AI Act and Author Remuneration - A Model for Other Regions?

Abstract

With the adoption of the AI Act, the EU has substantially enhanced the rules governing the training of generative AI systems and, more specifically, the interface with copyright protection. The AI Act clarifies that, from an EU perspective, reproductions carried out for AI training purposes have copyright relevance and require the authorization of rightholders. It also confirms the rights reservation system following from the 2019 Directive on Copyright in the Digital Single Market: declaring an “opt-out” in an appropriate – machine-readable – manner, copyright owners seeking to prevent the use of their works for AI training purposes can reserve their rights. Remarkably, the AI Act seeks to universalize this approach and achieve a “Brussels effect.” Regardless of whether the training has taken place in the EU or elsewhere, it imposes a market ban on AI systems that have not been trained in accordance with EU requirements, including the obligation to observe opt-outs declared under Article 4(3) CDSMD. To enable rightholders to police AI training processes, the AI Act introduces a new transparency obligation. Developers of generative AI systems must submit sufficiently detailed information on work repertoires that have been used for training purposes.

Before embarking on a discussion of these new rules, the analysis sheds light on the policy objective underlying this copyright package in the AI Act: the intention to ensure that authors are properly remunerated for the use of their works in AI training processes. It then discusses the extension of EU opt-outs to other regions and the training data transparency which the EU legislator deems necessary to enforce copyright in AI training contexts. Finally, the potential benefits to authors will be weighed against the regulatory burdens the AI Act imposes on AI trainers. As a final step, the analysis explores alternative solutions. While the AI Act focuses on the input dimension (the use of human works for AI training), it is conceivable to take the output dimension (the offer and commercialization of fully trained AI systems) as a reference point for remuneration systems. Following this alternative avenue, the remuneration obligation concerns the final stage when generative AI products and services are brought to the market. In contrast to the strategy underlying the AI Act, this alternative approach refrains from encumbering the AI training process with obligations to observe opt-outs, establish lists of training resources and pay remuneration. Before following in the footsteps of the AI Act, law and policymakers in other regions should evaluate the advantages of this alternative policy avenue.

Keywords: copyright, AI regulation, algorithms, transparency, opt-out, territoriality, cultural science, aesthetic theory, text and data mining, machine learning, fair use, collective rights management, freedom of expression, art autonomy, right to research, equitable remuneration, exceptions and limitations