When a US federal court ruled in 2025 that Anthropic could train its AI models on copyrighted books without permission, it never asked whether that conclusion complied with the Berne Convention, the copyright treaty binding more than 180 countries. The omission was not a minor procedural detail. It revealed a structural gap between domestic copyright exceptions and international obligations that affects every writer, photographer and publisher whose work has been scraped into a training dataset.

The Berne Convention's Article 9 grants authors the exclusive right to authorise reproduction of their works. Article 9(2) allows limited exceptions, but only if they satisfy a three-step test: the exception must apply to a "certain special case," it must not conflict with normal exploitation of the work, and it must not unreasonably prejudice the author's legitimate interests. All three conditions must hold simultaneously.

AI training necessarily involves reproduction. When a model ingests a copyrighted work, that work is stored and copied in a dataset. The Agreed Statements accompanying the WIPO Copyright Treaty confirm that electronic storage of a protected work constitutes reproduction. The relevant legal question is not whether the model's output resembles the original, as AI companies prefer to argue. It is whether the input process, the mass copying of works into training data, is authorised.

Why US fair use fails the Berne test

The Bartz v. Anthropic decision applied the fair use doctrine to AI training and found Anthropic's use "exceedingly transformative." Fair use considers four factors, including the purpose of the use and its effect on the market. The court did not examine whether its interpretation satisfies Berne's three-step test. Under US law, it had no obligation to do so: the Berne Convention is not self-executing, and courts apply domestic statutes rather than treaty text.

A broad fair use finding for commercial AI training fails all three steps. A WTO panel ruling in 2000 on US copyright exceptions interpreted "certain special cases" to mean exceptions that are clearly defined and narrow in scope. Fair use can apply to virtually any purpose across any industry. It is not limited to AI training. It is the opposite of a special case.

The second step is equally problematic. A licensing market for AI training already exists. The Associated Press licensed parts of its archive to OpenAI in 2023. The New York Times sued OpenAI the same year for training on its journalism without permission. When rightsholders are willing to license and prepared to litigate, that market is not hypothetical. Fair use allows AI companies to bypass it entirely.

The third step asks whether prejudice to authors is proportionate to the public interest served. AI models trained on copyrighted works can substitute for originals across research, education and entertainment, reducing the aggregate market value of entire categories of creative work. The harm is systemic. Mass unpaid copying for commercial development is difficult to square with proportionate prejudice.

Japan's non-enjoyment loophole

Japan's Article 30-4 of the Copyright Act permits using copyrighted works for data analysis when the purpose is "non-enjoyment," meaning the work is processed as data rather than consumed for its expressive content. The law excludes training that analyses data about a specific author, an attempt to narrow the exception for Berne's first step.

In practice, most AI training qualifies as non-enjoyment, making the exception so broad it undermines the rule. Whether a work is being "enjoyed" does not address the economic question at the heart of steps two and three: the author loses the opportunity to license the work regardless of the developer's stated purpose.

Japan's statute says use is not permitted when it would "unreasonably prejudice" the rightsholder. But it creates no general remuneration right and imposes no consent requirement. Individual authors face a structural proof problem: the harm from AI training is cumulative and spread across millions of works. The proof requirement is individual, but the injury is collective.

The EU's opt-out advantage

The EU's Directive on Copyright in the Digital Single Market permits text and data mining on lawfully accessible works for any purpose, but rightsholders can opt out. If they do, developers must either avoid the work or negotiate a licence. This satisfies Berne's first step because it is specific to text and data mining, not a general-purpose exception. It satisfies the second because the opt-out mechanism preserves a licensing market rather than undermining it. It satisfies the third because it gives authors control before their works are used, rather than requiring proof of harm after the fact.

The weakness is practical. Opting out requires authors to know where their works are being used and to implement machine-readable rights reservations. Individual creators, especially those outside the technology sector or in developing countries, may lack the resources to do so. The protection is formally available but practically inaccessible to many rightsholders.

A treaty built for a different era

Article 9(2) was drafted long before generative AI existed. The WIPO Copyright Treaty, adopted as a special agreement under Berne to address digital copyright, offers a precedent for treaty-level responses to technological change. A comparable agreement could clarify how the three-step test applies to AI training, recognise AI licensing markets as part of normal exploitation, and set minimum standards on access, rights reservations and compensation.

The largest obstacle is the United States. The world's biggest AI companies are American, and Washington has a record of resisting international copyright obligations that conflict with its fair use doctrine. If the US declined to join a new WIPO agreement, it would govern the use of copyrighted works everywhere except the jurisdiction where most AI training occurs.

Market pressure may be the only effective lever. EU adequacy requirements, bilateral trade provisions, and the regulatory gravity that has already pulled jurisdictions like Brazil toward EU standards could force convergence. Without it, the reproduction right the Berne Convention guarantees will continue to depend on where an AI company is headquartered rather than where a copyrighted work was created.

Organisations

Berne Convention · World Intellectual Property Organization · World Trade Organization · OpenAI · Anthropic · Associated Press