Misdirection After a Heist: Rejecting Necessity as a Justification for Copyright Exploitation in Training AI Models

Abstract

The unauthorized use of copyright works in training artificial intelligence (AI) models, and particularly large text, image, and audio prediction models, is catalyzing debate and litigation as copyright owners discover and challenge the legality and fairness of the wholesale appropriation of vast swathes of human creative expression. AI developers such as OpenAI Inc., Midjourney Inc., and Meta Platforms Inc. maintain that training AI models on unlicensed copyright works is necessary, and that using only non-copyright or licensed copyright works will threaten the development, functionality, and very existence of AI systems. This article tests the veracity of these assertions, which are reinforced in US-centric scholarship on AI and copyright, and finds them wanting. We explain how it is possible, and indeed desirable, to develop AI systems without exploiting copyright owners, and build the policy case for why and how this should occur.

Keywords: Intellectual property, Copyright, Artificial intelligence, Fair use, Generative AI, Large language models, Large image models