Text and Data Mining, Generative AI, and the Copyright Three-Step Test

Abstract In the debate on copyright exceptions permitting text and data mining (“TDM”) for the development of generative AI systems, the so-called “three-step test” has become a centre of gravity. The test serves as a universal yardstick for assessing the compatibility of domestic copyright exceptions with international copyright law. However, it is doubtful whether the international three-step test is applicable at all. Arguably, TDM copies fall outside the scope of the international right of reproduction and go beyond the test’s ambit of operation. Only if national or regional copyright legislation declares the test applicable does the question arise whether copyright exceptions supporting TDM for AI training constitute certain special cases that do not conflict with the normal exploitation of a work and do not unreasonably prejudice legitimate author or rightholder interests. As the following analysis will show, rules permitting TDM for AI training can satisfy all test criteria. An opt-out opportunity for copyright owners eliminates the risk of a conflict with the normal exploitation of a work and an unreasonable prejudice from the outset. A clear focus on specific policy goals, such as the objective of supporting scientific research, adds conceptual contours that dispel concerns about non-compliance. In the case of TDM provisions covering commercial AI development, equitable remuneration regimes can be introduced as a counterbalance to avoid an unreasonable prejudice.