Authored By: Ludovica Corsello
Trinity College Dublin
I. Introduction
In May 2026, a federal judge in California held a fairness hearing to decide whether to approve a $1.5 billion settlement between Anthropic and a class of authors whose books the company had used, without permission, to train its Claude models.1 The underlying ruling that led to this settlement was itself striking: the court found that training an AI model on legally acquired books was fair use, yet training it on millions of pirated ones was not. Judge Alsup did not resolve the fair use question once and for all; instead, he gave two different answers depending on how the training data had been obtained. This distinction exposes a broader problem in how copyright law encounters generative artificial intelligence.
Generative AI systems are built on vast quantities of text and images, much of it pulled from the internet without permission from rightsholders. This practice sits uneasily within a copyright framework that grants authors exclusive control over reproduction, yet was designed decades before AI training existed and therefore never anticipated that such training could infringe copyright. The result is a structural tension between the interests of AI developers, who require enormous datasets to train competitive systems, and the interests of creators, who receive neither compensation nor a meaningful say in how their work is used.
This article argues that the mechanisms currently available to regulate this practice, whether the European Union’s opt-out exception, the United Kingdom’s research-only carve-out, or the fair use doctrine applied in the United States, were built for human-scale copying and are structurally unable to give rightsholders meaningful control over how their work trains AI models.
II. Legal Framework
A. The European Union
The European Union addresses text-and-data mining through Articles 3 and 4 of the Digital Single Market Directive.2 Article 3 grants research organizations and cultural heritage institutions a mandatory exception for text-and-data mining carried out for scientific research, which rightsholders cannot override by contract.3 Article 4 permits text-and-data mining for any purpose, including commercial AI training, but this extension is subject to a major limitation: rightsholders may reserve their rights, usually through machine-readable means, and once they do, the exception no longer applies to their works.4 This opt-out mechanism is the central design feature of the EU approach, but it is also the most controversial point, since it places the burden of protection on the rightsholder rather than on the party intending to use the work.
The EU has since added a transparency obligation to this system. Providers of general-purpose AI models are now required, under the AI Act, to put in place a policy for complying with EU copyright law and to publish a sufficiently detailed summary of the content used to train their models.5 This does not alter the substantive scope of the Article 4 exception, but it is intended to make the opt-out mechanism enforceable in practice, by giving rightsholders the information needed to know whether their works were used.
B. The United Kingdom
The United Kingdom’s position is more restrictive: section 29A of the Copyright, Designs and Patents Act 1988 permits text-and-data mining only where it is carried out for non-commercial research, and the user already has lawful access to the work.6 Unlike the EU framework, there is no broader exception covering commercial use; therefore, training a commercial AI model on copyrighted works without a license falls outside section 29A.
The government ran a consultation from December 2024 to February 2025 to consider introducing a broader exception modelled on the EU’s opt-out approach. Following further consultation, however, it published the Report on Copyright and Artificial Intelligence in March 2026, confirming that this option was no longer being pursued.7 The report set out that no legislative reform is imminent, and that the government intends instead to focus on licensing, transparency, and technical standards while continuing to observe developments elsewhere. The outcome is that the position under section 29A remains the operative law, and commercial AI training in the UK continues to require a license in the absence of any applicable exception.
C. The United States
The United States has no dedicated law that addresses text-and-data mining specifically. Instead, when a court must decide whether training an AI model on copyrighted material is lawful, courts apply the general fair use doctrine codified at 17 U.S.C. § 107, which directs them to weigh four factors: the purpose and character of the use, the nature of the copyrighted work, the amount and substantiality of the portion used, and the effect of the use on the market for the original.8 Because this is a judicially applied standard rather than a clear and automatic rule, the scope of protection available to AI developers has been worked out case by case, a point taken up in the next section.
III. Case Law
Statutes can rarely be applied cleanly to a technology that did not exist when they were drafted. This is especially true in the United States, where courts must fit AI training into an open-ended fair use standard never written for the case of machine learning. Similarly, the United Kingdom and the European Union face their own difficulties, as the meaning of “lawful access” and the limits of the text-and-data mining exceptions remain contested. The real content of the law in this area, therefore, is found less in the statutory text than in how courts have applied it to specific disputes.
A. The United States
The most consequential ruling to date is Bartz v. Anthropic.9 In June 2025, Judge Alsup granted Anthropic summary judgment on fair use for training its models on books it had legally purchased, holding that this was a transformative use, but denied summary judgment on the separate question of books acquired through pirate sites, finding that the manner in which the books were obtained was not excused by the eventual transformative use. This distinction, between the use of a work and the means of its acquisition, is significant: it shows a court willing to find training itself lawful while still scrutinizing how the training data was acquired. Because Anthropic chose to settle rather than appeal, Alsup’s fair use holding was never tested at the Ninth Circuit and remains a single district court decision that binds no other court. This matters directly for the argument developed below: the case that produced the most detailed reasoning yet on AI training and fair use is not binding precedent anywhere outside its own courtroom.
Thomson Reuters v. Ross Intelligence presents a sharper conflict: in February 2025, Judge Bibas reversed his own earlier view and granted Thomson Reuters summary judgment, holding that Ross’s use of Westlaw headnotes to train a competing legal research tool was not transformative and infringed copyright. Ross has since taken an interlocutory appeal to the Third Circuit, making this the first time a U.S. Court of Appeals will hear arguments on whether AI training can be fair use. Unlike Bartz, where the transformative use finding favored the AI developer, the reasoning here turns on the commercial substitution point: Ross’s tool was built to compete directly with the product it was trained on, which cuts against a finding of fair use.10 Read together, the two cases suggest that US courts are treating “transformativeness” as highly sensitive to whether the resulting AI product competes with the market for the original works, rather than applying a uniform rule to AI training as such.
B. The United Kingdom
Getty Images v. Stability AI is the first UK ruling to address copyright and AI training directly.11 The court held that, even where a model was trained outside the country but made available in the UK, secondary infringement could in principle arise. However, Getty was still required to prove that its images were contained in the model, something it failed to demonstrate; no infringement was therefore established.
The significance of this case lies less in its outcome than in its reasoning: it confirms that a rightsholder pursuing a UK claim must show the model itself retains the copyrighted expression, a considerably harder burden than showing that the training process involved copying.
C. The European Union
Two of the most relevant decisions were produced by German courts. In Kneschke v. LAION, the Hamburg court held that assembling a dataset by analyzing lawfully accessible, non-reserved works fell within the scope of the EU’s TDM exception.12 In GEMA v. OpenAI, the Munich court took a different angle, holding that an AI model’s ability to reproduce exact song lyrics was itself evidence that the underlying works had been memorized during training, a finding relevant to whether copies are retained within a model rather than merely processed during it. Neither decision has yet reached the Court of Justice of the EU, though a preliminary reference on aspects of the Article 4 exception is pending.13
IV. Critical Evaluation
A. The Opt-Out Model’s Practical Limits
The EU’s Article 4 exception is, on its face, the most coherent response among the three jurisdictions: it preserves a rightsholder’s ability to withhold their work from training while allowing data mining to proceed by default. In practice, however, the mechanism assumes a capacity that most rightsholders simply do not have. As the UK’s own 2026 consultation responses illustrate, individual creators described implementing an opt-out across the full range of platforms on which their work might appear as significantly burdensome, and for large bodies of work, effectively impossible. The Kneschke v. LAION reasoning compounds this: by treating the exception as also covering the preparatory act of assembling a dataset from lawfully accessible, non-reserved works, German courts have extended the exception’s reach beyond training itself to the earlier stage of dataset curation, further widening the pool of material that falls outside a rightsholder’s control by default. An exception that depends on active, technically informed rights reservation by millions of individual authors and photographers is not, in any meaningful sense, protective of those least equipped to exercise it.
B. The UK’s Narrow Exception Solves Little
The UK’s approach avoids this enforceability problem by declining to introduce a comparable exception at all, but this comes at a different cost. Because section 29A applies only to non-commercial research, it offers UK rightsholders no additional protection against commercial AI training that would not already exist under ordinary copyright law, and it offers UK AI developers no meaningful domestic route to train models lawfully at scale. The consequence, acknowledged directly in the government’s own report, is that large developers simply train elsewhere, most commonly in the United States, where the position is more permissive. Restraint at the level of a single, mid-sized jurisdiction does not reduce the amount of UK-authored material used to train foreign models; it only ensures that the UK receives none of the resulting economic activity while its creators remain exposed to training conducted under a different country’s law.
C. Fair Use’s Costly Unpredictability
The US approach, resolved case by case rather than by statute, produces the least uniform outcomes among the three jurisdictions, but not necessarily the least protective ones. Thomson Reuters v. Ross shows that fair use fails where the resulting product substitutes commercially for the original; Bartz v. Anthropic shows it succeeds where the training is one step removed from any directly competing product, but only for lawfully acquired material. The difficulty is that neither ruling binds the other, and the Bartz settlement means the most detailed judicial reasoning on this question to date carries no precedential weight beyond its own courtroom. Developers and creators alike are left to plan around a doctrine whose outcome depends heavily on facts that only emerge after years of litigation, which is a poor basis for either investment decisions or licensing negotiations.
D. Addressing the Innovation-Incentive Argument
A natural objection is that broad exceptions, or at least fair use’s flexibility, are necessary because AI development depends on access to a wide variety of data that licensing alone cannot realistically supply, and that restricting it would push investment toward jurisdictions with fewer constraints. This objection is not wrong, and the UK report itself concedes the point: developers already choose their training jurisdiction based on comparative regulatory permissiveness, so unilateral restriction by any single country does little to protect that country’s own creators. But this concession supports this article’s thesis rather than undermining it. If no single jurisdiction’s copyright exception can, by itself, secure meaningful protection for its creators, precisely because developers can simply train elsewhere, then the problem is not that any one exception is poorly drafted; it is that the entire model of jurisdiction-by-jurisdiction copyright exceptions is not structurally capable of governing a technology whose training activity is inherently mobile across borders.
E. Comparative Synthesis
None of the three models, opt-out, narrow exception, or fair use, gives rightsholders reliable, enforceable control over whether their work trains a commercial AI system. The EU model fails because the burden of protection falls on those least able to bear it; the UK model fails because it protects on paper while displacing the activity it cannot reach; and the US model fails because it resolves the question too slowly and too inconsistently to give either side dependable guidance. What the comparison reveals is not that one jurisdiction has found the right balance and the others should follow it, but that the underlying legal tool, a domestic copyright exception aimed at a single act of reproduction, was built for a world of localized, traceable copying, and is not well suited to a technology whose entire value depends on aggregating inputs across jurisdictions at a scale no single national law can fully see or reach.
V. Conclusion
No jurisdiction’s exception gives rightsholders meaningful control over AI training, and the evidence above shows why.
The EU model fails because Article 4’s opt-out demands a technical capacity most creators lack, worsened by Kneschke v. LAION extending the exception’s reach to dataset curation itself. The UK model fails because section 29A’s narrow scope protects nothing against commercial training, while developers simply train elsewhere. The US model fails because Bartz v. Anthropic and Thomson Reuters v. Ross reached opposite results on closely related questions, and neither binds beyond its own courtroom.
That convergence is the proof: the gap cannot be a drafting flaw in any single jurisdiction. It is the structural limit identified above: a domestic exception built for one act of reproduction cannot govern an activity that aggregates inputs across borders faster than any one law can reach.
Bibliography
Table of Cases
Bartz v Anthropic PBC, No 3:24-cv-05417 (ND Cal, order on fair use, 23 June 2025)
Bartz v Anthropic PBC, No 3:24-cv-05417 (ND Cal, final approval fairness hearing held 14 May 2026)
GEMA v OpenAI, Landgericht München I, 11 November 2025, Case No 42 O 14139/24
Kneschke v LAION eV, Landgericht Hamburg, 27 September 2024, Case No 310 O 227/23
Thomson Reuters Enterprise Centre GmbH v Ross Intelligence Inc, No 1:20-cv-00613-SB (D Del, 11 February 2025)
Getty Images (US) Inc & Ors v Stability AI Ltd [2025] EWHC 2863 (Ch)
Table of Legislation
Copyright Act 1976, 17 USC § 107
Copyright, Designs and Patents Act 1988
Directive (EU) 2019/790 of the European Parliament and of the Council of 17 April 2019 on copyright and related rights in the Digital Single Market and amending Directives 96/9/EC and 2001/29/EC [2019] OJ L130/92
Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) [2024] OJ L1689
Other Sources
Intellectual Property Office, Copyright and Artificial Intelligence: Consultation (December 2024–February 2025)
Intellectual Property Office, Report on Copyright and Artificial Intelligence (March 2026)
