  • Do not use a competitor's editorial work as a labelling shortcut. The infringement in ROSS came from annotators who found headnotes "an easy way" to write questions, not from scraping. Annotation guidelines and vendor contracts should prohibit it, and audits should check compliance. ROSS was held responsible for its contractors' copying.
  • Document necessity, not convenience. A developer relying on intermediate copying in the US should be able to explain why the specific protected material was needed and why free sources would not do.
  • Respect access terms. As footnote 9 shows, borrowed or disguised accounts become evidence. In Türkiye they also create separate contractual and unfair-competition exposure.
  • Treat licensing activity as evidence. Rightsholders that rely on a training-data market should document actual licensing and internal use. Developers should expect their own licence deals to be cited as proof that the market exists.
  • Add a data-protection review for datasets built from Turkish judgments: anonymisation, a lawful basis for processing, and transfers abroad under KVKK.
  • Do not over-read the ruling. For generative models, ROSS is persuasive authority on necessity and licensing markets. It does not hold that generative training is unlawful.

What to watch next

  • Whether ROSS petitions the Supreme Court for certiorari. According to press reports of 2 October 2026, ROSS's counsel said it intends to seek Supreme Court review; we have found no petition on file yet.
  • The Delaware trial on the remaining issues and on damages.
  • Summary-judgment rulings in In re OpenAI (S.D.N.Y.) and Concord v. Anthropic (N.D. Cal.), and whether they follow or distinguish the Third Circuit on necessity and the licensing market.
  • The EU consultation, which closes on 3 November 2026, and any proposals on opt-outs, transparency and remuneration that would affect Turkish and Korean providers selling into the EU.
  • Whether the committee takes up Bill 2/3634 in Türkiye, and whether Korea's announced reform produces a TDM exception, an opt-out system or a remuneration scheme.

Frequently asked questions

Did the court hold that training AI on copyrighted material is unlawful?

No. It held that this defendant's copying of these headnotes, to train a non-generative tool that competed directly with the source, was not fair use. Footnote 7 expressly sets generative AI aside.

Why did the intermediate-copying argument fail?

The court read Oracle, Sega and Connectix as cases where copying was necessary to reach unprotected functional elements. ROSS could have trained on the free opinions and chose the headnotes because that was easier. As the court put it, "ease is not a justification for copying".

Can a market count if the owner has never licensed into it?

According to the Third Circuit, yes. The evidence showed a "rapidly developing" market for licensing headnotes as training data, and Thomson Reuters used its headnotes to train its own products. Two California courts took a narrower view where the use was transformative.

Would the result be the same in Türkiye?

Very probably, but by a different route. There is no fair use and no TDM exception. The questions would be hususiyet in the summaries, extraction from a database protected by Additional Article 8 FSEK, and unfair competition under the TTK. Each depends on the facts.

Is Korea closer to the US or the EU?

In method, closer to the US. Article 35-5 is an open four-factor clause, and the Supreme Court's 2024 judgment recognised markets that could reasonably be expected to develop. In policy, Korea is considering EU-style opt-out and licensing tools but has not enacted a TDM exception.

Burhan Doğuş Ayparlar's View

This section sets out my personal assessment as the founder of this site and an attorney at law (Türkiye).

The headlines call this the first appellate verdict on AI and fair use. The court tried hard to prevent that reading. It called the dispute "no more than an ordinary copyright case", set generative AI aside in a footnote, and decided facts that were about as unfavourable to a defendant as AI-training facts can be: a direct competitor, a refused licence, contractors writing from the rival's own summaries, and accounts obtained under false pretences. On those facts I think the result is correct.

What will last are two lines of reasoning. The first reads the intermediate-copying cases as cases about necessity. I think that is an accurate account of what Oracle, Sega and Connectix decided, and it changes the question in every training case to whether the specific protected material was needed or only convenient. Developers who chose pirated or scraped material because it was quicker will find the court's words about ease quoted against them. The second is the recognition of a training-data market that the rightsholder had not yet entered. The court tied that to a use it found minimally transformative. I would have preferred it to say so expressly, because without that limit the reasoning becomes circular, a risk the California courts had already identified.

For Türkiye, the decision is most useful as a mirror. American courts can weigh factors. Turkish courts have nothing to weigh, because FSEK offers a closed list of exceptions, none written for automated analysis, and a reproduction right that includes temporary copies. In practice, almost any unlicensed commercial training on protected Turkish content is infringing, whether the model matches headnotes or is a frontier language model. Turkish legal-tech firms building on commercial case-law databases should read ROSS as a description of their own exposure under FSEK Additional Article 8 and TTK Article 55, not as a foreign case with defences they could rely on at home.

That is why legislation matters more than litigation here. The only proposal on file, Bill 2/3634, would require licences through a single collective body for any AI training. Its premise is sound: rightsholders should be paid, and the Third Circuit's recognition of a training-data market supports that. But a regime that consists only of licensing, with no research exception and no commercial exception subject to opt-out, would leave Turkish universities and start-ups behind their European counterparts without clearly benefiting Turkish authors. In my view, the better course would follow what Türkiye did for databases in 2004 and track EU law. That means a research TDM exception, a commercial TDM exception subject to a machine-readable reservation, transparency about training sources, collective licensing for those who reserve their rights, and a separate answer to the data protection questions raised by training on judgments. The United States is answering this question through case law. Türkiye will have to answer it through legislation, and it should do so deliberately rather than leave the current position in place by default.

This article is for information only and does not constitute legal advice. It is based on the Third Circuit's opinion of 29 September 2026 (No. 25-2153), the District of Delaware's memorandum opinion of 11 February 2025 (No. 1:20-cv-613-SB), the orders in Bartz v. Anthropic (23 June 2025) and Kadrey v. Meta (25 June 2025), the texts of FSEK, the TTK and Bill No. 2/3634, Korean Supreme Court judgment 2021다272001 and the Korean Copyright Act, as available on the date of writing. The description of the Department of Justice's Statement of Interest is taken from the Third Circuit's footnote 7. The opinion was first issued under seal and later made public; this analysis uses the public version on the court's website. ROSS's stated intention to seek Supreme Court review, the date of final approval of the Bartz settlement, the Concord hearing date and the Korean policy developments come from press reports and may change. Statements about Turkish, EU, Korean and United States law are general in nature; specific cases require individual assessment. The analysis and assessments are the author's own.