ABSTRACT
Generative artificial intelligence has brought renewed urgency to the question whether using copyrighted works to train large language models constitutes fair use. A 2025 US Copyright Office Report and dicta in Kadrey v Meta Platforms proposed a broad theory of ‘market dilution’: even when particular outputs are not substantially similar to works in the training data, their aggregate competition may reduce sales of human-authored works. Bartz v Anthropic rejected that form of generalized competition as a cognizable copyright injury. This Article argues that neither the infringing status of an output nor a bare causal connection between training and lost demand resolves the fourth-factor inquiry. Courts should instead ask what feature of the challenged use caused the asserted loss. Market harm is cognizable when it is attributable to protected expression or invasion of a market copyright law reserves to the owner. It is not cognizable when demand is lost solely because follow-on works compete through genre, style at a high level of abstraction, facts, functional information, or other material copyright leaves free. Scale may magnify an injury after a protected market has been identified, but it cannot create an entitlement in an otherwise unprotected basis of competition. This source-of-harm rule preserves conventional claims involving memorization, regurgitation, substantial similarity, direct substitution, and properly defined licensing markets while preventing factor four from becoming a general right against technologically intensified competition.
Yoo, Christopher S, Market Dilution, Fair Use, and AI (September 28, 2026), University of Pennsylvania Law School, Public Law Research Paper No 26-25; Ohio State Technology Law Journal, forthcoming 2027.
Leave a Reply