This paper argues that by failing to acknowledge the complexity of modern research practices that are shifting from a single discipline to multiple disciplines involving many entities, some public, some private, the proposed AI Act creates mechanisms for regulatory arbitrage. The article begins with a semantic analysis of the concept of research from a legal perspective. It then explains how the proposed AI Act addresses the concept of research by examining the research exemption that is set forward in the forthcoming law as it currently exists. After providing an overview of the proposed law, the paper explores the research exemption to highlight whether there are any gaps, ambiguities, or contradictions in the law that may be exploited by either public or private actors seeking to use the exemption as a shield to avoid compliance with duties imposed under the law.

To address whether the research exemption reflects a coherent legal rule, it is considered from five different perspectives. The paper begins by examining the extent to which the research exemption applies to private or commercial entities that may not pursue research in a benevolent manner to solve societal problems, but nevertheless contribute to innovation and economic growth within the EU. Next, the paper explores how the exemption applies to research that takes place within academia but is on the path to commercialization. The paper goes on to consider the situation where academic researchers invoke the exemption and then go on to provide the AI they develop to their employing institutions or other public bodies for no cost. Fourth, the paper inspects how the exemption functions when researchers build high-risk or prohibited AI, publish their findings, or share them via an open-source platform, and other actors copy the AI. Finally, the paper considers how the exemption applies to research that takes place “in the wild” or in regulatory sandboxes.