AI Copyright Claims in Europe: Key Considerations

Frederiek Fernhout, Stijn de Jong and Branda Katan, 4 September 2026. 

As AI copyright litigation accelerates in the United States, technology companies operating in Europe face a structurally different legal landscape. Unlike in the US, fair use does not exist under Dutch or EU copyright law. The DSM Directive's text and data mining exceptions are narrow, subject to opt-out, and have been interpreted restrictively by German courts. The WAMCA, combined with developing case law on EU-wide jurisdiction, suggests that large-scale collective AI copyright and privacy claims in the Netherlands may be a question of timing rather than probability. This article sets out the key legal considerations, the most significant pending cases – including the Advocate General opinion in Like Company v. Google – and the steps technology companies may wish to consider to assess their exposure across European markets.

1. The scale of the problem

As of June this year, more than 70 AI copyright infringement lawsuits have been active or recently resolved in US and international courts. Here, we highlight a few of the most prominent ones. 

On 5 May this year, five of the world's largest publishers, Elsevier, Cengage, Hachette, Macmillan and McGraw Hill, filed a proposed class-action lawsuit in the Manhattan federal court against Meta Platforms, alleging that Meta pirated millions of copyrighted works ranging from scientific textbooks to novels in order to train its Llama large language models. The lawsuit opens a new front in the ongoing battle between content creators and AI developers. The latest significant filing came on 10 July 2026, when three of those same publishers, Hachette, Cengage and Elsevier, together with author Scott Turow, filed a similar proposed class action against Google, alleging that Google used millions of copyrighted books and journal articles, obtained in part from pirate sites, to train its Gemini models, and that it stripped or altered copyright management information to conceal the practice. Meta and Google are not alone in this: OpenAI, Anthropic, and Microsoft face similar claims brought by authors, news organisations, and visual artists. Anthropic became the first major AI company to settle, agreeing in 2025 to pay a group of authors USD 1.5 billion to resolve a class action  that could have cost it substantially more (‘Bartz v. Anthropic’). The New York Times has sued OpenAI and Microsoft. The legal and financial stakes across the industry are considerable.

For technology companies developing or deploying AI systems in Europe, US litigation is only part of the picture, and arguably not the most important part. The European legal framework is structurally different from US law and, in important respects, considerably more protective of rights holders.

2. The US situation: fair use doctrine

Meta’s defence against the five publishers to the proposed class action in the United States is already taking shape: 

"AI is powering transformative innovations... and courts have rightly found that training AI on copyrighted material can qualify as fair use. We will fight this lawsuit aggressively." 

It is clear that the US publishers' case will ultimately turn on whether Meta's use of copyrighted works for AI training constitutes "fair use" under Section 107 of the US Copyright Act. Fair use is a flexible, open-ended doctrine that US courts apply by weighing four factors: the purpose and character of the use (including whether it is commercial or transformative), the nature of the copyrighted work, the amount of the work used, and the effect of the use upon the potential market for the original work. The case law is still forming. In Bartz v. Anthropic, the court found that training on copyrighted works itself can qualify as fair use, though storing pirated copies of those works does not. The court's finding that the provenance of training data can give rise to independent liability has since featured prominently in subsequent pleadings against other AI developers. 

The most closely watched case at present is Thomson Reuters v. ROSS Intelligence, in which oral arguments were heard on 11 June 2026.  It is the first US appellate consideration of whether training an AI model on copyrighted works constitutes fair use. The case is generally seen as a highly instructive case for the rest of the field, with the district court finding no fair use. A ruling in the appellate case is expected before the end of 2026. Depending on its direction, it will either strengthen the hand of rights holders across every pending case or open the fair use question at appellate level for the first time. Either way, European courts will read it with interest. It will not, however, affect the analysis under Dutch or EU law.

3. Why the fair use argument does not work in the EU

Fair use is an exclusively US-based concept and has no counterpart in Dutch or EU copyright law. Under the Dutch Auteurswet and the EU framework established by the DSM Directive (Directive 2019/790), limitations on copyright are exhaustively listed and narrowly interpreted. If a use does not fit within one of those specific exceptions, it is an infringement, and there is no judicial ‘fair use’ discretion to recognise new exceptions on grounds of technological necessity or public benefit. A favourable US ruling therefore does not protect technology companies in EU proceedings.

The most directly relevant provisions in the DSM Directive are the two text and data mining (TDM) exceptions. Article 3 covers research organisations conducting scientific research. Article 4 covers any user for any purpose, but rights holders may opt out by expressing a reservation in machine-readable form. The most established mechanism is a robots.txt file hosted at the root of a website, which crawlers are expected to consult before accessing content. Once a valid reservation is in place, the exception no longer applies. Any subsequent scraping of that content is infringement. Rights holders are increasingly aware of this option, making a violation of this article more likely over time. 

Beyond this opt-out option, there is a limit on what the exceptions permit. They were designed to allow the extraction of information and insights from works, not the mass reproduction and permanent incorporation of entire texts into commercial model weights. The German Regional Court of Munich held, in proceedings brought by GEMA against OpenAI, that where a model memorises protected works and can reproduce them in outputs, the use goes beyond what the exception permits. The case is now before the Munich Court of Appeal. Because Article 4 is set out in the same Directive across all Member States, an appellate ruling against OpenAI in Germany would carry direct relevance for Dutch proceedings.

At CJEU level, Like Company v. Google (C-250/25), a referral from a Hungarian publisher, will address TDM in the context of generative AI directly. The Advocate General is expected to deliver an advisory opinion in September 2026, with a final ruling highly anticipated in 2027. Possible outcomes range from following the European Commission's recommendation and dismissing much of the referral, to narrowing the questions to the facts of the case, or hearing all four questions in full, potentially setting precedents on highly disputed legal issues – although critics have urged the CJEU to exercise caution. Whatever the outcome, CJEU rulings rarely provide the comprehensive clarity practitioners hope for, and this case is unlikely to be an exception.

The DSM Directive imposes real constraints on AI training in Europe, and the case law is developing quickly. But copyright law is only part of the picture. The EU AI Act adds a separate and reinforcing layer of obligations that technology companies need to understand alongside it.

4. European framework AI Act

The General-Purpose AI model obligations under the EU AI Act have applied since 2 August 2025. Enforcement fines apply from 2 August 2026, at up to 3% of global annual turnover or EUR 15 million, whichever is higher. The European AI Office's final GPAI Code of Practice, endorsed on 1 August 2025, requires providers to maintain documented policies for identifying and respecting opt-out rights reservations, to treat paywalls as equivalent to opt-outs, and to exclude content sourced from pirate websites. 

The mandatory training content summary under Article 53(1)(d) is particularly significant from a litigation perspective. It requires providers to publish a sufficiently detailed summary of the content used to train their models. This obligation sits alongside the copyright compliance policy requirement and is designed to address the information asymmetry that has long characterised large-scale AI training. 

Once published, those summaries may become evidence for rights holders and regulators assessing whether content was used without permission and whether opt-outs were respected. They may also lower the evidentiary burden for claimants in collective redress proceedings, providing precisely the documentary foundation that large-scale copyright claims require.

5. Competition law

EU competition law prohibits undertakings from abusing a dominant market position. Depending on the specific market assessment and the dominant firm’s position, claimants could seek to characterise unauthorised AI training as an abuse of dominance. 

Such a legal characterisation would be novel and untested. It would require establishing not only dominance, but also a sufficient link between the dominant position and the alleged abusive conduct, as well as demonstrating that the conduct forecloses competition or exploits market power. In essence, this would require elevating an intellectual property infringement into an EU competition law violation.

By way of illustration, consider a scenario in which a dominant search engine leverages its market position to index and cache content from publishers at scale, and subsequently uses that content without authorisation to train an AI model. The AI model then generates responses that reduce users’ need to visit the original sources, thereby undermining the publishers’ revenue base. In such a scenario, the dominant undertaking may be said to leverage its dominance on the search market to foreclose competition on the adjacent market for AI-generated content, with the unauthorised training constituting the means through which the abuse is effected.

Irrespective of the likelihood of success of such a claim, it is worth highlighting its potential legal consequences in civil litigation. If a breach of EU competition law is successfully alleged, the competition law concept of “undertaking” would apply. Under this concept, separate corporate entities within a corporate group can, under certain circumstances, be treated as forming a single undertaking that is liable for the infringement. This means that if antitrust liability were to be established and the relevant conditions satisfied, a local Dutch entity – such as a subsidiary or sister entity – could be held jointly liable for the group’s unauthorised AI training practices. Furthermore, this could be done under the Dutch rules for mass litigation (see below).

6. Collective redress in the Netherlands

Since January 2020, the WAMCA (Wet afwikkeling massaschade in collectieve actie) allows a representative organisation to bring a mass claim for the benefit of an entire group of affected parties who are represented unless they actively opt out. Unlike the previous regime, the WAMCA permits claims for damages. More than 100 cases have been brought under the WAMCA for tens of billions of euros in total, including large actions modelled on US class-action practice, such as privacy mass claims including against TikTok. Rights organisations are not starting from scratch. Foundations such as BREIN and React already have an active WAMCA infrastructure, experience in collective proceedings, and established relationships with the courts. The practical and procedural groundwork for a large-scale AI copyright claim in the Netherlands already exists.

Two aspects deserve particular attention.

Jurisdiction over EU-wide classes. Two routes may allow Dutch courts to accept jurisdiction over EU-wide classes, and technology companies should be aware of both. First, if claimants frame an AI training claim as involving abuse of market dominance, they may invoke the competition law concept of "undertaking," under which a Dutch anchor defendant could be held jointly liable for the training practices of the entire corporate group. Second, and more immediately, courts may base jurisdiction over EU defendants on Article 8 Brussels I recast without needing the "undertaking" concept. The current interpretation of this ground for jurisdiction by Advocate General Ibili in his opinion in the Stellantis case before the Dutch Supreme Court – where a judgment is expected by the end of September 2026 – suggests that if part of the class has a claim against the anchor defendant, the Dutch court may have jurisdiction over the claims of the entire class, EU-wide. If that interpretation is confirmed, it would make the Netherlands a significantly more attractive forum for EU-wide AI copyright and privacy mass claims. Moreover, Article 8(1) Brussels I recast has an equivalent in Article 7(1) Dutch Code of Civil Procedure for non-EU defendants. If Advocate General Ibili's view is upheld, this may tempt claim organisations to try and bring a worldwide rights infringements class action against, for instance a US technology company, using a Dutch subsidiary as anchor defendant.

Commonality: Claimants will argue that AI training claims have sufficient commonality to bundle them into a class action: the same training run constitutes a common underlying event affecting all represented rights holders simultaneously. That argument has some force. Defendants, however, have genuine grounds to contest homogeneity. Training datasets differ between model versions. The extent to which individual works were actually ingested, memorised or reproduced may vary considerably. Damage assessment will in most cases require individualised analysis. These are real arguments, and they should be developed early. 

Two types of claims are most plausible in the Dutch context: copyright infringement on behalf of a rights organisation and its members, analogous to the Authors' Guild action in the US; and GDPR-based privacy claims on behalf of Dutch internet users whose personal data was processed without an adequate legal basis during training data collection. Several GDPR mass claims are already pending in the Netherlands, where the framework is in several respects stricter than US law (see also our previous blog).

7. Conclusion

The US publishers' lawsuit against Meta is significant, but for technology companies operating in Europe it is only the starting point of the legal analysis. The European framework is structurally different: fair use is unavailable and TDM exceptions are narrow, subject to opt-out, and have been interpreted restrictively by German courts in ways that may influence Dutch proceedings. Moreover, if claimants successfully frame unauthorised AI training as a novel abuse of dominance under EU competition law, the “undertaking” concept could, under certain circumstances, render a local Dutch subsidiary or sister entity jointly liable for the entire group’s conduct. And the WAMCA, combined with various a broad interpretation of grounds for jurisdiction, means that large-scale collective AI copyright and privacy claims in the Netherlands are a question of when, not whether.

The Advocate General's opinion is expected to land in September 2026. The Dutch Supreme Court’s ruling in Stellantis is expected by the end of the same month. The Thomson Reuters appellate ruling follows before year’s end. The window for companies to audit their exposure, strengthen their compliance documentation and develop a coherent cross-border defence strategy is open now.