Large Language Models, Fair Use, and Protecting Author’s Rights: Re-Weighing the Four Factors of Section 107 of the Copyright Law

Abstract

This paper will argue that the four factors of Section 107 of the Copyright Act must be evaluated with more equal or comparable weights when discussing the usage of copyrighted materials to train Large Language models (LLMs). The question of fair use in the context of generative AI has repeatedly arisen in recent court cases, as authors bring class action suits against tech companies like Meta and Apple for using their works without permission to power human-like LLMs. Authors claim deprived revenue and market dilution, as they argue that AI models are creating artificial competition in three ways: first, in deprived licensing potential; second, in the form of replicated versions of their work both in regurgitated outputs and expressively similar summaries; and third, within competitive “original” content. This paper will first describe LLMs and the current concerns surrounding their training on copyrighted material. It will then define copyright law, set out the history and current cases facing courts, and conduct a fair use assessment on both the training of LLMs on copyrighted works and their outputs. The paper will propose a solution to the problem of courts accepting the training of LLMs as fair use without considering their outputs. Such a solution will center upon a de-emphasis of the historically salient transformative factor and instead propose a more balanced assessment of the four factors of Section 107. The analysis will use a qualitative approach founded upon copyright law and the fair use doctrine, precedent from court decisions, and ongoing cases to arrive at such a conclusion. The goal of this paper is to enhance the working understanding of LLMs relation to copyright law and support the rights of authors to protect their work in a rapidly evolving judicial landscape.