29 September, Philadelphia
On 29 September the US Court of Appeals for the Third Circuit, which sits in Philadelphia, decided the first appeal in the United States on the use of protected content to train artificial intelligence.1 Thomson Reuters, owner of the Westlaw legal database, accused ROSS Intelligence of copying its editorial notes to build a competing product. The three judges sided with Thomson Reuters: the notes are protected by copyright, and the way ROSS used them is not fair use, the exception that in the United States allows other people's works to be used without permission.
The notes are called headnotes: short summaries, written by Westlaw's editors, of the legal point in a judgment. ROSS wanted a search engine that would answer legal questions asked in plain language with passages taken from around ten million public judgments. To train it, it had another company, LegalEase, write about 25,000 memos, each with a question and four to six passages from judgments ranked by relevance. The questions started from Westlaw's headnotes, because they offered an easy way to frame them.1
ROSS made no secret of its aim. Its ads compared it with Westlaw, at prices in line with Westlaw's, and some law firms switched from one to the other.1 It shut down in 2021, citing the cost of the lawsuit, and has said it will ask for a rehearing and, in the end, try to reach the Supreme Court.2
Why ROSS lost
Fair use is decided by weighing four factors, and the court found three of them against ROSS.1
The first concerns purpose. ROSS argued that training an AI is a different activity from reading a headnote, and so gives the material a new purpose. The judges accepted there was a difference, but a small one: both companies used the headnotes to build a legal research tool, and ROSS did so to make money. A commercial use that, at best, changed very little.
The second factor, the nature of the work, leans slightly towards ROSS, because headnotes summarise facts and rules. The third, the amount copied, does not. ROSS took the headnotes in full and defended itself by saying they were 0.08% of Westlaw's 28 million. The court replied that copying them was not necessary, because the original judgments are public and free, and that convenience does not justify copying. Judge Stephanos Bibas, who decided the case at first instance in February 2025, had reached the same conclusion and asked the appeals court to rule.2
The fourth factor, the effect on the market, is the one that will be argued over longest. The court found two kinds of harm. The first is obvious: a competitor built with Westlaw's headnotes takes customers from Westlaw. The second is new. According to the judges there is a market for licensing headnotes as text to train AI, which they describe as developing fast, and Thomson Reuters already uses its own headnotes to train its products. The fact that it has never licensed them to anyone else does not prove the market does not exist. By copying them without permission, ROSS took away its chance to enter it.1
The public interest argument did not persuade the judges either: the judgments are already free and ROSS charged as much as Westlaw.1
Not every commentator agrees with the court. The lawyer Aaron Moss notes that the licensing argument risks going round in circles: any rights holder can say it would have charged for the use in dispute, and then every use becomes a lost market. In his view the court offers little evidence of that market, and the real question remains whether training an AI gives the work a new purpose.3
What the ruling leaves open
The court drew a clear line. ROSS was not generative AI: it returned passages from existing judgments without producing new text. In a footnote the judges note that on 1 September the Department of Justice intervened in the case against OpenAI in New York to argue that training a large language model is a use that adds something new, and they make clear that those arguments do not apply here.1 The law firm Ballard Spahr sums it up: the ruling does not say whether training a generative model on books or other works is fair use.4
On that question the lower courts have split. In June 2025 in San Francisco, Judge William Alsup found that training Claude on books changed their purpose so much that it was lawful, but ruled out Anthropic keeping a library of copies downloaded from pirate sites such as Library Genesis.5 Anthropic chose not to go to trial and agreed to pay 1.5 billion dollars, about 3,000 for each work. The settlement received final approval on 20 July 2026.6
A few days later, in the same court, Judge Vince Chhabria found for Meta with different reasoning. The authors lost because they had not proved the harm he considers most important, market dilution: the risk that text generated in huge quantities competes with the books used to train the model. Chhabria added that, in most cases, copying protected works to train generative AI will probably be illegal.7
The most closely watched case is the New York Times against OpenAI and Microsoft. On 4 September both sides asked Judge Sidney Stein to decide without a trial, and a ruling is expected in the coming months.8 The US Copyright Office, in its May 2025 report, had already set out where fair use ends: commercial use of vast amounts of works to produce content that competes with them, especially when they were obtained illegally, goes beyond it. It also recommended letting the licensing market grow without government intervention.9
Europe starts from a different rule
Europe has no fair use. Since 2019 the directive on copyright in the digital single market has allowed text and data mining of works to which one has lawful access, unless the rights holder has reserved their rights. For content published online, that reservation has to be expressed in an appropriate way, for example by machine-readable means.10 This is the opt-out system: everything can be used until someone says no.
The AI Act linked this rule to AI models. Since 2 August 2025, anyone developing a general-purpose model must have a policy for complying with EU copyright law and publish a summary of its training data. The code of practice published on 10 July 2025 asks signatories to respect robots.txt and other machine-readable opt-outs, to stay away from pirate sites and to give rights holders a point of contact for complaints. Since 2 August 2026 the Commission has been able to enforce these obligations, including with fines.11
European courts have begun to rule. On 11 November 2025 the Munich Regional Court found for GEMA, the German music rights society, against OpenAI: if a model has memorised song lyrics and returns them in response to simple prompts, that memorisation is a reproduction, and the text and data mining exception does not cover it. The ruling is not final.12 A week earlier in London, the High Court had largely found against Getty Images in its case against Stability AI. Getty had dropped its claims about training, which took place outside the UK, and the judge held that the finished model does not contain copies of the photos.13
Italy has gone further. Law 132 of 2025, in force since 10 October 2025, added article 70-septies to the copyright law: reproductions and extractions made through AI models and systems, including generative ones, are allowed only within the exceptions already provided for research and for mining with an opt-out. The same law made it a criminal offence, punishable with a fine of 51 to 2,065 euros, to extract text or data in breach of those rules, including through artificial intelligence systems.14
How the judges have ruled
| Case | Court and date | What was used | Outcome |
|---|---|---|---|
| Thomson Reuters v. ROSS | Court of Appeals for the Third Circuit, September 2026 | Westlaw editorial headnotes for a competing search engine | Not fair use, partly because of harm to the licensing market |
| Bartz v. Anthropic | Federal court in San Francisco, June 2025 | Purchased books and books downloaded from pirate sites | Training lawful, pirate library not; 1.5 billion dollar settlement |
| Kadrey v. Meta | Federal court in San Francisco, June 2025 | Books used to train Meta's models | Meta wins, because the authors did not prove market harm |
| GEMA v. OpenAI | Munich Regional Court, November 2025 | Song lyrics memorised by the model | Infringement: the data mining exception does not cover memorisation |
| Getty Images v. Stability AI | High Court in London, November 2025 | Getty photos used to train Stable Diffusion | The model contains no copies; training had taken place abroad |
What content is worth
Meanwhile the market the Philadelphia court recognised has already taken shape. In May 2024 News Corp, publisher of the Wall Street Journal, signed a deal with OpenAI that the Journal itself said could be worth about 250 million dollars over five years.15 Reddit, in its stock market listing prospectus, declared data licensing contracts worth 203 million dollars, running two to three years.16 Amazon pays the New York Times between 20 and 25 million dollars a year for articles, recipes and sports coverage, which it can also use to train its models.17 In Italy the best known agreement is still the one GEDI signed with OpenAI in September 2024, which includes using its titles' archives for training as well.18 What this means for newspaper traffic I covered in The news comes from the chatbot, the clicks never reach the newspaper.
The buyers, though, are not always the ones you would expect. Of the 230 public agreements between generative AI companies and media that Reporters Without Borders counted between 2023 and May 2026, half were signed by a single company, ProRata. OpenAI signed 20, Google 14.19
Who signs the deals with the media
For books the Italian picture is thinner still. In a survey of 97 publishers by the Italian Publishers Association, presented in December 2025, 27.7% had been approached by developers of language models about a licence. Of those, 3.7% had signed at least one contract, 37% had refused and 59.3% were still weighing it up.20
What I think
American judgments are public and free. What Westlaw sells, and what the court protected, is the work done on top: working out which principle in a judgment matters and summarising it in a few lines. It is the same difference as between a good article and a press release. When a model can read everything almost for free, value shifts to what takes effort to produce, namely checking and deciding what matters.
Nobody invents alone.
Eighth law of innovation, from Un perché alla volta (translated)
That is why I see the licensing market as good news. It puts a price on something that until now was simply taken, and replaces the lottery of lawsuits with contracts that can be negotiated. But it works on two conditions.
The first is that it pays the right people. Today it rewards those with the strength to negotiate, meaning the big groups, almost always English-speaking. Reddit's contract was signed by Reddit, and the people who wrote the posts are not part of it. In the AIE survey, the book publishers that had signed anything could be counted on one hand. The Anthropic settlement goes the other way: the money is calculated work by work and split between authors and publishers according to their contracts or splits agreed in advance.6 It is an imperfect model, but it reaches the people who wrote.
The second is that the market exists in practice and not only in legal briefs, and here Moss's objection is a serious one. What is needed is prices that can be known, at least by category, and contracts that measure use.
As for the question left open, I think the line between generative and non-generative AI will matter less than the question the court actually asked: does the trained product take the place of the original? ROSS wanted to replace Westlaw. An assistant that answers with the content of an article, so that nobody opens the article any more, does something similar, even when it rewrites it in its own words.
What to do now
- PublishersMake the no readableState the reservation of rights in machine-readable form, in robots.txt and in metadata. That is what the EU directive asks for and what the code of practice commits signatories to respect.
- Small publishersNegotiate togetherOn your own you sign little. Collective management already exists for music and books, and the GEMA case shows a collecting society can take an AI company to court and win, at least at first instance.
- Companies and brandsDecide what to openProduct pages and guides are worth letting models read. Research, proprietary data and creative archives only under contract. And for generated content used in campaigns, ask the supplier where the training data came from.
- Authors and creatorsRead the contractsA clause on using texts for AI is now worth money. Before signing a publishing or commissioning contract, it pays to know who controls it and how the fee is split.
The first step takes a few minutes. Anyone who publishes online can open their site's robots.txt file, found at the domain address followed by /robots.txt, and check which AI crawlers are allowed, which are blocked and which are not mentioned at all. For models that follow the European code, that file is what tells them what they can take.