Can Copyright Exceptions and Fair Use Defences Justify Unlicensed AI Training Under Intellectual Property Law?

In today’s data-hungry AI landscape, one of the most pressing legal questions is whether copyright and related rights can—or should—bend to accommodate unlicensed training of artificial intelligence models, particularly when those activities are framed as being for “research” or “learning.”

A new academic study by Professor Eleonora Rosati, published in the European Journal of Risk Regulation, offers a thorough, cross-jurisdictional investigation into this issue. Commissioned by the International Federation of the Phonographic Industry (IFPI), the study critically assesses whether existing intellectual property (IP) exceptions—especially those for research, teaching, or education—can legitimately be used as legal cover for training AI systems on protected content without permission.

The short answer? No, not in a way that is currently defensible across leading legal systems.

Why This Matters: AI, Data, and Copyright Tensions

AI systems, particularly those involving machine learning and generative capabilities, rely on vast datasets for training. These datasets frequently include copyright-protected works—books, songs, films, images, and even databases—due to the high quality and relevance of such material. As Professor Rosati notes, this raises profound tensions between the rights of content creators and the needs of AI developers.

Governments and courts are increasingly being called upon to strike a balance: should IP rights holders be able to control how their works are used in AI development, or should there be carve-outs—exceptions or limitations—that allow AI training for purposes like research or education?

Scope of the Study: A Global Lens on IP Exceptions

Rosati’s study focuses on two major questions:

1. Can unlicensed AI training fall under existing research or teaching/education exceptions to copyright?

2. Can commercial AI developers lawfully rely on those exceptions or broader fair use defences to justify their use of protected content?

To answer these, she examines both EU/UK legislation and case law, as well as systems with more flexible or open-ended limitations on copyright, such as those in the United States, Singapore, Japan, and South Korea.

Key EU and UK Legal Frameworks: Articles 3 & 4 DSM Directive, and Section 29A CDPA

Much of the European debate has focused on the DSM (Digital Single Market) Directive, particularly:

· Article 3, which provides a mandatory exception for text and data mining (TDM) carried out by research organisations for scientific research.

· Article 4, which provides a broader, optional TDM exception applicable to any user unless rights holders have opted out.

However, Rosati warns against equating TDM with full AI training. While TDM involves copying data for the purpose of mining patterns or information, AI training generally involves broader uses, including reproductions, transformations, and potentially redistribution—activities which go beyond what Article 3 or 4 contemplate.

Further, while the UK’s Section 29A of the Copyright, Designs and Patents Act (CDPA) allows for TDM for non-commercial research, this too is narrow and currently not aligned with the broader use cases required by commercial AI developers.

Hamburg Court Case: LAION and the Misapplication of Article 3

A key part of Rosati’s analysis critiques a recent District Court of Hamburg decision involving LAION, an organization involved in building datasets for AI training. The court accepted that LAION qualified as a “research organization” under Article 3 DSM, without critically assessing whether their collaboration with commercial AI developers undermined that status.

Rosati warns that this sets a dangerous precedent by potentially allowing commercial entities to bypass rights clearance by partnering with academic or research institutions, thus exploiting a loophole in the law. The ruling also conflates TDM with the broader, more intrusive practice of AI training—raising concerns about legal overreach.

US Fair Use Doctrine: Flexible, but Not Unlimited

Unlike the EU, the United States relies on the fair use doctrine under Section 107 of the Copyright Act to assess whether unlicensed uses of protected content are permissible. This includes uses for “research,” “teaching,” or “commentary,” among others.

However, Rosati argues that AI training—even if nominally done for research—often fails the fair use test because:

· The use is commercial, especially if the AI system is sold or monetized;

· The training data includes entire works, not just excerpts;

· The effect on the market for the original works is potentially harmful;

· The use, although transformative in a broad sense, may not qualify as transformative enough under US case law.

In other words, fair use does not provide a blanket exemption for AI developers—and courts will likely apply strict scrutiny to such claims.

Other Jurisdictions: Singapore, South Korea, Japan

Countries like Singapore (Section 244 of its Copyright Act), Japan (Article 30-4), and South Korea have more permissive, fair use–style exceptions or specific TDM provisions. Yet even in these jurisdictions, the scope of permissible use is still bounded, and commercial exploitation of training outcomes remains a legal grey zone.

Final Verdict: Licensing Is Not Optional

Rosati concludes that no existing exception or defence, either in Europe or globally, fully permits unlicensed AI training, even when such training is for “research” or “learning.” Most legal frameworks simply were not designed to accommodate the scope and scale of modern AI development.

Therefore, licensing is not only advisable—it is ultimately unavoidable, especially for AI developers intending to launch models in jurisdictions with strong copyright enforcement, such as the EU.

Looking Ahead: Legislative Reform or Legal Risk?

This study lands at a time when governments are actively reassessing the interaction between IP and AI. The UK Government’s consultation on copyright and AI has concluded but is still awaiting a formal response. Meanwhile, European Commission President Ursula von der Leyen has hinted at legislative proposals aimed at making Europe “the home of innovation again.”

But innovation cannot come at the cost of legal certainty. Until and unless laws evolve to explicitly cover AI training scenarios, AI developers should proceed with caution—and with licences in hand.