The Globalization of Copyright Exceptions for AI Training

Abstract

Generative AI, machine learning and other computational uses of copyrighted works pose profound questions for copyright law. This article conducts a global survey of multiple countries with different legal traditions and local conditions to explore how they have attempted to answer these questions in relation to the unauthorized use of copyrighted works for AI training.

Although the world has yet to achieve international consensus on this issue, an international equilibrium is emerging. Jurisdictions with common law and civil law traditions, and with varying economic conditions, technological capabilities, political systems and cultural backgrounds, have found ways to reconcile copyright law and AI training. In this equilibrium, countries recognize that text and data mining, computational data analysis and AI training can be socially valuable and may not inherently prejudice the copyright holders' legitimate interests. Such uses should therefore be allowed without express authorization in some, but not all, circumstances.

We identify three forces driving toward this equilibrium: (1) the centrality of the idea-expression distinction; (2) global competition in AI; and (3) the race to the middle. However, we also address factors that may upset this emerging equilibrium, including ongoing copyright litigation, partnerships and licensing deals in the United States as well as legislative and regulatory efforts in both the United States and the European Union, including the EU AI Act.

A key lesson of our cross-country survey is that globally, the binary policy debate that assumes that text and data mining and AI training must be categorically condemned or applauded has been eclipsed by a more granular debate about the specific circumstances in which the unauthorized use of copyrighted works for AI training should be allowed or prohibited. Countries that have hesitated until now to modernize their copyright laws in the area of AI training have several templates open to them and little reason for hesitation.

Keywords: Copyright, generative AI, Fair Use, Limitations and Exceptions, Nonexpressive Use, Text Data Mining (TDM), Machine Learning, Three-Step Test