A US Appeals Court Ruled Against ROSS Intelligence on 29 September 2026, Delhi Refused ANI an Injunction Against OpenAI on 24 July and Munich Ruled Against Suno on 31 July, and None of the Three Settles Whether Training a Chatbot on Copyrighted Text Is Lawful

On 29 September 2026 a three-judge panel of the US Court of Appeals for the Third Circuit ruled that ROSS Intelligence, a legal-research startup, was not making fair use when it copied Westlaw’s editorial headnotes to train an AI search tool. In July two other courts had ruled on AI training: the Delhi High Court on 24 July, refusing the news agency ANI an injunction against OpenAI, and a Munich court on 31 July, ruling for the German music society GEMA against the AI music maker Suno. Each court decided its own facts under its own country’s law, and none of the three says whether training a chatbot on copyrighted text is lawful in general.

The Third Circuit ruling on 29 September

ROSS built a search engine that answered plain-language legal questions with passages from court opinions. To train it, a contractor called LegalEase wrote about 25,000 legal memos, and its writers used thousands of Westlaw headnotes, the short summaries of points of law that Thomson Reuters’s editors write, to frame the questions. The district court in Delaware, with Judge Stephanos Bibas of the Third Circuit sitting by designation, granted Thomson Reuters partial summary judgment on 2,243 headnotes. At the parties’ request it then sent two questions up on an interlocutory appeal, meaning an appeal taken before the case is finished: whether the headnotes are original enough to be protected, and whether ROSS’s use was fair.

Judges Restrepo, Montgomery-Reeves and Bove heard the argument on 11 June 2026. Judge Montgomery-Reeves wrote the opinion, which affirms on both questions. LawNext reported it was first filed under seal, and Jones Walker says it was public on 30 September. The headnotes have “some creative spark,” the panel held, because editors chose which points of law to include and how to word them.

On fair use, the panel weighed the four statutory factors and found that three of them went against ROSS. The use was commercial, ROSS charged prices comparable to Westlaw’s, and the panel called it “minimally transformative, at best,” because both companies used the headnotes to build a legal-research platform. ROSS had argued that it took only 0.08% of Thomson Reuters’s 28 million headnotes. The panel answered that the court opinions behind the headnotes were free to copy, so copying the headnotes was not necessary: “Unlike necessity, ease is not a justification for copying.” The one factor that went ROSS’s way, by a slight margin, was the nature of the work, because headnotes are more factual than fictional.

On market harm, the panel counted two markets: the market for legal-research platforms, where ROSS aimed to replace Westlaw, and a possible market for licensing headnotes as AI training data. It found on the evidence that this licensing market is “rapidly developing,” and that ROSS had taken Thomson Reuters’s chance to enter it. That Thomson Reuters had not yet licensed its headnotes to anyone did not disprove the market, the panel said.

What the panel set aside

The opinion says ROSS’s tool was not a generative AI, “meaning it would not create any new expression.” A footnote addresses the US Department of Justice’s statement of interest filed on 1 September 2026 in the New York litigation against OpenAI (No. 1:25-md-3143). The Justice Department argued there, relying on the 2025 California decision in Bartz v. Anthropic, that training a large language model, which can generate original responses, is transformative. The panel wrote that “The concerns raised in that separate case do not apply here,” and added that the department had not asserted its interests in the ROSS appeal. It described its own case as “no more than an ordinary copyright case.”

The panel also did not decide whether headnotes that quote an opinion word for word are copyrightable, or rule on Westlaw’s Key Number System, which ROSS never challenged in its briefs.

The 2025 district rulings in Bartz v. Anthropic and Kadrey v. Meta, both from the Northern District of California, are the decisions the panel set its own case apart from. The Delhi judgment records that Bartz separated books a company had bought lawfully from copies taken from shadow libraries, which are sites that host pirated books. Womble Bond Dickinson’s 7 October analysis headlines the ruling as the first federal appellate fair use decision on AI training, and Ballard Spahr’s 2 October alert says it leaves generative AI questions unresolved.

Munich on 31 July: GEMA against Suno

The 42nd civil chamber of the Munich I Regional Court granted GEMA’s claims for an injunction, disclosure of Suno’s revenue and damages largely, according to the court’s own press release (Az. 42 O 763/25). The case covered six songs, including “Rasputin,” “Big in Japan” and “Forever Young.” Song lyrics were not part of it.

The court made three findings. Copies of the works made during training in the United States were not covered by US fair use, it held, and it distinguished Bartz and Kadrey because simple, open-ended prompts produced outputs substantially similar to the originals. The works were also reproducibly contained in Suno’s model versions 3.5 and 4, stored on servers in Germany, and the German text-and-data-mining exception did not cover that storage. Suno, not its users, was responsible for the outputs, the court held.

The judgment is not final. The release gives no damages figure, and GEMA’s own page, dated 31 July, says OpenAI has appealed the Munich court’s November 2025 ruling against it to the Higher Regional Court in Munich. Legal 500 describes that earlier case as one about song lyrics memorised by OpenAI’s models. Simmons & Simmons notes that the German judgment leaves the EU-wide text-and-data-mining question to the pending Court of Justice case below.

Where the EU’s top court stands

That case is Like Company v Google Ireland (C-250/25). A Hungarian court, the Budapest Környéki Törvényszék, referred it, and the Court of Justice heard it on 10 March 2026 before its 15-judge Grand Chamber. The EU’s IP Helpdesk lists the questions: whether a chatbot that displays content partly identical to a protected press publication communicates it to the public, whether training a large language model is a reproduction, and whether the text-and-data-mining exception in Article 4 of the 2019 copyright directive covers either act. The same page said an Advocate General’s opinion was expected on 3 September 2026. We did not find a published opinion or a judgment, so neither is reported here.

Delhi on 24 July: ANI against OpenAI

Justice Amit Bansal’s 135-page judgment, reserved on 27 March 2026 and pronounced on 24 July, dismissed ANI’s application for an interim injunction (I.A. 45300/2024 in CS(COMM) 1028/2024). Every finding in it is prima facie, meaning a first view taken only to decide that application. ANI won one point: the court kept jurisdiction even though OpenAI’s servers are in the United States, because the training claim “cannot be entirely divorced” from the output claim, and the outputs are produced within the court’s jurisdiction.

On training, the judgment accepts that the reproduction right in section 14 of the Copyright Act, 1957 covers storing a work by electronic means, and that OpenAI stored ANI’s articles at least temporarily. It then holds that this storage falls within section 52(1)(a), the fair dealing clause for “private or personal use, including research,” and so is not infringement under section 51. The court recorded that the training cut-off dates were April 2022 for GPT-4 and April 2024 for GPT-4o, earlier than the articles ANI had named, and said there was “no factual foundation in the plaint” for the training claim as pleaded. It ruled on the storage question anyway because storage was admitted.

On outputs, the judgment says the answers ChatGPT gave using retrieval, which pulls information from outside sources before the model replies, were not substantially similar to ANI’s articles, and that ANI “failed to satisfy” the court that memorisation or regurgitation had happened. On the balance of convenience, it notes that ANI could block crawlers and had not, that OpenAI said it had blocked ANI’s site, and that ANI’s own licence offer of USD 7.5 million, dated 3 October 2024, shows the loss can be counted in money. The judge also wrote that licences from many sources would make building a large language model economically unviable, and cited a NITI Aayog paper of October 2025 on the benefits of AI. Paragraph 5 of the judgment notes that Sarvam AI launched India’s first domestic large language model on 18 February 2026, and says India is considered a forerunner in the field. Tuput’s reports on Indian-language AI and the IndiaAI Mission cover the domestic side.

What the Delhi judgment leaves for trial

Paragraph 274 says the observations are made only to decide the application and “would have no bearing on the final outcome of the suit.” The judgment does not test a dataset built from stolen or paywalled copies: paragraph 200 records that ANI did not allege OpenAI took its works from unauthorised sources or by breaking a paywall, the fact on which Bartz had drawn its line. The other open question is memorisation. The judgment notes that ANI cited the Munich court’s November 2025 decision, which found song lyrics memorised in OpenAI’s models, but the Delhi court found no proof of memorisation on the record before it. A correction order of 27 July 2026 changed one case name in paragraph 80 from “Bartz v. Anthropic” to “Richard Kadrey v. Meta Platforms.”