by Catarina Conran

Written on 27 April 2025

Why this matters

Large‑scale generative AI has become possible in recent years only because billions of works—books, news articles, photographs, code, cultural heritage materials, and more—are freely available online. Yet the copyright rules that govern these works were never written with machine‑learning in mind. As Large Language Models (LLMs) seep into education, research, and everyday creativity, we urgently need clear answers to three questions:

1. When AI developers copy protected works to train their models, is that lawful?
2. How do we keep the commons vibrant when vast commercial value is extracted from it?
3. How, if at all, should authors be remunerated when their works are used for AI training?

We don’t have definitive answers to these questions just yet, as courts, policymakers, and scholars are still working to determine how existing copyright principles should govern generative AI.

Against that backdrop, this blog post unpacks the current state of copyright law in this age of AI, reviews the headline lawsuits that may help to answer some of these questions, and proposes ways to channel AI’s power toward openness and reciprocity.

Copyright & AI – a short primer

1. Authorship & originality

Human authorship remains the bedrock of most national copyright regimes. In March of this year, the US Court of Appeals for the D.C. Circuit confirmed that a work “authored exclusively by artificial intelligence” cannot receive copyright protection (Thaler v. Perlmutter). Rather, copyright vests only in the human choices behind the final expression. Thus, while AI-assisted works can be copyrighted, copyright protection can only be extended to the portions of the work over which a human exercised creative control. Accordingly, the US Copyright Office now asks applicants to explain how they used AI and to disclaim the non‑human portions.

In Canada, the EU, and Australia, the requirement for copyright protection is framed as “intellectual creation”—again implying a human mind. Meanwhile, the UK’s Copyright, Designs and Patents Act assigns copyright to the person “who made the necessary arrangements,” although that rule is now under review.

2. Exceptions to copyright protections (Fair use & TDM exceptions)

Copyright owners have the right to prevent the unauthorized reproduction of their works. Training an LLM necessarily involves making temporary (and often permanent) copies of the training data, which infringes upon the reproduction right. Thus, AI developers rely on exceptions to copyright protection to train their LLMs.

• In the US, AI developers rely on the flexible fair use defense. Fair use is determined according to the following four factors:
  • Purpose & character of the use – Is model training transformative, like Google’s search‐indexing, or exploitative, like bulk copying for a competing product?
  • Nature of the copyrighted work – Was the work that the AI trained on heavily fact-based, or highly creative? Copying of fact‑heavy works is more likely to be fair use, while copying of highly creative works is more likely infringement.
  • Amount & substantiality of the copying – How much of the work was copied by the AI during training? It is common for the entire work to be ingested by AI crawlers, which may go against fair use.
  • Market effect on the copyrighted work – Does the model substitute for, or devalue, the original market for the work? This factor is central in the New York Times and Authors Guild cases (see next section for more).
• Instead of the fair use test, AI developers in the EU and the UK rely on the Text‑and‑Data Mining (TDM) exceptions, which permit non‑commercial scientific research bodies and cultural‑heritage institutions to mine any work to which they have lawful access, with no opt‑out for rights holders. EU members may also choose to permit TDM for commercial uses, unless a rights holder has elected to reserve their rights. Most EU countries have now implemented both articles, meaning that unless a work carries a clear opt‑out flag, LLM developers can train on it across the EU.

3. Moral rights & attribution

In many jurisdictions, creators enjoy moral rights—inalienable rights to be credited for their work and to object to revision, alteration, or distortion of the work. Outside the US these are granted by the Berne Convention; in the US they exist only in narrow form under the 1990 Visual Artists Rights Act and do not cover books, music, or journalism. As AI scales, safeguarding attribution and reputational integrity—even where not legally required—remains vital to a healthy commons.