AI Training is Fair Use: The Beginning of the End of the Copyright Assault on Gen AI

Two federal courts overseeing claims against the developers of generative artificial intelligence (GenAI) have pointed the way to resolving these infringement actions by finding that the training of GenAI models is a transformative fair use under copyright law. While the two opinions differed in tone and scope, this article takes these rulings as the starting point for a discussion on resolving the ongoing copyright claims against AI developers, signaling what may be the beginning of the end of the copyright assault on GenAI.

The goal of this article is to inject urgency into resolving these matters. It asserts that uncertainty over the legal status of AI training is a drag on innovation and development in this vital economic sector. While massive investments are pouring into this field, the money flows to an extraordinarily small number of players whose resources allow them to run the risks posed by class actions and multi-party actions demanding damages that might cripple even the largest companies. With the threat of destruction by copyright infringement action removed, AI development could expand and flourish among even the smallest of innovators.

Ending the infringement actions requires more than just a recognition that indiscriminately drawing data from existing works without permission and without licensing to create a generative artificial intelligence expression machine is fundamentally transformative under factor one of the copyright fair use test. Plaintiffs have fought to sell a theory of the case that keeps AI developers in the defendants’ seats, even though the parties responsible for the production of outputs and for any resulting market harm are the end-users of the technology.

This article asserts that the proper theory of these infringement cases is that GenAI developers made a general-purpose technology that can create an infinite variety of new, original expression, but end-users of the technology can choose to use it to compete with the plaintiff artists and creators in their same style and in their same medium, at massively reduced costs and massively increased speeds. And sometimes end-users will use the technology to create infringing works. Far from being a unique 21st century high technology story, this story is the same as that of photocopy machines, Betamax and VCR devices, scanners, image-editing software, and internet search engines, all of which are capable of making duplicates of expressive works that can be put to uses that infringe on the original works and harm their markets. Yet, the designers of these copying technologies are not sued for copyright infringement because of the disconnect between the action of creating a useful tool and the action of an end-user who co-opts the tool for their own purposes.

The designers of these GenAI models made them powerful and extraordinarily fluent tools for creating new expression with a “further purpose or different character, altering the first with new expression, meaning, or message,” but in the end, GenAI systems are just tools. They are not artists or authors and do not automatically regurgitate infringing content. Rather, they are tools capable of being used by end-users who may act purposefully to create substantially similar and potentially infringing works that can be used to compete with the plaintiffs.