Assessing the feasibility of collective licensing of in-copyright works as training data for generative AI systems

Copyright owners have sued several developers of large-scale generative AI systems for copyright infringement because of their uses of massive quantities of in-copyright works as training data for building AI models. Fair use will be the main defense against these charges. If fair use defenses succeed, developers will be free to continue to commercially exploit models already built on copyrighted data as well as to use these data to train new models or fine-tune existing ones. If copyright owners prevail, developers may be liable for billions of dollars of damages. Developers could also be enjoined from further model development on in-copyright works and even ordered to destroy models trained on infringing works. Numerous commentators have proposed collective licensing as a compromise solution to the copyright-training-data dilemma. Other commentators have questioned the feasibility of such a compromise. This article discusses several proposals for collective licensing to enable development of generative AI systems while providing some compensation to copyright owners. It assesses the complex normative, economic, and practical problems that must be addressed if such a regime is to become feasible. It discusses the implications of a licensing mandate not only for the large firms whose models are widely used today, but also for start-ups, research centers, and higher education developers of generative AI systems, as well as the general public.