AI Copyright and International Law: Training Data Under Berne and TRIPS

Introduction

International copyright law does not categorically prohibit the use of copyrighted works to train artificial intelligence systems, but it does not grant unrestricted permission for that use. The legal position depends on the reproduction rights established by the Berne Convention, the obligations imposed by the Agreement on Trade-Related Aspects of Intellectual Property Rights (TRIPS), and the exceptions recognized under applicable national legislation. Neither treaty expressly establishes a general authorization for AI training.

Generative AI systems are developed using extensive collections of text, images, music, software, and other materials. These collections may contain copyrighted works, licensed content, public-domain materials, and factual information that copyright does not protect. Acquiring, storing, and processing such material can involve acts of reproduction that require authorization unless an applicable exception permits them.

The principal international rules were established before contemporary machine learning emerged. Article 9 of the Berne Convention protects the author's exclusive reproduction right while allowing carefully limited exceptions. TRIPS incorporates most substantive Berne obligations and imposes further restrictions on copyright exceptions through Article 13. The central question is whether national laws permitting text and data mining, including commercial AI training, remain within those international limits.

National approaches have diverged. The European Union recognizes specific text and data mining exceptions, the United States relies substantially on fair use, and Japan and Singapore have adopted provisions accommodating computational analysis. Courts have also begun to distinguish the copying of training material from the reproduction of protected expression within models and generated outputs.

These developments reveal a distinction between international obligations and domestic copyright liability. States must comply with their treaty commitments when establishing copyright protection and exceptions. AI developers, meanwhile, must determine whether their activities are permitted under the applicable national law. The legality of a particular training operation and the international compatibility of the legislation authorizing it are related but separate legal questions.

1. When AI Training Engages Copyright Protection

Copyright becomes relevant to artificial intelligence development when protected expression is reproduced or otherwise used in a manner engaging an exclusive right. The presence of copyrighted material in a training dataset does not, by itself, establish infringement. The legal assessment depends on the material involved, the acts performed, and the conditions under which the developer obtained or processed it.

Training datasets may contain novels, scientific publications, photographs, journalistic articles, musical compositions, and computer programs. Some materials remain protected by copyright, while others are publicly licensed or have entered the public domain. Certain collections combine protected expression with information that copyright law does not independently protect.

The technical purpose of AI training does not eliminate the significance of copyright. Developers may copy works to identify statistical relationships rather than distribute the original expression. Nevertheless, copyright law generally examines whether a protected act occurred, not simply whether the developer intended to provide the original work to a human audience.

1.1 Protected Works, Unprotected Facts, and Datasets

Article 2(1) of the Berne Convention defines literary and artistic works broadly, encompassing productions in literary, scientific, and artistic fields regardless of their mode or form of expression. Its categories include books, lectures, musical compositions, films, drawings, photographs, and other creative works. The Convention establishes minimum standards for protecting qualifying expression rather than exclusive ownership of all information conveyed through a work (Berne Convention, 1971).

TRIPS Article 9(2) expressly distinguishes protected expression from ideas, procedures, methods of operation, and mathematical concepts. Berne Article 2(8) also excludes news of the day and miscellaneous facts having the character of mere items of press information. These distinctions are particularly relevant to AI systems trained on journalism, scientific research, and other information-intensive materials.

A factual statement appearing in a newspaper article does not ordinarily become copyrightable merely because it was reported by a journalist. The article's original expressive content may nevertheless remain protected. Extracting a fact and reproducing the copyrighted article containing that fact are legally different activities.

Original compilations require a separate assessment. Berne Article 2(5) recognizes protection for qualifying collections of literary or artistic works, while TRIPS Article 10(2) protects compilations of data or other material that constitute intellectual creations through their selection or arrangement. Such protection does not automatically extend to individual facts or otherwise unprotected contents.

A dataset may also implicate rights beyond copyright in its constituent works. European Union law, for example, recognizes a distinct database right under Directive 96/9/EC where its statutory conditions are satisfied. That right is not a universal form of protection established by Berne or TRIPS. The legal status of a training corpus depends on the rights recognized in the jurisdiction concerned.

1.2 Reproduction Across the AI Training Process

Generative AI development involves several technical operations. Training material may be collected through web crawling, acquired from commercial suppliers, downloaded from repositories, converted into standardized formats, filtered, divided into smaller units, and stored for computational processing. Some of these operations create digital copies of protected expression.

Downloading a complete copyrighted work and retaining it in a training repository ordinarily raises a reproduction question. Temporary copies made during preprocessing or computational operations may also engage reproduction rights, although domestic laws differ in their treatment of transient and intermediate copying. An activity does not cease to involve reproduction merely because the copies are intended for automated analysis.

Other stages require more precise characterization. Tokenization and numerical processing transform information into representations used by machine learning systems. Whether a particular operation creates a legally relevant reproduction depends on the material retained, the technical process, and the governing legal standard. The existence of numerical representations does not establish that each contains a protected reproduction of the underlying works.

The question becomes more difficult where model parameters retain recoverable expression. Research on model memorization has demonstrated that some systems can reproduce portions of their training material. Whether this constitutes an additional reproduction of a protected work depends on identifiable expression, technical evidence, and applicable copyright doctrine.

The distinction between copies used to construct a training corpus and protected material retained within a trained model is particularly important. A statutory exception may authorize preparatory reproductions without necessarily authorizing every subsequent embodiment or dissemination of protected expression. Dataset acquisition, computational analysis, model retention, and generated outputs must be examined separately when the facts require it.

2. The Berne Convention and the Right of Reproduction

The Berne Convention establishes the principal international framework for protecting literary and artistic works. First adopted in 1886 and subsequently revised, it obliges contracting states to recognize specified rights while leaving substantial aspects of implementation to national legislation. Its reproduction provisions were formulated before modern machine learning, but their language is not confined to the technologies available when they were adopted.

The Convention does not operate as a comprehensive international copyright statute regulating every transaction between authors and AI developers. Its principal obligations concern the protection that contracting states must provide. Individual infringement claims, defenses, remedies, and questions of jurisdiction ordinarily operate through the relevant domestic legal system.

For AI training, the most significant provisions concern the exclusive reproduction right, the application of copyright to digital storage, and the international protection of foreign works. These rules establish the legal foundation against which national exceptions for computational copying must be assessed.

2.1 Article 9(1): Reproduction in Any Manner or Form

Article 9(1) of the Berne Convention recognizes authors' exclusive right to authorize reproduction of their works in any manner or form. The formulation is technologically broad. It is not limited to printed copies, commercial publication, or reproductions intended for immediate human consumption.

This language provides a substantial basis for treating digital copying during AI development as copyright-relevant. A company that downloads copyrighted books and stores electronic copies to construct a training corpus may exercise the reproduction right even if its ultimate objective is to develop a statistical model rather than distribute the books themselves.

The existence of an exclusive right does not establish that every reproduction infringes copyright. Authorization may be granted through licensing, and domestic legislation may recognize exceptions permitting particular acts. Berne Article 9(2) allows contracting states to create exceptions to the reproduction right within internationally defined limits.

A doctrinal disagreement concerns whether every copy made exclusively for computational analysis falls within the minimum international reproduction right. Senftleben (2026) argues that text and data mining copies may fall outside its scope because they concern computational extraction rather than the conventional exploitation of protected expression. This interpretation challenges an expansive understanding of reproduction; it is not an exemption expressly established by Berne.

The Convention's broad wording and subsequent international treatment of digital reproduction support a different interpretation, under which many intermediate copies engage the reproduction right and require authorization or an applicable exception.

Even where a particular activity arguably falls outside the minimum treaty right, national legislation may establish broader protection. The distinction between international minimum standards and domestic reproduction rights remains essential.

2.2 Digital Reproduction Under the WIPO Copyright Treaty

The WIPO Copyright Treaty, adopted in 1996, supplements the Berne Convention in the digital environment. Article 1(4) requires contracting parties to comply with Articles 1–21 and the Appendix of Berne. The WIPO Copyright Treaty is a separate international instrument whose obligations bind its contracting parties rather than automatically extending every provision to all Berne members.

The agreed statement concerning Article 1(4) confirms that the Berne reproduction right and its permitted exceptions apply to works used in digital form. It also recognizes that storing a protected work digitally in an electronic medium constitutes reproduction within the meaning of Berne Article 9. This clarification is directly relevant to the construction and retention of AI training datasets (WIPO, 1996).

The agreed statement does not determine whether every computational operation within a machine learning system creates a reproduction of protected expression. It establishes the legal significance of digital storage, while the classification of particular processes may depend on their technical characteristics and applicable domestic law.

Article 10 of the WIPO Copyright Treaty governs limitations and exceptions through the three-step test. Its accompanying agreed statement recognizes that contracting parties may appropriately extend existing exceptions into the digital environment and devise new exceptions suited to digital networks. Those possibilities remain subject to the treaty's conditions.

The WIPO Copyright Treaty does not create a specific authorization for generative AI training. Its contribution is to confirm the continuing application of copyright protection to digital reproduction while recognizing that national legislation may accommodate technological developments through properly formulated exceptions.

2.3 Article 5: National Treatment and Territorial Protection

Article 5 of the Berne Convention establishes national treatment for qualifying foreign works. Authors generally enjoy, outside the country of origin, the rights that the protecting country's law grants to its own nationals, together with the rights specially granted by the Convention. Article 5(2) also prohibits formalities as a condition for the enjoyment and exercise of Convention rights.

Protection is independent of protection in the country of origin, subject to the Convention's qualifications. Article 5(2) further provides that the extent of protection and the means of redress are governed by the law of the country where protection is claimed, apart from the Convention's own provisions (Berne Convention, 1971).

These principles are relevant to datasets containing works originating in several countries. A copyrighted photograph does not become freely reproducible merely because it appears on a foreign website. Conversely, a copyright exception recognized in one jurisdiction does not automatically govern every reproduction of that photograph elsewhere.

Article 5 does not resolve every question of private international law arising from cross-border training. The applicable law, competent court, and location of alleged infringement may involve domestic conflicts rules and facts concerning the copying activity. Treating Berne Article 5(2) as a complete worldwide choice-of-law code would overstate its effect.

The international obligations must also be understood through established principles of treaty interpretation under the Vienna Convention on the Law of Treaties. Treaty language, context, and object and purpose remain relevant to determining the scope of copyright protection. Economic or technological policy preferences cannot independently establish a new exception to an agreed treaty obligation.

3. TRIPS and the Three-Step Test for AI Training

The TRIPS Agreement incorporates international copyright standards into the World Trade Organization's legal framework. Its provisions require WTO members to provide minimum protection while allowing specified limitations and exceptions. For AI training, its principal significance lies in the relationship between national permissions for computational copying and the international restrictions imposed on those permissions.

The central limitation is the three-step test. Berne Article 9(2) applies the test to exceptions from the reproduction right, while TRIPS Article 13 extends its application more broadly to limitations and exceptions to exclusive rights. Each provision addresses the circumstances in which states may restrict copyright protection without breaching their international obligations.

The test is cumulative. An exception must concern certain special cases, avoid conflict with the normal exploitation of protected works, and avoid unreasonable prejudice to legitimate rights-holder interests. These requirements were interpreted by a WTO panel in 2000, although that dispute did not involve artificial intelligence.

Applying the conditions to generative AI raises questions that existing treaty language does not expressly resolve. Computational analysis may involve copying undertaken for purposes different from the original works, while commercial AI systems may affect established or emerging licensing markets. The legal significance of those effects depends on the exception's scope and the evidence concerning the protected works.

3.1 TRIPS Articles 9 and 13: International Obligations

TRIPS Article 9(1) requires WTO members to comply with Articles 1–21 of the Berne Convention and its Appendix, excluding the moral rights established by Article 6bis and rights derived from that provision. The exclusion of moral rights from TRIPS obligations does not remove the economic reproduction rights protected by Berne.

TRIPS clarifies protection for materials commonly incorporated into AI training datasets. Article 9(2) excludes ideas, procedures, methods of operation, and mathematical concepts from copyright protection as such. Article 10 protects computer programs as literary works and qualifying compilations of data without automatically extending protection to otherwise unprotected contents.

Article 13 restricts limitations and exceptions to exclusive rights through the three-step test. Its wording differs from Berne Article 9(2), particularly in referring to the legitimate interests of rights holders rather than authors. Both provisions establish cumulative conditions, although their respective scope and language must be distinguished.

The WTO panel in United States – Section 110(5) of the US Copyright Act examined the compatibility of American statutory exemptions concerning the public communication of musical works. Its report, adopted on July 27, 2000, interpreted the three conditions and considered their application to two distinct exemptions. The dispute did not concern digital datasets, machine learning, or generative AI (WTO Panel, 2000).

TRIPS Articles 7 and 8 provide relevant interpretative context. They recognize technological innovation, the dissemination of knowledge, social and economic welfare, and certain public-interest measures. Their objectives do not independently authorize reproductions that fail to satisfy the Agreement's applicable limitations on exceptions.

The distinction between international obligations and private liability is essential. TRIPS binds WTO members, and its dispute settlement procedures address disputes between members. An alleged incompatibility between a national exception and Article 13 is not identical to a finding that a private developer has infringed copyright under domestic law. This distinction follows the broader principles governing the sources and operation of international law.

3.2 Certain Special Cases: Defining Permissible Training

The first condition requires copyright limitations and exceptions to be confined to certain special cases. In the WTO panel's interpretation, an exception must be clearly defined and sufficiently limited in scope and reach. The inquiry concerns the boundaries of the permitted use rather than simply whether legislators have identified a particular technological activity.

This requirement has direct implications for text and data mining legislation. An exception restricted to scientific research by specified institutions has relatively identifiable beneficiaries and purposes. A broader exception available to commercial AI developers may also be defined with legal precision, but its substantive scope requires closer examination.

Relevant considerations include the acts permitted, the categories of works covered, the circumstances in which copying may occur, and restrictions on retention or subsequent use. These conditions may establish the boundaries of an exception without requiring legislation to identify every individual application in advance.

The first condition does not establish that an exception is permissible merely because its wording is precise. A provision may remain insufficiently confined even where its terms are comprehensible. Equally, an exception does not necessarily fail solely because it accommodates commercial activity or a developing technology.

Lucchi, Dornis and Sierek (2026) question whether treating all generative AI training as a single permitted category provides sufficient specificity. Their analysis favors distinctions concerning the materials and uses involved. That position is an academic interpretation of the treaty requirement, not a judicial finding that every broad AI training exception violates TRIPS.

3.3 Normal Exploitation and the AI Licensing Market

The second condition prevents exceptions that conflict with the normal exploitation of protected works. Its application to generative AI is difficult because training can involve uses different from conventional reading, viewing, or listening while also creating commercial demand for access to copyrighted material.

In United States – Section 110(5), the WTO panel examined normal exploitation by reference to the exercise of exclusive rights to derive economic value from protected works. It recognized that the concept is narrower than every theoretically possible exercise of copyright. The analysis also considered forms of exploitation that have economic or practical significance or are likely to acquire it (WTO Panel, 2000, paras. 6.166–6.185).

The panel's interpretation arose from a dispute concerning music broadcasts. Applying its reasoning to AI training requires examining the economic characteristics of the disputed activity rather than assuming that the panel had already resolved the question. Markets for computational access may differ materially from markets for conventional distribution of protected expression.

Licensing arrangements for books, news content, photographs, and other materials demonstrate that training-data access can have commercial value. Yet the existence of a potential licensing opportunity does not automatically establish that every reproduction conflicts with normal exploitation. The nature of the works, the development of the relevant market, and the economic significance of the use require consideration.

Generated outputs introduce a related question. A model may produce material that does not reproduce recognizable protected expression, yet its outputs may compete with products or services offered by rights holders. Such competition is not automatically copyright infringement. Its possible relevance to the second step depends on the relationship between permitted training reproductions and the normal exploitation of protected works.

Lucchi, Dornis and Sierek (2026) argue for particular attention to licensing markets, direct substitution, and market dilution through functionally equivalent AI-generated content. Their proposed analysis considers safeguards against outputs that compete closely with protected works. These are arguments about the interpretation and application of existing treaty obligations, not conditions expressly written into Berne or TRIPS.

Evidence of lost training-data licensing revenue may support one argument, while evidence of substitution by generated content may support another. The second condition requires a legally grounded assessment of normal exploitation rather than an assumption that all economic competition involving artificial intelligence is prohibited.

3.4 Legitimate Interests and Unreasonable Prejudice

The third condition prohibits exceptions that unreasonably prejudice legitimate interests. Berne Article 9(2) refers to the legitimate interests of authors, whereas TRIPS Article 13 refers to those of rights holders. The test does not prohibit every adverse economic effect arising from a copyright exception.

In its analysis of TRIPS Article 13, the WTO panel considered the circumstances in which financial prejudice could become unreasonable. The interpretation recognizes a threshold: not every reduction in potential income establishes treaty incompatibility. The scale, character, and consequences of the permitted use influence the assessment (WTO Panel, 2000).

For generative AI training, relevant circumstances may include the amount of protected material reproduced, the purpose of training, the retention of copies, and the economic consequences of the permitted activity. Safeguards restricting redistribution, storage, or subsequent exploitation may reduce the prejudice associated with a particular exception.

Compensation mechanisms present another possibility. A statutory license, collective remuneration arrangement, or levy might reduce economic prejudice where an exception otherwise satisfies the international requirements. Their significance depends on the scope of permitted uses, the affected rights, and the conditions attached to compensation.

Payment cannot automatically remedy a failure to satisfy the second condition. An exception that conflicts with normal exploitation does not become compatible merely because rights holders receive financial compensation. All three conditions must be satisfied.

Lucchi, Dornis and Sierek (2026) argue that remuneration should address prejudice under the third step only after conflicts with normal exploitation have been excluded. Their approach also considers technical safeguards and the competitive characteristics of model outputs. This differs from proposals that would authorize broad training uses and attempt to address the resulting economic consequences entirely through compensation.

3.5 Competing Interpretations of the Three-Step Test

The scholarly disagreement over AI training concerns the scope of reproduction rights, the meaning of normal exploitation, and the conditions under which national exceptions can accommodate computational analysis. It cannot be resolved simply by characterizing one position as favorable to innovation and the other as favorable to copyright owners.

Senftleben (2026) advances a more permissive interpretation of the international framework. His argument questions whether reproductions undertaken exclusively for computational analysis necessarily fall within the international reproduction right. Under this interpretation, the application of the three-step test may depend on whether the activity is covered by that international right or by additional protection established under national or regional law.

Senftleben also examines the compatibility of text and data mining exceptions with the three-step test where that test applies. Rights reservations, substitution effects, and remuneration are relevant to his analysis. His position does not establish that every computational use is exempt from copyright or that commercial training always satisfies international requirements.

Lucchi, Dornis and Sierek (2026) place greater emphasis on the limitations imposed by the three-step test. Their analysis treats training-related reproductions as activities capable of engaging copyright protection and proposes differentiated conditions for permissible exceptions. The economic exploitation of works, competition from generated outputs, and protection against market displacement receive particular attention.

Their argument does not amount to a categorical prohibition on training copyrighted material. They identify circumstances in which exceptions may be permissible, including uses accompanied by appropriate safeguards or compensation. They also distinguish situations in which authorization is necessary because an exception would interfere with normal exploitation.

The competing interpretations lead to different conclusions about legislative discretion. Under a narrower understanding of international reproduction rights, certain computational operations may fall outside the treaty minimum. Under the broader approach, those operations engage protected rights, and national permissions must satisfy the applicable conditions.

Neither interpretation has been adopted as a universally binding resolution of generative AI training under Berne and TRIPS. The disagreement concerns the interpretation and application of established treaty obligations, not their existence. It explains why legislative approaches and domestic judicial decisions cannot independently settle the international limits of copyright exceptions.

4. National Exceptions for Copyrighted AI Training Data

International copyright treaties establish common standards, but national laws differ considerably in how they permit computational uses of protected works. These differences affect the legality of collecting training material, making intermediate copies, retaining datasets, and using copyrighted content for commercial model development.

The European Union has adopted express text and data mining exceptions. The United States applies its general fair use doctrine, while Japan and Singapore recognize provisions addressing computational analysis. The United Kingdom maintains a narrower exception for noncommercial research and has reconsidered proposals for broader reform.

These domestic approaches cannot be treated as interchangeable. Their conditions concerning lawful access, commercial purposes, rights reservations, contractual restrictions, and permitted reproductions vary. Their compatibility with international copyright obligations also raises questions distinct from whether a developer can rely on them in domestic proceedings.

4.1 European Union: Text and Data Mining Under EU Law

Directive (EU) 2019/790 on Copyright and Related Rights in the Digital Single Market establishes two principal exceptions for text and data mining. Article 3 concerns scientific research conducted by qualifying research organizations and cultural heritage institutions. Article 4 provides a broader exception for reproductions and extractions of lawfully accessible works and other protected subject matter.

Article 3 requires member states to permit specified reproductions and extractions for scientific research by eligible institutions. Copies must be stored with an appropriate level of security and may be retained for scientific purposes, including verification of research results. Rights holders may adopt necessary measures to protect the security and integrity of networks and databases, but those measures must remain within the statutory limits.

Article 4 is not confined to noncommercial research. It applies to qualifying text and data mining activities involving lawfully accessible material, provided rights holders have not expressly reserved the relevant uses in an appropriate manner. Reproductions and extractions made under the exception may be retained for as long as necessary for the permitted mining activity.

Lawful access and reservation of rights are separate requirements. Public availability may establish lawful access in certain circumstances, but it does not eliminate the possibility of an effective reservation under Article 4(3). For online content, the Directive identifies machine-readable means as an appropriate method of reserving rights. Recital 18 discusses metadata and website or service terms as relevant mechanisms (European Union, 2019).

Whether every form of generative AI training qualifies as text and data mining requires attention to Article 2(2). The provision defines text and data mining by reference to automated analytical techniques used to analyze digital text and data to generate information, including patterns, trends, and correlations. The legal characterization of training must consider the activities actually undertaken and the scope of the statutory permission.

In Like Company v. Google Ireland, Advocate General Maciej Szpunar expressed the view in November 2026 that training uses could generally fall within the Article 4 exception, subject to the relevant conditions. His observations were made in connection with preliminary questions he considered inadmissible. They do not constitute a binding judgment of the Court of Justice.

The EU AI Act adds obligations under Article 53 for providers of general-purpose AI models, including policies concerning Union copyright law and summaries of training content. These regulatory requirements do not replace the copyright exceptions established by Directive 2019/790. Their place within the wider regulatory system is addressed separately in EU AI Act Explained.

4.2 United States: Fair Use and Commercial AI Training

United States copyright law contains no general statutory exception expressly authorizing all generative AI training. Instead, many disputes are assessed under the fair use doctrine codified in 17 U.S.C. § 107. The doctrine requires consideration of the circumstances of the particular use rather than applying a categorical permission to a named technology.

Section 107 identifies four factors: the purpose and character of the use, the nature of the copyrighted work, the amount and substantiality copied, and the effect on the potential market for or value of the work. Courts assess the factors together. Commercial use is relevant but does not automatically exclude a finding of fair use.

The purpose of training is a central consideration. A developer may argue that copying works to identify statistical relationships serves a function different from distributing or consuming their expressive content. Rights holders may argue that the copying supports commercially competing products and interferes with licensing or other markets for protected works.

The U.S. Copyright Office examined these questions in Copyright and Artificial Intelligence, Part 3: Generative AI Training, released in pre-publication form on May 9, 2025. The report concluded that several stages of generative AI development involve uses implicating exclusive copyright rights and that fair use is central to assessing whether those uses are permitted (U.S. Copyright Office, 2025).

The report distinguishes among training activities rather than treating every use of copyrighted material as legally equivalent. It recognizes that some uses may be transformative and permissible, while commercial copying capable of producing competing expressive material may weigh against fair use. Its analysis considers the circumstances of acquisition, the purpose of copying, the character of generated outputs, and potential market effects.

The Copyright Office's report is an institutional interpretation and policy assessment, not a judicial ruling. Courts determine fair use in disputes within their jurisdiction. The decisions concerning Anthropic, Meta, and ROSS illustrate why different technologies and evidentiary records can produce different outcomes under the same statutory doctrine.

American fair use also cannot be equated directly with the international three-step test. Section 107 provides an open-ended, fact-sensitive inquiry, whereas TRIPS Article 13 expresses the conditions governing permissible national limitations and exceptions. The relationship between them is an issue of international treaty compatibility, not proof that every domestic fair use ruling necessarily determines the international position.

4.3 Japan and Singapore: Computational Analysis Exceptions

Japan's Copyright Act contains a provision specifically relevant to computational uses. Article 30-4 permits certain uses of works where the purpose is not to enjoy, or cause another person to enjoy, the thoughts or sentiments expressed in the work. The provision accommodates activities such as information analysis, subject to statutory qualifications.

The distinction between enjoyment and non-enjoyment is central. Using a work to extract information for computational processing may qualify where the protected expression is not being used for its expressive purpose. A training activity that also seeks to reproduce or provide enjoyment of the original expression may require a different legal assessment.

Article 30-4 excludes uses that would unreasonably prejudice the copyright owner's interests, considering the nature or purpose of the work and the circumstances of its exploitation. The Japan Agency for Cultural Affairs' May 2024 General Understanding on AI and Copyright in Japan discusses this limitation, including the possible significance of markets for analysis-ready materials and uses designed to generate expression closely resembling protected works.

The Agency's guidance is not legally binding and does not establish the result of every AI copyright dispute. It provides an institutional interpretation of existing Japanese law while recognizing that specific technical practices and economic circumstances may change the assessment (Agency for Cultural Affairs, 2024).

Singapore's Copyright Act 2021 adopts a more express statutory treatment of computational data analysis. Section 243 includes the use of computer programs to identify, extract, and analyze information, as well as using works as examples to improve a program's operation. Its statutory illustration specifically addresses the use of images to train a program to recognize images.

Section 244 permits copying for computational data analysis or preparation for that analysis, subject to conditions including lawful access. Copies may not be used for unrelated purposes, and their supply to others is restricted to specified circumstances, including verification of results and qualifying collaborative research or study.

Singapore also limits contractual restrictions on the permitted activity. Section 187 renders contractual terms void insofar as they purport to exclude or restrict computational data analysis uses that the Act permits. This does not remove the separate statutory requirements, including lawful access to the material from which the copy is made.

Japan and Singapore consequently accommodate computational analysis through different legislative techniques. Japan emphasizes the purpose for which protected expression is used and the risk of unreasonable prejudice. Singapore defines computational analysis expressly and establishes conditions for reproductions undertaken for that purpose. Neither system supplies unconditional permission to reproduce all copyrighted content obtained from any source.

4.4 United Kingdom: Research Exceptions and Reform

The United Kingdom's principal text and data mining exception appears in section 29A of the Copyright, Designs and Patents Act 1988. It permits copying by a person with lawful access to a work where the copy is made for computational analysis solely for noncommercial research. The provision also requires sufficient acknowledgment unless that is impractical or otherwise impossible.

The exception places restrictions on subsequent use. Copies made under section 29A cannot be transferred to another person or used for another purpose without authorization from the copyright owner. Contractual terms attempting to prevent or restrict copying permitted by the section are unenforceable to that extent.

Its scope is narrower than a general exception for commercial generative AI development. An AI company cannot rely on section 29A merely because its activity involves computational analysis. Whether a particular use qualifies depends on the statutory conditions, including the required noncommercial research purpose.

In December 2024, the government launched a consultation examining possible changes to copyright law and AI training. Its initially preferred option was a broader exception with a mechanism allowing rights holders to reserve their works against training uses. That proposal attracted substantial opposition from creative industries and criticism concerning implementation.

The government's March 18, 2026 Report on Copyright and Artificial Intelligence stated that a broad exception with an opt-out was no longer its preferred approach. It identified continuing gaps in the evidence and proposed further examination of alternatives. The accompanying impact assessment did not establish a replacement preferred option (United Kingdom Government, 2026).

The policy change did not itself amend section 29A. Current statutory permissions must remain separate from consultations, impact assessments, and proposed reforms. The United Kingdom illustrates how the economic interests of rights holders and AI developers can influence legislative choices without determining the content of existing law.

5. Judicial Interpretation of AI Training and Copyright

National courts have begun to distinguish the reproduction of copyrighted works for AI development from the reproduction of protected expression in trained models and generated outputs. The disputes concern different technologies, types of material, statutory rights, and exceptions. Their outcomes cannot be reduced to a general judicial conclusion that commercial AI training is always lawful or unlawful.

The legal significance of a decision also depends on its procedural status. A ruling on summary judgment may resolve particular issues without completing every aspect of litigation. An interlocutory determination may address the evidence available at an early stage, while a settlement resolves claims without necessarily establishing judicial findings concerning their merits.

Recent decisions in the United States, Germany, India, and the United Kingdom illustrate these distinctions. They also show that courts may reach different conclusions when applying domestic copyright provisions to training datasets, model memorization, and AI-generated content.

5.1 United States: Fair Use, Piracy, and Market Harm

In Bartz v. Anthropic, the United States District Court for the Northern District of California considered the acquisition and use of copyrighted books in developing large language models. In June 2025, Judge William Alsup concluded that the particular use of books for training qualified as transformative fair use. The ruling distinguished that activity from the acquisition and retention of pirated books in a permanent digital library.

The distinction was decisive. The court treated training and library acquisition as legally separable activities, rather than assuming that a favorable fair use conclusion for one necessarily extended to the other. Copies acquired through piracy presented different legal questions from copies lawfully obtained and reproduced for the training activities addressed by the ruling.

The litigation subsequently proceeded toward settlement. On July 20, 2026, Judge Araceli Martínez-Olguín granted final approval of the class action settlement concerning pirated-source books and entered judgment. The settlement resolved the covered claims; it did not establish a general judicial authorization for commercial training using copyrighted books (Bartz v. Anthropic, 2025; 2026).

A different evidentiary question arose in Kadrey v. Meta Platforms. In June 2025, Judge Vince Chhabria granted Meta's motion for partial summary judgment concerning the fair use issues presented. The decision emphasized weaknesses in the plaintiffs' showing of relevant market harm and expressly recognized that other claims supported by stronger evidence might produce different outcomes.

The finding did not establish that training generative models on copyrighted literary works is invariably fair use. Its reasoning depended on the claims, evidence, and statutory factors before the court. The decision illustrates the significance of demonstrating actual or reasonably supported potential market consequences rather than relying solely on assertions that AI development threatens authors' economic interests (Kadrey v. Meta Platforms, 2025).

The United States Court of Appeals for the Third Circuit addressed another form of AI-related copying in Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc. In an opinion filed on September 29, 2026, the court affirmed partial summary judgment holding that ROSS's copying of protected Westlaw headnotes was not fair use. The appeal concerned the originality of the headnotes and the application of the statutory fair use doctrine.

ROSS developed an AI-powered legal research system using training materials derived from Westlaw headnotes. The system was designed to identify relevant passages in judicial opinions in response to legal questions. It was not a conventional generative language model producing new expressive content.

The Third Circuit found that the protected editorial headnotes contained sufficient originality and that ROSS had used them for a highly similar commercial purpose. The court considered the limited transformative character of the copying, the substantial material taken, and the effects on relevant markets. Its reasoning reflected the specific relationship between the competing legal research products (Thomson Reuters v. ROSS Intelligence, 2026).

These decisions demonstrate the importance of distinguishing technical purposes, methods of acquisition, and market relationships. Bartz addressed particular training uses and separately treated pirated-book acquisition. Kadrey turned substantially on the evidentiary record concerning market harm. ROSS involved copying protected editorial material to develop a directly competing legal research service.

None establishes a universal rule under Berne or TRIPS. The judgments apply American copyright law to the disputes before the courts, and their reasoning must be assessed in relation to the particular works and activities involved.

5.2 European Decisions: Reproduction and Model Outputs

European litigation has addressed whether protected material can be reproduced within trained AI models and whether generated responses infringe copyright or related rights. These issues differ from the initial copying used to assemble training datasets. A copyright exception permitting preparatory computational reproductions does not necessarily determine the legality of protected expression retained or reproduced later.

In GEMA v. OpenAI, the Regional Court of Munich I issued a judgment on November 11, 2025, concerning protected German song lyrics. The proceedings involved allegations that lyrics represented by the collecting society GEMA had been memorized within language models and reproduced through chatbot responses. The court substantially upheld the copyright-related claims, while rejecting the separate claims concerning general personality rights.

The court found that the disputed lyrics were reproducibly contained in the models and could be extracted through generated responses. It considered this legally relevant reproduction under German copyright law and Article 2 of Directive 2001/29/EC. The findings concerned identifiable protected lyrics and the evidence presented, rather than a presumption that every model reproduces every work incorporated into training.

The judgment also distinguished preparatory text and data mining from the retention of protected expression within model parameters. The court accepted that necessary preparatory reproductions, including format conversion and temporary storage, can fall within the German text and data mining provisions. It concluded that the reproductions established within the models were not covered by those exceptions because they went beyond the extraction of information contemplated by the statutory framework.

The court additionally found that the reproduction of protected lyrics in chatbot outputs infringed the relevant exploitation rights. Its reasoning addressed the defendants' responsibility for the models and the generation of protected material following ordinary user requests. The judgment was a first-instance decision and was not final when issued (Regional Court of Munich I, Case 42 O 14139/24, 2025).

A separate European development occurred on November 3, 2026, when Advocate General Maciej Szpunar delivered his opinion in Like Company v. Google Ireland, Case C-250/25. The preliminary reference arose from litigation involving the Google Gemini chatbot and the rights of a Hungarian press publisher. The proceedings principally concern the related right established for press publishers by Article 15 of Directive 2019/790.

The referring court raised questions concerning the alleged use of press publications in model training and their subsequent retrieval or reuse in generated answers. Advocate General Szpunar considered the training-related questions hypothetical and recommended that they be declared inadmissible. His assessment rested partly on the absence of a sufficient factual basis establishing that the claimant's publications had been used in training.

In additional reasoning, the Advocate General expressed the view that training of large language models could generally fall within the text and data mining exception under Article 4 of Directive 2019/790, subject to its conditions. He separately considered whether real-time retrieval and reuse of press publications could infringe publishers' rights. That distinction reflects the different legal treatment of training operations and the subsequent use of protected publications.

An Advocate General's opinion is not a judgment of the Court of Justice. As of November 6, 2026, the Court had not delivered its ruling in Like Company. The opinion provides a significant interpretation of EU copyright and related rights, but its observations concerning model training do not establish a binding conclusion that every generative AI training activity satisfies Article 4.

The English High Court addressed different questions in Getty Images v. Stability AI, decided on November 4, 2025. The litigation originally included allegations concerning the unauthorized use of protected photographs in developing Stable Diffusion. Getty abandoned its primary training-related copyright claim after acknowledging the absence of evidence that the relevant training and development had occurred in the United Kingdom.

The claim concerning infringing outputs and the database-right claim were also abandoned, leaving issues including secondary copyright infringement, trademarks, and passing off for determination. The judgment did not decide that the foreign training activity was lawful under every potentially applicable legal system. Its significance for training-related copyright claims lies particularly in the importance of territorial evidence and the precise conduct alleged (Getty Images v. Stability AI, [2025] EWHC 2863 (Ch)).

An additional approach emerged in India. In ANI Media Pvt. Ltd. v. OpenAI OpCo LLC, the Delhi High Court considered allegations concerning the use of news content to train language models and reproduce material in generated responses. In an interlocutory judgment delivered on July 24, 2026, Justice Amit Bansal concluded, on a prima facie assessment, that the training-related storage addressed by the proceedings fell within the fair-dealing provision of section 52(1)(a) of the Indian Copyright Act 1957.

The court considered the purpose of the activity, its fairness, potential economic competition, and public-interest considerations. It also identified limitations in the factual evidence concerning the alleged training of particular published articles. The court declined interim injunctive relief, but the decision did not finally resolve all merits questions (ANI Media v. OpenAI, 2026).

These proceedings demonstrate divergent approaches within domestic copyright systems. The German judgment emphasizes identifiable reproductions retained within models and generated outputs. The Indian ruling gives considerable weight, at the interlocutory stage, to research-related fair dealing and public-interest considerations. The European Advocate General's opinion addresses the interpretation of EU text and data mining law without yet constituting a binding judgment.

Their differences cannot be reconciled simply by choosing one jurisdiction as the correct model for international copyright law. They concern distinct statutory provisions, protected rights, technical activities, and procedural questions. Berne and TRIPS provide the international standards against which national rules may be assessed, but these decisions do not themselves settle the international interpretation of the three-step test.

6. Cross-Border Training and Copyright Authorization

Generative AI development frequently involves copyrighted works acquired, stored, and processed across several jurisdictions. A company may collect material from websites hosted abroad, process datasets through distributed computing infrastructure, and offer the resulting model internationally. The technical integration of these activities does not eliminate the territorial character of copyright protection.

Different stages of the same development process may engage different legal questions. The acquisition of a protected work, the creation of a training copy, the retention of material within a model, and the distribution of generated content cannot always be treated as one legally indivisible activity. The applicable rights and exceptions must be identified for the particular conduct concerned.

Authorization is equally important. A developer's right to access material does not necessarily include permission to reproduce it for commercial training. Licensing arrangements may provide such authorization, while statutory exceptions may permit particular uses without individual consent. The legal effect depends on the governing law and the conditions attached to the permission.

6.1 Territoriality and the Location of Copyright Infringement

The international protection of copyright operates primarily through national legal systems. Berne establishes common standards and national treatment, but it does not replace territorial rights with a single worldwide copyright entitlement. This is particularly significant where alleged reproductions occur outside the jurisdiction in which legal protection is sought.

A foreign website's accessibility does not establish that every reproduction of its contents is lawful. Equally, the presence of a copyrighted work in a dataset used abroad does not automatically establish infringement under the law of every country where the resulting model becomes available. The relevant rights, conduct, territorial connections, and applicable-law rules require separate consideration.

The distinction between training location and model deployment is illustrated by Getty Images v. Stability AI. The abandonment of the training-related copyright claim in the English proceedings reflected the absence of evidence that the relevant training occurred within the United Kingdom. That procedural development did not establish the legality of the same conduct under other national copyright systems.

Distributed processing introduces additional factual difficulties. Copies may be created on servers in different countries, transferred between service providers, or retained temporarily during computational operations. The legal significance of those activities depends on which protected acts occur and which laws govern them.

These issues overlap with the wider regulation of cross-border data flows in international law, although copyright has its own protected subject matter and infringement standards. Privacy, contractual obligations, data protection, and international trade law may govern aspects of the same operation without determining whether a copyrighted work has been reproduced unlawfully.

The EU AI Act adds obligations for providers making general-purpose AI models available within the Union, including copyright compliance policies. Its market-access framework does not automatically establish that every reproduction undertaken abroad infringes substantive EU copyright law. The territorial operation of copyright rights and the regulatory application of the AI Act involve distinct legal questions.

6.2 Licensing, Rights Reservations, and Enforcement

Licensing provides a means of obtaining authorization for uses that would otherwise require copyright-owner consent. An agreement may permit the reproduction of specified works, their incorporation into training datasets, and their retention or processing for defined purposes. The scope of authorization depends on the rights controlled by the licensor and the conditions of the agreement.

Large training corpora create difficulties because they may contain works controlled by numerous authors, publishers, collecting societies, and other rights holders. Ownership information can be incomplete, and different materials may be subject to different contractual terms. Permission to use one collection does not establish authorization for every work incorporated into a wider dataset.

Collective licensing and standardized agreements may reduce transaction costs where organizations have authority to license the relevant rights. Their effectiveness depends on the rights represented, the territorial scope of the authorization, and the ability to identify covered uses. Neither Berne nor TRIPS establishes a universal licensing mechanism for AI training.

Rights reservations provide another means of controlling computational uses where domestic legislation recognizes them. Under Article 4 of Directive 2019/790, an appropriate express reservation prevents reliance on that exception for the reserved use. This operates differently from Japan's non-enjoyment exception, Singapore's computational analysis provisions, and American fair use.

Contractual conditions also vary in effect. Singapore's Copyright Act restricts attempts to contract out of specified permitted uses, while the United Kingdom protects copying that satisfies section 29A against contrary contractual terms. These provisions do not remove the separate statutory conditions attached to the permissions. In the European Union, the Article 4 rights-reservation mechanism must be distinguished from a contractual prohibition outside the scope of that exception.

Enforcement can be difficult where rights holders lack information about training datasets. A generated response resembling protected expression may provide evidence relevant to reproduction, but it does not necessarily establish the precise manner in which the material was obtained, copied, or retained. Conversely, an inability to identify copyrighted material within a model's outputs does not conclusively establish that no protected works were reproduced during training.

Training-data records, provenance information, contractual documentation, and relevant technical evidence may assist in determining which activities occurred. Disclosure obligations can improve access to information without automatically establishing infringement. The legal assessment still requires identification of protected rights, relevant acts, and any applicable authorization or exception.

Compensation must also be distinguished from permission. A voluntarily negotiated licensing fee reflects contractual authorization, whereas a statutory remuneration mechanism depends on legislation. Proposals for levies or collective compensation should not be presented as obligations already imposed by international copyright treaties.

7. WIPO and International Copyright Cooperation

The World Intellectual Property Organization provides an institutional forum for examining copyright issues arising from new technologies. Its Standing Committee on Copyright and Related Rights brings together member states and participating organizations to consider copyright protection, limitations and exceptions, and possible international cooperation. The Committee's discussions do not themselves create new treaty obligations.

On October 23, 2026, WIPO published the Scoping Study on AI Training and Copyright, document SCCR/49/6, as material associated with the Standing Committee's forty-ninth session. The session is scheduled to take place in Geneva from November 30 to December 4, 2026. As of November 6, the meeting has not occurred, and no outcome can be attributed to those forthcoming proceedings.

The publication of the study places the relationship between AI training and copyright within a formal international institutional process. The study is not an international treaty, an adopted interpretation of Berne or TRIPS, or a binding decision concerning the legality of training. Its institutional status must remain distinct from the legal force of existing copyright instruments.

International cooperation faces several substantive questions. States must consider the protection afforded to digital reproductions, the conditions under which computational analysis may be exempted from exclusive rights, and the effect of national exceptions on international licensing markets. These questions involve existing treaty provisions as well as choices about future legislation.

Greater convergence could reduce legal uncertainty for developers and rights holders operating across borders. Common approaches to licensing information, rights reservations, and dataset documentation might also facilitate lawful transactions. Yet differences between domestic copyright systems reflect distinct legislative decisions that cannot be removed without agreement on the relevant substantive standards.

The arguments for reform also differ. Rights holders emphasize the economic exploitation of protected works, control over reproductions, and the ability to obtain remuneration. AI developers and researchers emphasize access to information, computational analysis, and predictable legal permissions. International copyright law requires these interests to be addressed through the applicable treaty obligations rather than through an assumption that one category of interest always prevails.

Proposals concerning remuneration, transparency, and technical safeguards may influence future national legislation or international discussions. Their legal effect depends on whether they are adopted through the appropriate instruments and institutions. WIPO discussions should not be confused with binding obligations already established under Berne, TRIPS, or the WIPO Copyright Treaty.

Copyright governance also differs from the broader discussion of artificial intelligence ethics in international law. Ethical principles may influence policy, but they do not independently confer reproduction rights or establish statutory exceptions. Copyright protection requires analysis of identifiable rights, legal obligations, and permitted uses.

The principal international question remains the extent of national legislative discretion. Existing treaties provide minimum standards and limitations on exceptions. Their application to generative AI training is contested, but disagreement over interpretation does not eliminate the binding obligations that states have accepted.

Conclusion

Copyrighted works may be used to train artificial intelligence systems where the relevant activities are authorized, fall outside the applicable exclusive rights, or satisfy a valid exception under governing domestic law. International copyright law does not establish an unrestricted right to train AI systems using protected material, nor does it categorically prohibit every form of computational analysis.

The Berne Convention protects the reproduction right and permits carefully limited exceptions. TRIPS incorporates the relevant Berne obligations and imposes the three-step test through Article 13. The WIPO Copyright Treaty confirms the application of reproduction rights to digital storage while recognizing the possibility of appropriately framed exceptions for the digital environment.

The principal unresolved questions concern the scope of reproduction during computational processing, the compatibility of broad AI training exceptions with normal exploitation, and the economic consequences of using protected works to develop commercial models. National laws and judicial decisions provide different answers to particular disputes, but none establishes a uniform international permission or prohibition.

The distinction between international treaty obligations and domestic copyright liability remains decisive. A state may authorize particular computational uses through national legislation while facing separate questions concerning the international compatibility of its exceptions. A developer must rely on the law governing its activities rather than treating international copyright treaties as an independent authorization to reproduce protected works.

Existing international law supplies substantive constraints on national copyright policy. What remains contested is how those constraints apply to the technical operations, licensing markets, and commercial consequences of generative AI training. That distinction identifies both the protection already required by Berne and TRIPS and the questions that national courts, legislatures, and international institutions have yet to resolve.

References

Agency for Cultural Affairs (2024) General Understanding on AI and Copyright in Japan: Overview. Tokyo: Japan Copyright Office, Government of Japan. Available at: https://www.bunka.go.jp/english/policy/copyright/pdf/94055801_01.pdf (Accessed: 6 November 2026).

Agreement on Trade-Related Aspects of Intellectual Property Rights (TRIPS Agreement) (1994) Annex 1C to the Marrakesh Agreement Establishing the World Trade Organization, adopted 15 April 1994, entered into force 1 January 1995. Available at: https://www.wto.org/english/docs_e/legal_e/27-trips_01_e.htm (Accessed: 6 November 2026).

Berne Convention for the Protection of Literary and Artistic Works (1886) Adopted 9 September 1886, revised at Paris on 24 July 1971 and amended 28 September 1979. Available at: https://www.wipo.int/wipolex/en/text/283698 (Accessed: 6 November 2026).

Copyright Act 1957 (India) Act No. 14 of 1957, enacted 4 June 1957, as amended. Available at: https://www.wipo.int/wipolex/en/legislation/details/22949 (Accessed: 6 November 2026).

Copyright Act 1970 (Japan) Act No. 48 of 6 May 1970, as amended. Available at: https://www.wipo.int/wipolex/en/legislation/details/24012 (Accessed: 6 November 2026).

Copyright Act 2021 (Singapore) Act No. 22 of 2021. Available at: https://sso.agc.gov.sg/Acts-Supp/22-2021/Published/ (Accessed: 6 November 2026).

Copyright Act of 1976 (United States) Public Law 94-553, 19 October 1976, 90 Stat. 2541, codified at 17 U.S.C. § 107 (Fair Use), as amended. Available at: https://uscode.house.gov/us/usc/119-118/t17/s107/ (Accessed: 6 November 2026).

Copyright, Designs and Patents Act 1988 (United Kingdom) 1988 c. 48, 15 November 1988, section 29A, as amended. Available at: https://www.legislation.gov.uk/ukpga/1988/48/section/29A (Accessed: 6 November 2026).

Court of Justice of the European Union (2026) Like Company v. Google Ireland Limited, Case C-250/25, Opinion of Advocate General Maciej Szpunar, 3 November 2026, ECLI:EU:C:2026:919. Available at: https://infocuria.curia.europa.eu/tabs/redirect/juris/documents.jsf?num=C-250%2F25 (Accessed: 6 November 2026).

Department for Science, Innovation and Technology and Department for Culture, Media and Sport (2026) Report on Copyright and Artificial Intelligence. London: UK Government, ISBN 978-1-5286-6308-3. Available at: https://www.gov.uk/government/publications/report-and-impact-assessment-on-copyright-and-artificial-intelligence/report-on-copyright-and-artificial-intelligence (Accessed: 6 November 2026).

Department for Science, Innovation and Technology and Department for Culture, Media and Sport (2026) Copyright and Artificial Intelligence: Impact Assessment. London: UK Government, ISBN 978-1-5286-6317-5. Available at: https://www.gov.uk/government/publications/report-and-impact-assessment-on-copyright-and-artificial-intelligence/copyright-and-artificial-intelligence-impact-assessment (Accessed: 6 November 2026).

Directive 96/9/EC of the European Parliament and of the Council (1996) On the legal protection of databases, 11 March 1996, Official Journal of the European Communities, L 77, 27 March, pp. 20–28. Available at: https://eur-lex.europa.eu/eli/dir/1996/9/oj (Accessed: 6 November 2026).

Directive 2001/29/EC of the European Parliament and of the Council (2001) On the harmonisation of certain aspects of copyright and related rights in the information society, 22 May 2001, Official Journal of the European Communities, L 167, 22 June, pp. 10–19. Available at: https://eur-lex.europa.eu/eli/dir/2001/29/oj (Accessed: 6 November 2026).

Directive (EU) 2019/790 of the European Parliament and of the Council (2019) On copyright and related rights in the Digital Single Market and amending Directives 96/9/EC and 2001/29/EC, 17 April 2019, Official Journal of the European Union, L 130, 17 May, pp. 92–125. Available at: https://eur-lex.europa.eu/eli/dir/2019/790/oj (Accessed: 6 November 2026).

High Court of Delhi (2026) ANI Media Pvt. Ltd. v. Open AI OpCo LLC, judgment on application for interim injunction, 24 July 2026, corrected 27 July 2026, CS(COMM) 1028/2024, I.A. 45300/2024. Available at: https://www.wipo.int/wipolex/en/judgments/details/3893 (Accessed: 6 November 2026).

High Court of England and Wales (2025) Getty Images (US) Inc and Others v. Stability AI Limited, judgment, 4 November 2025, [2025] EWHC 2863 (Ch). Available at: https://www.judiciary.uk/judgments/getty-images-v-stability-ai/ (Accessed: 6 November 2026).

Landgericht München I (2025) GEMA v. OpenAI, judgment, 11 November 2025, Case 42 O 14139/24, GRUR-RS 2025, 30204. Available at: https://www.gesetze-bayern.de/Content/Document/Y-300-Z-GRURRS-B-2025-N-30204 (Accessed: 6 November 2026).

Lucchi, N., Dornis, T.W. and Sierek, P.T. (2026) ‘The three-step test in international copyright—a global framework for generative AI training’, American Business Law Journal, 63(3), pp. 265–298. Available at: https://doi.org/10.1111/ablj.70022 (Accessed: 6 November 2026).

Regulation (EU) 2024/1689 of the European Parliament and of the Council (2024) Laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), 13 June 2024, Official Journal of the European Union, L, 2024/1689, 12 July. Available at: https://eur-lex.europa.eu/eli/reg/2024/1689/oj (Accessed: 6 November 2026).

Senftleben, M. (2026) ‘Text and data mining, generative AI, and the copyright three-step test’, IIC – International Review of Intellectual Property and Competition Law, 57, pp. 67–107. Available at: https://doi.org/10.1007/s40319-026-01680-2 (Accessed: 6 November 2026).

United Kingdom Government (2024) Copyright and Artificial Intelligence: Consultation. London: Intellectual Property Office, Department for Science, Innovation and Technology and Department for Culture, Media and Sport, CP 1205. Available at: https://www.gov.uk/government/consultations/copyright-and-artificial-intelligence (Accessed: 6 November 2026).

United States Court of Appeals for the Third Circuit (2026) Thomson Reuters Enterprise Centre GmbH and West Publishing Corp. v. ROSS Intelligence Inc., opinion, 29 September 2026, No. 25-2153. Available at: https://www2.ca3.uscourts.gov/opinarch/252153p.pdf (Accessed: 6 November 2026).

United States Copyright Office (2025) Copyright and Artificial Intelligence, Part 3: Generative AI Training. Pre-publication version, 9 May 2025. Washington, DC: U.S. Copyright Office. Available at: https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf (Accessed: 6 November 2026).

United States District Court for the Northern District of California (2025) Bartz et al. v. Anthropic PBC, order on fair use, 23 June 2025, Case No. 24-cv-05417-WHA, Docket No. 231. Available at: https://docs.justia.com/cases/federal/district-courts/california/candce/3%3A2024cv05417/434709/231 (Accessed: 6 November 2026).

United States District Court for the Northern District of California (2025) Kadrey et al. v. Meta Platforms, Inc., order on cross-motions for partial summary judgment, 25 June 2025, Case No. 23-cv-03417-VC, Docket No. 598. Available at: https://law.justia.com/cases/federal/district-courts/california/candce/3%3A2023cv03417/415175/598/ (Accessed: 6 November 2026).

United States District Court for the Northern District of California (2026) Bartz et al. v. Anthropic PBC, order granting final approval of class action settlement and judgment, 20 July 2026, Case No. 24-cv-05417-AMO, Docket No. 680. Available at: https://docs.justia.com/cases/federal/district-courts/california/candce/4%3A2024cv05417/434709/680 (Accessed: 6 November 2026).

Vienna Convention on the Law of Treaties (1969) Concluded 23 May 1969, entered into force 27 January 1980, 1155 UNTS 331. Available at: https://treaties.un.org/pages/showdetails.aspx?objid=080000028003902f (Accessed: 6 November 2026).

WIPO Copyright Treaty (WCT) (1996) Adopted at Geneva on 20 December 1996, entered into force 6 March 2002, including agreed statements concerning the treaty. Available at: https://www.wipo.int/wipolex/en/text/295166 (Accessed: 6 November 2026).

World Intellectual Property Organization (2026) Scoping Study on AI Training and Copyright, SCCR/49/6, 23 October 2026. Geneva: WIPO. Available at: https://www.wipo.int/meetings/en/doc_details.jsp?doc_id=665485 (Accessed: 6 November 2026).

World Intellectual Property Organization (n.d.) Standing Committee on Copyright and Related Rights (SCCR) [online]. Available at: https://www.wipo.int/en/web/sccr (Accessed: 6 November 2026).

World Trade Organization, Dispute Settlement Panel (2000) United States – Section 110(5) of the US Copyright Act, Report of the Panel, 15 June 2000, WT/DS160/R, adopted 27 July 2000. Available at: https://www.wto.org/english/tratop_e/dispu_e/cases_e/ds160_e.htm (Accessed: 6 November 2026).