October 22, 2026 · 25 min read · Copyright and AI

AI training and copyright after Thomson Reuters v. Ross: what it means for lawyers in Europe

CopyrightAI TrainingText and Data MiningLegal Databases

On 29 September 2026 the United States Court of Appeals for the Third Circuit became the first federal appeals court to decide whether copying protected material to train an AI system can be fair use. The case concerned the profession's own tools. ROSS Intelligence had trained a legal research engine with memos built from Westlaw headnotes, and the court held that its use of 2,243 of them was not fair use (1). In Europe the same question is now before the highest courts. The German Federal Court of Justice heard the appeal in the LAION dataset case on 3 September and will give its decision on 17 December 2026 (2). An appeal is pending against the Munich judgment that held OpenAI liable for song lyrics memorised by its models (3). The Court of Justice of the EU heard its first case on generative AI and copyright in March and has not yet ruled (4). Since 2 August 2026 the European AI Office has been able to enforce the AI Act's copyright duties against the providers of general-purpose AI models (5).

Law firms are involved on several sides. They use tools trained on material whose legal status is contested, subscribe to legal databases whose licences increasingly deal with automated use, and publish analysis that crawlers collect. Their clients build, buy and license models. This guide reads Ross from a European perspective, sets out where the Netherlands, Germany and France stand, and turns the rules into decisions a firm can take now.

The short answer: Ross does not apply in Europe, but its central distinction does. Legislation and court decisions are free to use: Dutch and German copyright law exclude them expressly, and France publishes its judgments as reusable open data. What publishers add, from headnotes to annotations and commentary, is protected, and so are the databases that hold it. Europe has no general fair-use defence. Copying to train AI is lawful only under a licence or within a statutory exception, chiefly the text and data mining exception, which a rightholder can switch off for online content with a machine-readable reservation. A firm should know what its tools were trained on and what its database licences allow, reserve its own published work properly if it wants to, and be ready to advise clients on the AI Act's copyright policy and training summary.

What the Third Circuit decided in Ross

ROSS set out to build a legal research engine that would answer questions put in plain language with relevant passages from judicial opinions. It was not a generative system; it retrieved passages from a bank of some ten million opinions, which are not protected by copyright in the United States. To train it, ROSS commissioned about 25,000 memos from a contractor, each pairing a legal question with passages from opinions graded by how well they answered it. The contractor's drafters used Westlaw headnotes to frame the questions because, in the words of the record, they offered "an easy way" to do so. ROSS advertised prices "in line with" Westlaw and hoped to replace it (1).

The appeal was interlocutory. Judge Stephanos Bibas, a circuit judge sitting by designation in the District of Delaware, had granted Thomson Reuters partial summary judgment in February 2025 and certified two questions for appeal: whether the headnotes were original and whether ROSS's use was fair. A panel of Judges Restrepo, Montgomery-Reeves and Bove heard argument on 11 June 2026 and affirmed on both points, in an opinion by Judge Montgomery-Reeves that describes the dispute as "no more than an ordinary copyright case" (1).

On originality, the court found the necessary "creative spark" in the editors' choices: which points of law in an opinion were important enough for a headnote, and how to state each one so that it stands alone, accurately, within a guideline of 800 characters. ROSS argued that protecting headnotes would give Thomson Reuters a monopoly over the law. The court's answer is the sentence most likely to be quoted in Europe: "Headnotes are not law; judicial opinions are." The opinions themselves remain "free for publication to all". The court rejected the merger argument because there are many ways to express a point of law, noting that a competing publisher writes different headnotes for the same opinions.

On fair use, three of the four statutory factors went against ROSS. The use was commercial and "minimally transformative, at best", because ROSS used the headnotes for the same ultimate purpose as Westlaw: a platform that helps users find relevant law. The intermediate-copying cases on software interoperability did not help, because copying there was necessary to reach unprotected functional elements, whereas ROSS had free access to the opinions themselves. "Unlike necessity, ease is not a justification for copying." Each headnote was a complete work, copied in full without need. And the copying harmed both Westlaw's market and a potential market for licensing headnotes as AI training data, which the court called "rapidly developing", noting that Thomson Reuters already uses its headnotes to train its own AI search products. Only the second factor, the largely factual nature of headnotes, favoured ROSS.

Two footnotes matter as much as the holding. In footnote 7 the court set generative AI aside. It noted a Department of Justice statement of interest, filed on 1 September 2026 in the consolidated OpenAI litigation in New York, which relied on a 2025 California district court decision to argue that training a large language model is transformative, and it cited a second California decision that explored the same question. Those concerns, the court said, do not apply here, because ROSS's platform "cannot generate original expression". Ross therefore does not decide whether training a general-purpose model is fair use. In footnote 9 the court recorded that ROSS had tried to access Westlaw with the credentials of a law firm that had invested in it, despite notice that the terms of service prohibited this; that an employee had asked about an account while posing as a solo practitioner; and that a student account had been used to reach Westlaw. To the extent good faith still counts, it counted against ROSS. That footnote is the part of the case most directly about law firms.

Why the American answer does not carry over to Europe

Fair use is an open standard: a court weighs four factors and decides whether a use is fair. European copyright law works differently. The rights of reproduction and communication are harmonised, and the exceptions to them form a closed list in EU directives, applied subject to the three-step test: certain special cases, no conflict with a normal exploitation of the work, no unreasonable prejudice to the rightholder's legitimate interests. A European court does not ask whether training an AI system is fair. It asks whether a specific exception covers the specific acts of copying, and whether its conditions are met.

For training, the relevant exceptions are the two text and data mining exceptions in the 2019 Directive on copyright in the Digital Single Market (6). Article 3 allows research organisations and cultural heritage institutions to mine works for scientific research. Article 4 allows anyone to make reproductions and extractions of lawfully accessible works for text and data mining, unless the rightholder has reserved that use. The definition is broad: any automated analytical technique aimed at analysing text and data in digital form to generate information such as patterns, trends and correlations.

On the facts of Ross, a European court would therefore have asked different questions: whether the headnotes were copied for automated analysis or reproduced as expression in new memos; whether those who copied them had lawful access to Westlaw, a question that footnote 9 makes pointed; and whether the publisher had reserved text and data mining. Competition between the AI product and its source, the decisive point in Ross, would come in only through the three-step test and its concern for the normal exploitation of the work.

There is also a difference in reach. Fair use is a defence under United States law for acts done there. The AI Act looks at the market instead: recital 106 states that any provider placing a general-purpose AI model on the EU market should comply with the copyright-policy obligation "regardless of the jurisdiction in which the copyright-relevant acts underpinning the training of those general-purpose AI models take place" (7). A model trained lawfully under fair use in California still needs a policy for respecting EU reservations if it is placed on the market in the Union.

Legislation and judgments are free; headnotes, annotations and commentary are not

The Ross court's line between law and commentary on law runs through European statutes as well, where it is written into the text. Article 11 of the Dutch Copyright Act provides that there is no copyright in laws, decrees and regulations issued by public authorities, nor in court decisions and administrative decisions (8). Section 5(1) of the German Copyright Act goes one step further: laws, regulations, official decrees and notices, decisions and officially drafted headnotes to decisions enjoy no copyright protection (9). Under German law, then, a headnote written by the court is free and one written by a publisher is not. The Munich judgment against OpenAI illustrates the difference. On the Bavarian state's public case-law portal, its headnotes are marked "redaktioneller Leitsatz", editorial headnote, because they were written by a publisher, not by the court (3).

France reaches the same practical result by another route. Article L.111-13 of the Code of Judicial Organisation requires the decisions of the ordinary civil and criminal courts to be made available to the public free of charge in electronic form, after removal of the names of the individuals involved, and subjects their reuse to the general rules on the reuse of public information (10). The same article contains a rule that matters for every legal AI product in France: the identity data of judges and court staff may not be reused for the purpose or with the effect of evaluating, analysing, comparing or predicting their actual or supposed professional practices. Breach is a criminal offence. Training a system on French case law is permitted; building a judge-by-judge prediction from it is not.

What publishers add is a different matter. Case notes, annotations, commentaries and handbooks are literary works if they meet the EU originality standard: they must be their author's own intellectual creation, reflecting free and creative choices. In Funke Medien, a case about military status reports, the Court of Justice said that purely informative documents, whose expression is dictated by the information they contain and their technical function, are not protected (11). A doctrinal commentary will normally qualify. A publisher's headnote is protected if its selection and wording reflect such choices, which is essentially how the Third Circuit reasoned. A headnote that merely repeats the court's words may not be.

Then there is the database right, which has no American equivalent. The maker of a database that required a substantial investment in obtaining, verifying or presenting its contents can prohibit the extraction or reuse of a substantial part of it, or the repeated and systematic extraction of insubstantial parts. Legal databases are the textbook case: even where the underlying decisions are free, the publisher's collection of them can be protected (12) (9) (13). The text and data mining exception extends to the database right as well, under the same conditions.

The text and data mining exception and the opt-out

Article 4 of the Directive has been implemented closely in all three jurisdictions: in article 15o of the Dutch Copyright Act and article 4a of the Dutch Databases Act, in section 44b of the German Copyright Act, and in article L.122-5-3 of the French Intellectual Property Code (8) (12) (9) (14). The structure is the same everywhere.

Lawful access. The person mining must have lawful access to the material: freely available online content, or content available under a subscription or licence. Circumventing a paywall does not qualify, and nor does logging in with someone else's account in breach of the licence.
Copies for mining only. Copies may be kept as long as necessary for the mining (the Netherlands), must be deleted when no longer needed (Germany) or destroyed at the end of the mining (France).
No reservation. The exception does not apply if the rightholder has expressly reserved text and data mining "in an appropriate manner, such as machine-readable means in the case of content made publicly available online". According to recital 18 of the Directive, for online content only machine-readable means are appropriate, "including metadata and terms and conditions of a website or a service"; in other cases a contract or a unilateral declaration will do (6).

The German courts have given "machine-readable" real content. In the LAION case, a photographer's image had been downloaded in 2021 from a photo agency's website to check whether it matched its caption, as a step in building a public dataset of image and text pairs that can be used to train generative models. The agency's website prohibited access by "automated programs, applets, bots or the like", in natural language only. The Hamburg Court of Appeal held in December 2025 that the download was covered by both the general exception and the research exception, that the mining rules are not to be read narrowly so as to exclude steps that prepare AI training, and that a reservation is machine-readable only if a machine can not merely capture the text but interpret it, so that an automated process actually leaves the reserved content alone. The photographer had not shown that the natural-language prohibition was machine-readable at the time of the download. The court put the burden of proving the absence of a reservation on the user, but the burden of proving machine readability on the rightholder (15). The Federal Court of Justice will give its decision on the photographer's further appeal on 17 December 2026 (2).

Contract matters as much as the reservation. Article 7(1) of the Directive makes contract terms contrary to the research exception, and to the exceptions for teaching and preservation, unenforceable, but does not list article 4 (6). The national laws follow. In the Netherlands the research exception in article 15n cannot be set aside by agreement, while article 15o contains no such protection (8). In Germany section 60g protects the research exception in section 60d against contractual restriction, but not section 44b (9). A licence for a legal database can therefore validly exclude text and data mining for commercial purposes. France adds a twist for databases: article L.342-3 of the Intellectual Property Code declares void any clause contrary to the mining exception for databases, but that exception applies only on the conditions of article L.122-5-3, which include the absence of an appropriate objection by the rightholder (13). A clear licence term objecting to mining is likely to count as exactly that.

Underneath all this sits an unresolved question: is training a generative model text and data mining at all? The Hamburg court treated preparatory copying for AI training as mining and left open whether the later training itself also is, because the check of each image against its caption already was (2). The Munich court accepted that the exception covers the copies made when compiling a training corpus, but held that it does not cover reproductions that end up in the model itself (3). The Court of Justice has been asked the question directly.

What European courts have said so far

Netherlands: an opt-out that named only some bots

The Dutch reference point for article 15o is a judgment that is not about AI training but is directly instructive. In October 2024 the Amsterdam District Court rejected claims by DPG Media and two Mediahuis companies against HowardsHome, a press-monitoring service that collected their online articles to produce alerts for its customers (16). The court quoted the explanatory memorandum to the Dutch implementing act: the right to read texts in principle includes the right to analyse them automatically to discover patterns, trends and correlations. The publishers relied on a prohibition of automated searching. HowardsHome argued, and the publishers did not dispute, that their robots.txt excluded only certain AI bots, such as GPTBot, ChatGPT-User, CCBot and anthropic-ai. The court held that a machine-readable reservation had therefore not been established, so article 15o applied; article 4a of the Databases Act led to the same result. It is a first-instance judgment. As far as we are aware, no Dutch court has yet ruled on the training of a generative model.

Germany: LAION and GEMA

Germany has produced the two most developed decisions in Europe so far. The LAION judgments described above concern the stage before training: building a dataset. The Munich Regional Court's judgment of 11 November 2025 in GEMA v OpenAI concerns the model itself (3). GEMA showed that two versions of OpenAI's models reproduced the lyrics of nine songs in response to simple prompts. The court held that the lyrics were memorised, and that memorisation in a model is a reproduction: it does not matter how the model stores what it has learned, because the lyrics are reproducibly contained in it and therefore embodied in it. The mining exceptions cover the copies made to compile the training corpus, but not further reproductions in the model, which do not serve only to prepare an analysis. The operators were also held responsible for infringing outputs, because a simple prompt does not shift control of the output to the user. The court granted an injunction, ordered the defendants to provide information and declared them liable in damages; the rest of the action was dismissed. The appeal is pending before the Munich Court of Appeal under reference 6 U 3662/25 e.

The Court of Justice: Like Company v Google

The first reference on generative AI and copyright comes from Hungary, in a dispute between a press publisher and Google over its chatbot. The Budapest court asked four questions (17): whether displaying, in a chatbot's answers, text partly identical to a press publication is a communication to the public, and whether it matters that the chatbot only predicts the next word; whether training a chatbot built on a large language model is a reproduction; if so, whether reproducing lawfully accessible works for that purpose falls within the text and data mining exception; and whether, when a user's prompt matches or refers to a press text and the answer displays it, the provider makes a reproduction. The Court held a hearing on 10 March 2026. Advocate General Szpunar is assigned to the case. On 22 October 2026 the Court's register lists the case as pending, and no Opinion has been published (4). Its answers to the second and third questions will bind courts throughout the Union.

France: a pending action and a bill on the burden of proof

In France the main dispute is an action brought by the national publishers' association SNE, the authors' society SGDL and the authors' and composers' union SNAC against Meta before the third chamber of the Paris Judicial Court, for infringement and economic parasitism, announced in March 2025 (18). No judgment on the merits has been reported. The more significant French development may be legislative. On 8 April 2026 the Senate adopted a bill introducing, in civil proceedings, a presumption that a protected work was used by the provider of an AI model or system whenever an indication relating to its development, its deployment or its output makes such use plausible, unless the provider proves otherwise. The Senate text would apply to pending proceedings. The National Assembly's cultural affairs committee adopted the bill on 2 June 2026; the legislative record shows no vote in plenary yet, so it is not law (19). If it is enacted as the Senate adopted it, the hardest point for a claimant, proving that a particular work was in the training data, would shift to the provider in French civil proceedings.

The AI Act layer: a copyright policy and a training summary

The AI Act does not change copyright law, but it turns respect for it into a regulatory duty for the providers of general-purpose AI models. Article 53(1)(c) requires them to put in place a policy to comply with EU copyright law, and in particular to identify and comply with reservations under article 4(3) of the Directive, "including through state-of-the-art technologies". Article 53(1)(d) requires them to publish a "sufficiently detailed summary about the content used for training", using a template from the AI Office (7). These duties have applied since 2 August 2025; models already on the market before that date have until 2 August 2027. The Commission's enforcement powers, including fines, started on 2 August 2026, and the Commission's announcement of that date named the copyright policy and the training summary among the duties of every provider (5).

The General-Purpose AI Code of Practice, published on 10 July 2025, explains what the copyright policy is expected to contain (20). Signatories commit to crawl only lawfully accessible content, without circumventing paywalls, and to exclude websites recognised as persistently infringing (Measure 1.2). They commit to use crawlers that read and follow robots.txt as specified in the IETF's RFC 9309, and to respect other appropriate machine-readable protocols, such as asset-based or location-based metadata, that are standardised or widely adopted (Measure 1.3). Providers that also run a search engine are encouraged to make sure that respecting a mining reservation does not harm the indexing of the content concerned. Signatories also commit to technical safeguards against outputs that reproduce training content in an infringing way and to a contact point and complaints procedure for rightholders (Measures 1.4 and 1.5). The training summary follows a template published on 24 July 2025, in three parts: general information, data sources, and relevant aspects of data processing (21).

A firm that uses an AI tool does not take on these duties. They matter to it in two other ways: as a source of information when choosing tools, and as the regulatory framework for clients who develop or fine-tune models.

The firm as a user of AI tools

The copyright risk for a firm that simply uses a reputable tool is modest. Training data is in the first place the provider's problem, and the Munich court held the operator responsible for infringing outputs generated by simple prompts. Four practical issues, however, land on the firm's side of the line.

What the tool was trained on, and with what. Providers of general-purpose models must now publish training summaries (for models already on the market before August 2025, by August 2027), so a vendor can be asked where they are. A legal AI vendor should be able to say which models it relies on, which legal content it has licensed and on what terms, and whether anything the firm puts into the tool is used to train or improve models for others. The last point is primarily a matter of professional secrecy rather than copyright, but it is also where the rights of third parties come in. The French Conseil national des barreaux makes the point in its practical guide on AI: an avocat using generative AI must check that the provider's terms do not allow it to appropriate or reuse what is submitted, because the avocat often does not hold the rights in that material, such as the other side's submissions, contracts or doctrinal articles written by others (22).

What the firm's database licences allow. Copying chapters of a commentary or a run of annotated decisions from a subscription database into a general AI tool is a reproduction of protected material, and it may also breach a licence that restricts automated or bulk use. Because the commercial mining exception can be excluded by contract, the licence terms decide what the firm may do, not the exception. Read them before building internal knowledge tools on licensed content, and ask publishers what their AI terms permit.

Outputs. AI-generated text can reproduce protected material, and the German Federal Bar's AI guidance warns that such text, although itself in principle unprotected, can infringe third parties' rights (23). A firm whose client alert, brief or website text contains reproduced passages is itself reproducing them, and communicating them to the public if it publishes them. The safeguard is the one lawyers already apply to citations: check what comes out. Some providers offer contractual protection. Microsoft's Customer Copyright Commitment, for example, is a provision of its Product Terms under which it defends customers against certain third-party intellectual property claims relating to output content; for Azure OpenAI and other configurable services, cover depends on required mitigations being in place (24). Such indemnities help, but they come with conditions the firm has to meet.

Credentials. Footnote 9 of Ross describes a company trying to use the research credentials of a law firm that had invested in it, contrary to the database's terms. Sharing research logins with technology partners, portfolio companies or vendors building tools for the firm will usually breach the licence and, in Europe, can defeat the lawful access on which every mining exception depends.

The European bars point the same way. The CCBE's guide on generative AI of October 2025 lists ownership of input and output as a key concern, notes that copyrighted or unlicensed data may have been used for training and that inputs containing protected works can lead to recognisable outputs, and warns that tool terms may let the vendor reuse what lawyers submit (25).

The firm as a rightholder

Law firms publish a great deal: client alerts, articles, guides to new legislation, commentary on judgments. The CCBE guide states the consequence bluntly. Crawling and scraping mean that firms' intellectual assets, "case analyses, legislative guides, and specialised research publications", are harvested into training datasets that may later power services competing with lawyers, and lawyers "should remain vigilant" about what they publish (25). Whether a firm wants to stop that is a business decision; some firms want their analysis to appear in AI answers. A firm that does want to reserve its rights should do it in a way that would hold up in court.

A robots.txt that covers the relevant crawlers. The Code of Practice commits signatories to follow robots.txt (20). The Amsterdam judgment shows the limit: a file that names only some AI bots did not count as a reservation against a crawler it did not name (16). Decide whether to name crawlers one by one, which needs maintenance, or to restrict more broadly, which affects search engines too.
A reservation in the website terms. Recital 18 counts website terms as a machine-readable means (6), but the Hamburg court required that a reservation can be interpreted automatically (15). A sentence in the terms helps, but is unlikely to be enough on its own.
Machine-readable metadata. Protocols such as the TDM Reservation Protocol, published as a W3C Community Group report rather than a formal standard, let a site declare a reservation in a form crawlers can process (26).
Evidence. Keep dated copies of the robots.txt, the terms and the metadata. In LAION the question was what the reservation looked like in the second half of 2021, and the rightholder bore the burden of showing it was machine-readable then (15).

Two limits remain. A reservation works for the future and does not undo collection that has already happened. And it does not bind research organisations mining for scientific research, whose exception cannot be reserved against at all: LAION, a non-profit association, won partly on that basis.

A related question is whether the firm owns what it produces with AI. Under the EU standard, a text is protected only to the extent that it reflects a human author's own free and creative choices (11). The German Federal Bar's guidance states that text generated by an AI system is in principle not protected for lack of a personal intellectual creation (23). A lawyer who selects, structures and rewrites an AI draft can add protectable expression; a template produced by a prompt and left largely untouched may have little protection of its own. For a firm's know-how, copyright was never the main safeguard. Confidentiality, access control and contracts with staff and vendors protect it better.

Advising clients who build, buy or license models

The third position is the one most firms will be paid for. Clients that develop or fine-tune models, clients that buy them, and clients whose content is being collected all need advice that combines copyright, contract and the AI Act.

Developers need to show lawful access to their sources, detect and honour reservations at the time of crawling and keep records that prove it, document where their datasets come from, reduce memorisation and filter outputs in the light of the Munich judgment, and, if they place a general-purpose model on the EU market, maintain the copyright policy and publish the training summary (7) (21). Licensing high-value content directly avoids depending on an exception whose scope the Court of Justice has yet to settle.
Buyers need contract terms: warranties about training data and licences, an indemnity for claims about output and its conditions, cooperation if a claim arises, access to the provider's training summary and copyright policy, and a clear rule that the buyer's inputs are not used to train models for others.
Rightholders, including publishers and media clients, need effective reservations, monitoring of whether they are respected, and a litigation strategy that takes account of the evidence problem. In France, the Senate bill would change that strategy if it becomes law (19).

Ross makes one further point that holds in Europe too. The court treated the emerging market for licensing content as training data as a market the copier had harmed. European law reaches the same place by a different path, through reservations and licences, but the commercial reality is the same: content that can be licensed for training has a price, and using it without a licence or a clear exception is a litigation risk.

A firm copyright policy for AI in eight decisions

Tools are chosen with their training in view. For every AI tool, the firm knows which models it uses, where their training summaries are, what legal content is licensed, and whether firm inputs train anything.
Database licences are read for AI terms before licensed content is used in any AI tool or internal knowledge system.
No licensed commentary goes into general tools beyond short quotations, unless the licence allows it.
Research credentials are personal and never shared with vendors, partners or companies the firm has invested in.
Output is checked for reproduced text before it leaves the firm, in the same review that checks citations.
The firm decides whether to reserve its published work and, if so, does it in machine-readable form and keeps dated evidence.
Vendor contracts cover training data and output: warranties, indemnity conditions and the use of the firm's inputs.
Someone follows the case law. The Federal Court of Justice's decision of 17 December 2026, the Munich appeal, the Court of Justice's ruling in Like Company and the French bill could each change parts of this list.

How this fits the way LexVera supports legal work

Ross turned on the difference between the law and what is written about it. LexVera is built for Dutch and European law: Dutch legislation and case law, EU law and the case law of the European Court of Human Rights. Its research works from legislation and court decisions as published by public authorities, the layer of legal material that copyright leaves open to everyone, and it keeps each proposition tied to the passage it rests on, so that the reviewing lawyer opens the official text rather than relying on a summary of it.

A firm's own documents and know-how stay within the firm's workspace, under the access rules the firm sets for its matters and teams. LexVera supports the preparation; a qualified lawyer decides what becomes advice, which is the standard set out in our editorial policy.

Frequently asked questions

Does Thomson Reuters v. Ross apply in Europe?

No. It applies American fair use, which European law does not have. Its central distinction does carry over, though: court decisions are free to use, while the editorial material publishers add to them is protected. In Europe, copying for AI training is lawful only within a statutory exception or under a licence, and the AI Act requires providers of general-purpose models placed on the EU market to have a policy for complying with EU copyright law, wherever the training took place.

May AI tools be trained on court decisions and legislation?

As far as copyright is concerned, largely yes. Dutch and German law expressly exclude legislation and court decisions from copyright, and German law extends that to headnotes written by the court itself. France publishes its court decisions as reusable open data. Other limits remain: data protection, the database right in a publisher's compilation, and in France a criminal ban on using judges' identity data to evaluate or predict their professional practices.

May AI be trained on headnotes, annotations or legal commentary?

Only with a licence or within the text and data mining exception. That exception requires lawful access and lapses if the rightholder has reserved its rights in an appropriate way, which for online content means a machine-readable reservation. Licence terms can also exclude text and data mining for commercial purposes, because, of the two mining exceptions, EU law protects only the research exception against contractual override.

How can a law firm stop AI developers from training on its website content?

By reserving its rights in a machine-readable way: a robots.txt that covers the relevant crawlers, a text and data mining reservation in the website terms and machine-readable metadata. Courts in Amsterdam and Hamburg found that a robots.txt naming only some AI bots, and a prohibition written only in natural language, did not suffice in the cases before them. A reservation does not undo past collection and does not bind non-commercial research organisations.

Is a document drafted with AI protected by copyright?

Only to the extent that it reflects a human author's own free and creative choices, which is the EU standard. Text generated by an AI system without such choices is in principle not protected, as the German Federal Bar also notes. A lawyer's selection, structuring and rewriting can be protected. AI output can also infringe someone else's copyright, so it must be checked before it is used.

What should a firm ask its legal AI vendor about training data?

Which models the tool uses and where their public training summaries are, which legal content the vendor licenses and on what terms, whether the firm's own inputs and documents are used to train or improve anything, what safeguards prevent the tool from reproducing protected text, and whether the vendor indemnifies the firm against copyright claims about output and on what conditions.

Sources and methodology

This guide reflects legislation, court decisions, official notices, bar guidance and regulatory documents available on 22 October 2026. Several of the decisions discussed are under appeal or awaiting a higher court: the Federal Court of Justice will rule in the LAION case on 17 December 2026, the GEMA judgment is under appeal, the Amsterdam judgment is a first-instance decision, and the Court of Justice has not yet ruled in Like Company. The French bill described is not law. Statutory provisions are paraphrased in English; the quoted wording of the Directive and the AI Act is the official English text. This is general professional information, not advice on a specific matter.