Article

Copyright’s Latent Space: Generative AI and the Limits of Fair Use

BJ Ard 

Associate Professor of Law, University of Wisconsin Law School, and Affliate Fellow, Yale Information Society Project. The author thanks Oren Bracha, Robert Brauneis, Carys Craig, R. Feder Cooper, Rebecca Crootof, Deven Desai, Justin Hughes, Mark Lemley, Glynn Lunney, Michael Murray, David Nimmer, Jacob Noti-Victor, Blake Reid, Matthew Sag, Frederic Sala, Benjamin Sobel, Madhavi Sunder, Charlotte Tschider, Molly Shaffer Van Houweling, Christopher Yoo, and Peter Yu for feedback on this project, along with participants at the Fifth Annual Art Law Works-in-Progress Colloquium, 2024 Copyright Scholarship Round-table, UW Integrating Robots into the Future of Work Colloquium, 2024 Legal Scholars Roundtable on AI, Texas A&M School of Law Transformation in IP and Technology Law Symposium, 2024 Works-in-Progress for IP Colloquium, and faculty workshops at UGA and UW law schools. The author also thanks Kate Bishop, Peter Feider, John Lavanga, and Rosemary Patton for excellent research assistance. Research support was provided by the Office of the Vice Chancellor for Research and Graduate Education at UW with funding from the Wisconsin Alumni Research Foundation.

30 May 2025


Generative AI poses deep questions for copyright law because it defies the assumptions behind existing legal frameworks. The tension surfaces most clearly in debates over fair use, where established tests falter in the face of generative systems’ distinctive features. This Article takes up the fair-use question to expose copyright’s limitations as well as its latent commitments, particularly its allowances for the exploitation of non-authorial value. Fair use’s transformative use paradigm, which compares the purpose of the use with that of the original work, faces difficulty evaluating copying during the training of AI models. Close examination of the technology—from training through the operation of completed systems—reveals that the purpose of copying may be contingent because a model’s capabilities and ultimate uses are indeterminate at the time of training. This hurdle can be sidestepped by recognizing that purpose serves as a proxy for determining whether the use intrudes on markets rightly belonging to the copyright owner. However, this raises the question of which markets those are. Answering the market question requires delving into copyright’s latent space—the unarticulated principles and commitments embedded in its jurisprudence. This Article identifies a dividing line between market value that stems from an author’s creative choices and market value that does not, with courts permitting users to tap into the latter even to the copyright owner’s detriment. The reoriented test would ask whether a user exploits non-authorial value like that which stems from facts, tropes, and third-party investment versus the authorial value arising from an artist’s creative decisions. The precise line remains to be hashed out—courts have historically drawn the line differently across creative fields to balance copyright’s competing objectives in specific contexts. The fair-use question also reveals deeper structural limitations of the copyright regime. Concretely, the argument that copyright’s pro-artist policies compel denial of fair use misses that AI systems trained on licensed works may still displace human creators. The lack of unauthorized use takes the problem outside copyright’s domain. The core problem is not the duplication of specific works, but the ability to produce comparable works cheaply and quickly. The challenge cannot be resolved through the mere extension or denial of fair use. Instead, it demands we put copyright in dialogue with other regimes for promoting the arts, blunting the misuse of these tools, and confronting the technology’s capacity to consolidate power.

To read this Article, please click here: Copyright’s Latent Space: Generative AI and the Limits of Fair Use.