Restitution for Machine Learning: Unjust Enrichment and the Valuation of Training Data in Generative AI Litigation

Abstract

Generative artificial intelligence has exposed a structural mismatch between conventional copyright damages and the economic architecture of machine learning. Copyright plaintiffs increasingly allege that protected books, news articles, photographs, illustrations, code, and other expressive works were copied into training datasets and processed to build commercially valuable models. Yet a conventional loss-based damages inquiry may understate the economic significance of training where the claimant cannot show that the model displaced sales of the original work. This article develops a restitutionary account of machine-learning training based on unjust enrichment. It asks whether the legal focus should shift, in appropriate cases, from the claimant's lost market income to the defendant's non-consensual acquisition and use of a valuable input. The analysis is grounded in the uploaded reference set, which identifies the growing literature on AI and intellectual property, comparative approaches to generative AI, and the conceptual foundations of restitution, and is supplemented by recent primary legal developments. The article proposes a valuation architecture built around a hypothetical license, adjusted for the marginal contribution of the claimant's work, dataset scarcity, substitutability, training-stage importance, downstream commercialization, and the defendant's own technological contribution. It also examines disgorgement, reasonable royalty, constructive trust, continuing royalties, and the defense of change of position. The article concludes that a generalized entitlement to model profits would be overbroad, but that a narrowly tailored restitutionary or restitution-informed remedy can be justified where acquisition was unauthorized, the defendant received a measurable benefit, the benefit was causally connected to the claimant's protected input, and retention without compensation would defeat the allocation principles of copyright and private law. The paper proposes an "unjust enrichment by extraction" framework, while emphasizing that such a doctrine should not convert copyright into a general property right over facts, ideas, or statistical learning.