Training data and copyright

The training copy, the weights and the output

Training an AI model starts with copies: a developer collects works, such as articles, photographs or song lyrics, from the web or from a source that has licensed them, stores them in a collection called a training dataset and passes them through the model during training. Each of those copies is a reproduction in the ordinary sense of copyright law, so the law asks whether a licence or an exception permits it. In the European Union and the United Kingdom the main exceptions in issue are those for text and data mining, the automated analysis of works in digital form to find information such as patterns, trends and correlations, alongside a narrow exception for temporary copies made as part of a technical process.

A trained model is, at bottom, a very large set of numbers called weights, which the model applies to its input to calculate its output. Training adjusts those millions or billions of numbers a little at a time, example after example, until the model’s outputs resemble the material it was trained on. The finished model is shipped as a file of those numbers without the dataset, so a photograph or a lyric is not kept in it as a picture or as text. Copyright law then asks about the weights themselves: whether a work is embodied in them in a form from which it can be perceived or reproduced. The output, meaning what the model produces when someone uses it, raises the familiar question whether a particular answer, image or song reproduces a protected work.

Training can, however, set the numbers so closely to one work, often because the work appeared many times in the training data, that the model can reproduce it, or long passages of it, closely enough to recreate the original. That behaviour is called memorisation, explained in Sampling, hallucination and memorisation, an Intermediate module, and it is what links the weights to a particular work. The leading European judgments came to different results on records that differed on memorisation. In Getty Images v Stability AI the photo agency Getty Images did not claim that the model’s weights stored its photographs. In GEMA v OpenAI, by contrast, the Munich Regional Court found song lyrics reproducibly contained in the models, meaning held in a form from which the models could produce them again, and held that memorisation embodied them in the models’ parameters, another name for their weights.

Collecting, training and use can each happen in a different country, which bears on whose law applies, since copyright is territorial. Under section 16(1) of the Copyright, Designs and Patents Act 1988, the owner of a UK copyright has the exclusive right to do the acts that copyright restricts “in the United Kingdom”. In the European Union, under Article 8(1) of the Rome II Regulation, the law that applies to the infringement of an intellectual property right is “the law of the country for which protection is claimed”. Each country’s copyright covers acts done in that country, so a claim over a copy usually has to be brought under the copyright of the country where the copy was made. The place where a developer trained its model therefore helps decide which country’s exceptions apply to the training copies, as the London and Munich cases described below both show.

United Kingdom

The UK’s exception for text and data mining is narrow. Section 29A of the Copyright, Designs and Patents Act 1988 permits a person with lawful access to a work to copy it so that a computational analysis can be made of what the work records. That computational analysis has to be “for the sole purpose of research for a non-commercial purpose”, and the copy has to be accompanied by a sufficient acknowledgement, unless that is impossible. Any copy made under the section infringes copyright if it is transferred to another person, or used for another purpose, without the copyright owner’s authority, which bears on datasets shared with developers. A contract term that tries to prevent copying for that purpose is unenforceable under the same section. Training a model for a commercial product, by contrast, appears to fall outside the exception on its terms, since its purpose is commercial.

Section 28A of the same Act adds a narrow exception for temporary copies that are transient or incidental and form an integral and essential part of a technological process. Each such copy has to serve only to enable a lawful use of the work, or its transmission in a network between third parties by an intermediary, and lack independent economic significance. The government’s report of 18 March 2026 notes that some developers have cited both exceptions, “though these contain a number of conditions and will only be applicable when those conditions are met”, so a developer that does not meet them needs a licence, or an exception in another country.

That report on copyright and AI, which section 136 of the Data (Use and Access) Act 2025 required the government to prepare, observed that “Often models will be trained in jurisdictions with the most flexible copyright laws”. In its consultation, which ran from December 2024 to February 2025, the government had preferred a broad exception for data mining that a rightholder could opt out of through machine-readable means, similar to the exception in EU law. Machine-readable means are, in ordinary usage, signals written for software to read, such as a notice in a website’s files telling automated collection tools not to copy its pages. Most respondents rejected that option, and the report recorded that “a broad copyright exception with opt-out is no longer the government’s preferred way forward”. The report proposed to gather further evidence and to consider other approaches, and said that the current law would continue to apply in the absence of new legislation.

Output raises a separate question in UK law, about who owns a work that a computer made. Under section 9(3) of the 1988 Act, the author of a literary, dramatic, musical or artistic work that is computer-generated is “the person by whom the arrangements necessary for the creation of the work are undertaken”. Section 178 defines a computer-generated work as one generated by computer in circumstances in which the work does not have a human author. In its report of March 2026 the government proposed removing that protection for wholly computer-generated works, “in the absence of evidence of its ongoing value”, while copyright would continue to protect works made with the assistance of AI.

The leading UK case on generative AI and copyright is Getty Images v Stability AI, decided by Joanna Smith J in the High Court on 4 November 2025. Getty Images and companies in its group sued Stability AI over Stable Diffusion, an image generator whose later versions Stability published for download. Stable Diffusion’s first versions were released by a research group at universities in Munich and Heidelberg; Getty argued that Stability was responsible for that release too, and the judge held that Stability was not directly liable for it. Every version of Stable Diffusion was trained on parts of LAION-5B, a collection of about 5.85 billion links to images on the web paired with their captions, and training required downloading the linked images. Stability accepted that at least some images from Getty’s websites were used in training. Every image on those websites shows a Getty Images or iStock watermark, and a version without the watermark is available only to a customer who takes a licence, which became the basis of Getty’s claim over trade marks described below.

Getty’s pleaded case covered copyright in training and in the model’s output, database right, secondary infringement of copyright, trade marks and passing off. Shortly before closing submissions Getty abandoned its claim over training, accepting that the evidence did not show any training of Stable Diffusion in the United Kingdom. Getty also gave up its claim over output since, as it acknowledged, Stability’s blocking of the prompts used to produce Getty’s examples of infringing images had substantially achieved the relief Getty would have been entitled to on that claim. Getty accepted that its claim over its database, tied to the claims over training and output, could not then be advanced. Stability’s case throughout was that training took place on Amazon’s cloud servers outside the United Kingdom, but the judge made it clear that she did not need to decide where training took place.

What remained of the copyright case was secondary infringement, the wrong committed by someone who imports or deals in an infringing copy of a work. Sections 22 and 23 of the 1988 Act make it an infringement to import into the United Kingdom, possess in the course of business or distribute an article that is, and that the defendant knows or has reason to believe is, an “infringing copy”. Under section 27(3), an imported article is an infringing copy if making it in the United Kingdom would have infringed copyright. Getty did not say that the model stored copies of its works; its argument was that making the weights, a process that involved copying its works during training, would have infringed copyright had it happened in the United Kingdom, so the model imported into the country was an infringing copy.

Joanna Smith J accepted Getty’s argument that an “article” can be intangible, holding that an electronic copy stored in an intangible medium such as cloud storage is capable of being an infringing copy. The judge rejected Getty’s argument that the model was an infringing copy, on the reasoning that an infringing copy has to be a copy, and that an article which has not contained or stored a copy of the work at any stage does not qualify. The weights of Stable Diffusion were altered by exposure to Getty’s works during training, but in her words they are “purely the product of the patterns and features which they have learnt over time during the training process”. The judge recorded that Getty did not claim that the weights reproduced its works and that the evidence did not show any of its works to have been memorised, so the claim of secondary infringement failed.

Getty did better on its claim over trade marks, though only in part. Users of early versions of Stable Diffusion, working through Stability’s own online services, had generated images bearing iStock watermarks. Users of the version 2 models had generated images with Getty Images watermarks, and the judge held that both kinds of watermark infringed Getty’s registered marks. The claims failed for Stability’s later models, the Stable Diffusion XL series and version 1.6, since the evidence did not show a single user in the United Kingdom generating a watermark with them. Version 1.6, despite its number, came out after version 2 and was built on the design of the XL series. The claim under section 10(3) of the Trade Marks Act 1994, which protects marks with a reputation, failed for every version. In her words, the findings in Getty’s favour were “historic and extremely limited in scope”.

On 16 December 2025 the judge granted Getty permission to appeal on the infringing copy point, calling it “a pure question of law” of statutory construction which had not been considered by any court before, while refusing Stability permission to appeal on trade marks. The claims Getty abandoned were dismissed, with words in the order recording that they had been fought through the whole trial until closing submissions and then abandoned, so that the court could not later be said to have decided them. In a separate judgment on costs the judge held Stability to be the overall winner of the litigation, while awarding Getty 75 per cent of its costs of the trade mark claim. The Court of Appeal had not published a judgment on the appeal by 29 October 2026, according to Find Case Law, the official database of judgments in England and Wales.

For practice, whether training or output infringes copyright in the United Kingdom remains open after the judgment, since Getty abandoned its claims on both, and the place where training happened largely set the scope of the case. On the High Court’s reading of section 27, a model is not an infringing copy unless its weights store, or at some stage stored, a copy of the work, and the Court of Appeal may read the section differently. Evidence that a model has memorised particular works may therefore become central to later claims in the United Kingdom.

European Union

EU copyright law has two exceptions for text and data mining, both in the Digital Single Market Directive, Directive (EU) 2019/790. Article 2(2) defines text and data mining as “any automated analytical technique aimed at analysing text and data in digital form in order to generate information which includes but is not limited to patterns, trends and correlations”. Article 3 requires Member States to let research organisations and cultural heritage institutions make reproductions of works to which they have lawful access, in order to perform text and data mining for the purposes of scientific research. Article 4 requires a wider exception, open to any user and not limited to research, for reproductions of lawfully accessible works made for text and data mining, which may be kept for as long as that purpose requires.

The Article 4 exception applies on condition that the rightholder has not expressly reserved the use of its works “in an appropriate manner, such as machine-readable means in the case of content made publicly available online”. Recital 18 adds that, for content made publicly available online, a reservation should only be considered appropriate if made by machine-readable means, “including metadata and terms and conditions of a website or a service”. Metadata is information attached to a file or a web page that describes it, such as who made it and on what terms it may be used. Training a generative model analyses works for patterns, which appears to fit the definition in Article 2(2). The Munich court accepted that language models, programs that write text by predicting one word after another, in principle come within the scope of the exceptions, which in its view cover the copies needed to assemble a training corpus. Training also aims at a model that produces new material, and a model that memorises works may keep more than information about them, a point the Munich court relied on, as described below. Whether training a generative model counts as text and data mining at all, and what makes a reservation machine-readable, are questions that German courts have begun to address and that the Court of Justice has been asked about in part.

The EU’s AI Act adds duties for the providers of general-purpose AI models, models able to perform a wide range of distinct tasks, such as the models behind chat assistants. Under Article 3(3), a provider is any person or body, commercial or not, that develops such a model, or has one developed, and places it on the market under its own name or trade mark, whether for payment or free of charge. Article 53(1)(c) requires a provider to put in place a policy to comply with Union law on copyright and related rights. The policy has in particular to identify and comply with a reservation of rights expressed under Article 4(3) of the Digital Single Market Directive, “including through state-of-the-art technologies”. Recital 106 says that every provider placing such a model on the EU market should meet that duty “regardless of the jurisdiction in which the copyright-relevant acts underpinning the training of those general-purpose AI models take place”. On the recital’s reading, training outside the EU therefore does not excuse a provider from respecting EU opt-outs.

Under Article 53(1)(d) the provider also has to publish a sufficiently detailed summary of the content used to train the model, following a template from the AI Office, the part of the European Commission that supervises these models. The duties in Article 53 have applied since 2 August 2025 under Article 113, although providers of models placed on the market before that date have until 2 August 2027 to comply under Article 111(3). Article 101 lets the Commission fine a provider of a general-purpose AI model up to 3 per cent of its worldwide annual turnover in the preceding financial year or 15 million euros, whichever is higher, where the Commission finds that the provider infringed the AI Act intentionally or negligently. Article 113, in listing the provisions that have applied since 2 August 2025, expressly excepts Article 101, so the Commission’s power to fine appears to run from the AI Act’s general date of application, 2 August 2026.

The Commission published its template for the training summary on 24 July 2025. According to its explanatory notice, a provider lists the main datasets used to train the model and describes the data scraped from the web, meaning collected from websites automatically by software. Its summary names the internet domains in the top 10 per cent by the amount of content scraped. Small and medium-sized businesses may give a shorter list of domains. The provider also describes the measures it took to respect reservations under Article 4(3), including the opt-out protocols it honoured, which are the standard signals by which a website tells collecting software not to copy its pages. The notice also recommends that providers tell rightholders, on request and voluntarily, whether content from other domains was used.

Article 53(4) lets a provider rely on a code of practice to show that it complies, and the Commission and the AI Board have confirmed that the General-Purpose AI Code of Practice, published on 10 July 2025, is an adequate voluntary tool for that purpose. Its copyright chapter sets out what the signatories commit to. Their web crawlers, the programs that collect pages from the web, are to read and follow robots.txt, a file on a website that tells crawlers which pages they may collect, and to comply with other appropriate machine-readable protocols for reserving rights. Signatories also commit not to get round paywalls or similar technical protections, and to leave out websites that courts or public authorities in the EU or the European Economic Area have recognised as persistently infringing copyright on a commercial scale. Each signatory further commits to put in place appropriate and proportionate technical safeguards to prevent its models from generating outputs that reproduce protected training content in an infringing manner, and to give rightholders a point of contact and a way to lodge complaints. The Commission’s list of signatories, read on 29 October 2026, includes Amazon, Anthropic, Google, Microsoft, Mistral AI and OpenAI; xAI has signed only the chapter on safety and security, and Meta does not appear on the list.

The German cases

German copyright law has its own exceptions for text and data mining. Section 44b of the Copyright Act permits copies of lawfully accessible works for text and data mining generally, and treats a rightholder’s reservation for content online as effective only if it is machine-readable. Section 60d permits such copies for scientific research by research organisations, cultural heritage institutions and individual researchers working for non-commercial purposes. Three judgments of German courts have tested those provisions against AI, and the judgments in all three cases are not yet final. The first concerned a dataset built to train image generators, and the other two concerned models that reproduced song lyrics and music.

Kneschke v LAION was brought by Robert Kneschke, a photographer, against LAION, a German non-profit association that publishes, free of charge, a dataset of 5.85 billion links to images on the web paired with descriptions of the images, which developers can use to train generative models. To build the dataset in the second half of 2021, LAION took an existing collection of image addresses and descriptions, downloaded each image, checked with software that the description matched the image and kept only the links and descriptions that passed. One of the images was a preview of a photograph by Mr Kneschke on a photo agency’s website, which forbade access by “automated programs, applets, bots or the like”, and he sued to stop LAION copying it.

The Hamburg Regional Court dismissed the claim in September 2024, and on 10 December 2025 the Hanseatic Higher Regional Court in Hamburg dismissed his appeal. According to the court’s press release and the Federal Court of Justice’s summary of the appeal judgment, the download was a reproduction, but section 44b permitted it. Checking whether an image matched its description was itself an automated analysis that produced information, so the appeal court did not need to decide whether training a generative model also counts as text and data mining. The agency’s reservation counted as the photographer’s, but section 44b requires a reservation for content online to be machine-readable. The Digital Single Market Directive and the German Copyright Act do not define machine-readable, and in ordinary usage the term describes a notice that software collecting content can read and act on. The appeal court asked whether the notice was machine-readable as at the time of the download, and what software can read has changed over time. The photographer had not shown that the agency’s notice, written in ordinary language, was machine-readable in the second half of 2021, when LAION made the download.

Section 60d, the research exception, also covered the download, the Higher Regional Court held, agreeing with the Regional Court, since LAION is a research organisation and building the dataset was a methodical step towards later knowledge. In the appeal court’s view, that commercial developers could also use the dataset did not change LAION’s status. LAION’s open-source approach makes the results of its work available to everybody, commercial developers included, and the court found that a private company did not have decisive influence over it. The judges also found the use consistent with the three-step test in Article 5(5) of the Information Society Directive, the EU’s main copyright directive of 2001, which allows an exception only in certain special cases that do not conflict with a normal exploitation of the work and do not unreasonably prejudice the rightholder’s legitimate interests.

Permission for a further appeal was given by the Higher Regional Court, and the Federal Court of Justice, Germany’s highest court in civil matters, heard the appeal on 3 September 2026 and has set 17 December 2026 as the date for pronouncing its decision. Under Article 267 of the Treaty on the Functioning of the European Union, a court whose decisions are not open to further challenge under national law is in principle required to refer a question of EU law to the Court of Justice where a decision on it is necessary to give judgment. Under the Court of Justice’s established case law, restated in Consorzio Italian Management, such a court is relieved of that duty where the question is irrelevant, where the Court of Justice has already interpreted the provision, or where the correct interpretation is so obvious that reasonable doubt is excluded. The Federal Court of Justice may therefore refer questions about Article 4 to the Court of Justice, or decide the appeal itself if it finds that one of those grounds applies. In the meantime, a rightholder who relies on a notice written in ordinary language may need evidence of what software was able to read at the time of the copying, since the appeal court judged machine-readability as at that time, while opt-outs expressed through robots.txt or similar protocols are likely to be on firmer ground.

Song lyrics were at stake in GEMA v OpenAI, brought by GEMA, the society that collects royalties in Germany for composers, lyricists and music publishers, against two companies in the OpenAI group. The claim concerned song lyrics by nine well-known German songwriters, among them “Atemlos” by Kristina Bach and “Wie schön, dass du geboren bist” by Rolf Zuckowski. OpenAI’s chatbot runs on language models, which write text by predicting one word after another, as How a chatbot writes an answer, a Beginner module, explains. GEMA’s case was that the lyrics had been memorised by OpenAI’s language models and came out, largely word for word, in answers to simple prompts. OpenAI argued that its models do not store or copy particular training data, that users who write the prompts are responsible for the output, and that any copying was covered by the exception for text and data mining.

On 11 November 2025 the Munich Regional Court largely upheld GEMA’s claims for an injunction, information and damages, according to its press release. Comparing the lyrics in the training data with the chatbot’s answers, the court found the lyrics reproducibly contained in two of OpenAI’s models, versions 4 and 4o, and given the length and complexity of the lyrics it ruled out chance as the explanation. Memorisation, in the court’s view, embodies the lyrics in the model’s parameters, which is a reproduction under German law and under Article 2 of the Information Society Directive, a provision covering reproduction by any means and in any form. A language model’s weights set how likely each possible next word is, so a memorised lyric is held as probabilities that make the model likely to produce it word by word. Storage as mere probability values did not alter that conclusion, the court said, citing the case law of the Court of Justice for the point that a work is reproduced when it can be perceived indirectly, with the help of technical means.

Copies made while assembling a corpus for training, such as converting a work into a digital format or holding it in a computer’s working memory, fall within the exception for text and data mining, the court accepted, since they serve later analysis and do not harm the author’s interest in exploiting the work. Training that reproduces works in the model goes beyond extracting information, in the court’s view, so the exception did not apply to it. Extending an exception by analogy means applying it to a situation its words do not cover because that situation is similar. The judges declined to extend the exception by analogy, since the risk of memorisation comes from the developer’s side. Extending the exception, which leaves authors unpaid, would in the court’s view leave the rightholder alone to bear a risk that the developer creates. OpenAI’s defences of incidental inclusion and consent also failed, the second because training a model is not a usual and expected use that a rightholder has to anticipate. The chatbot’s answers reproduced the lyrics and made them available to the public, an act that Article 3(1) of the Information Society Directive reserves to the author. That right of communication to the public covers putting a work where members of the public can access it at a time and place they choose. OpenAI was responsible for the answers, the court held, since simple prompts produced them and OpenAI operated the models trained on the lyrics and was responsible for their architecture, the overall design of the models, and for the memorisation.

GEMA lost only its claim that attributing altered lyrics to their authors infringed their personality rights, a complaint about answers that changed the words of a song while still naming its authors. The press release states that the judgment is not final, and the Munich courts had not published a further statement on the case by 29 October 2026. In practice, the judgment reasoned from the answers to the models: the answers matched the lyrics in the training data, their length and complexity ruled out chance, and the court inferred that the models contained the lyrics. A developer whose model reproduces long protected text may therefore face claims in Germany over the model itself as well as over its answers.

Music, as distinct from lyrics, was the subject of GEMA v Suno, in which GEMA sued Suno, the provider of an AI music generator, over six well-known songs, among them “Atemlos durch die Nacht”, “Daddy Cool”, “Forever Young” and “Rasputin”. Suno had copied recordings from YouTube by stream-ripping, getting round a technical protection that YouTube uses to prevent downloads, trained its models in the United States and stored them on servers in Germany. GEMA prompted the generator with each song’s original lyrics, a musical style and the title, without any instructions on melody, harmony, rhythm or arrangement.

On 31 July 2026 the same court largely upheld GEMA’s claims, according to its press release. In terms close to those of the OpenAI judgment, it found the works reproducibly contained in versions 3.5 and 4 of Suno’s models and held that storing them there was a reproduction that the exception for text and data mining did not cover. Outputs generated in Germany reproduced the works and communicated them to the public, and Suno was responsible for them, since simple and open-ended prompts produced them. Offering the model and the app was itself held to infringe the right of communication to the public. Under a German statute on collecting societies the court also took jurisdiction over the copies Suno made in the United States for training, which it held unlawful under the law of the United States as the law of the country for which protection was claimed; this site covers EU and UK law and leaves that part of the judgment aside.

Like the OpenAI judgment, the Suno judgment is not final, according to the court’s press release. For practice, the ruling suggests that training a model abroad may give limited protection to a developer whose model is stored or offered in Germany and can reproduce protected works.

The pending reference to the Court of Justice

Generative AI and copyright are now before the Court of Justice of the European Union in Like Company v Google Ireland, C-250/25, a reference for a preliminary ruling from the Budapest Környéki Törvényszék, a Hungarian court, lodged on 3 April 2025. A preliminary reference is the procedure under Article 267 of the Treaty on the Functioning of the European Union by which a national court asks the Court of Justice how EU law should be interpreted, so the answers in this case are likely to guide courts across the EU. The applicant, Like Company, is suing Google Ireland Limited, and the questions referred concern a chatbot built on a large language model and the content of press publishers’ web pages.

Article 15 of the Digital Single Market Directive requires Member States to give publishers of press publications established in a Member State rights over the online use of their publications by providers of online services. The rights in question are those of reproduction and of making available to the public set out in Articles 2 and 3(2) of the Information Society Directive, and the questions referred are framed under Article 15 and ask about those two acts. The press publishers’ rights do not cover “individual words or very short extracts”, and they expire two years after publication, counted from the following 1 January. Article 3(1) of the Information Society Directive, described above, gives authors the right over any communication to the public of their works, “including the making available to the public” of them, so making available is one form of the wider act. According to the notice of the reference published by the Court, the Hungarian court asks whether a chatbot’s answer that displays text partly identical to a publisher’s web pages, long enough to be protected, is a communication to the public. The referring court asks too whether its first question should be answered differently because the chatbot merely predicts the next word from observed patterns, the way language models write. The Hungarian court also asks whether training a chatbot built on observing and matching patterns is a reproduction, and if so, whether reproducing lawfully accessible works in that way falls within the exception for text and data mining in Article 4. Its last question asks whether, where a user’s prompt matches or refers to a press publication and the answer displays some or all of it, that display is a reproduction by the provider of the chatbot.

The Court’s case file, read on 29 October 2026, lists the case as in progress, with the request and its notice as the only documents published. The Court’s judicial calendar, read the same day, lists the case before the Grand Chamber, a formation of the Court with more judges than its ordinary chambers, with the opinion of Advocate General Szpunar scheduled for 3 November 2026. Under Article 252 of the Treaty on the Functioning of the European Union, an Advocate General’s duty is to make reasoned submissions on a case in open court, with complete impartiality and independence; the Court had not published that opinion or a judgment by 29 October 2026. The Court of Justice’s answers on whether training is a reproduction and whether Article 4 covers it would bear directly on the issues that the Hamburg and Munich courts approached differently. The Hamburg court left open whether training is text and data mining, while the Munich court held that training which reproduces works in the model is not. The Hungarian court’s questions do not mention the opt-out in Article 4(3), so whether a reservation is machine-readable may come to the Court of Justice by another route, possibly from the Federal Court of Justice in Kneschke.