Citation: Matthew Sag, Fairness and Fair Use in Generative AI, 92 Fordham Law Review 1887 (2024)
In a nutshell:
In Fairness and Fair Use in Generative AI, Matthew Sag argues that generative AI training should generally qualify as “non-expressive use” and thus is likely to constitute fair use under U.S. copyright law, provided the models do not communicate the original expression from their training data to new audiences.
Summary
Fairness and Fair Use in Generative AI develops a principled framework for evaluating fair use claims in generative AI by grounding the analysis in copyright’s fundamental architecture rather than abstract policy judgments. The article cautions against treating fair use as a general public policy instrument to balance AI’s broader societal costs and benefits, arguing instead that fair use analysis should focus on whether the challenged use threatens the copyright owner’s interest in controlling how their original expression reaches the public.
One of the core insights of Fairness and Fair Use in Generative AI that copyright law centers on protecting original expression, not prohibiting all copying. When generative AI models learn abstract patterns from training data without reproducing original expression in their outputs, this constitutes “non-expressive use”—analogous to reverse engineering, plagiarism detection, and search engine indexing cases where courts have consistently found to be fair use.
Nonetheless, while well-designed generative AI models will typically meet the requirements of fair use, the article acknowledges room for fairness considerations under the fourth factor when uses undermine copyright’s economic incentives through unlawful access, systematic indirect expressive substitution, or a failure to implement appropriate safeguards.
Why read this article?
Fairness and Fair Use in Generative AI provides an excellent overview of the “non-expressive use” principle that has emerged from decades of copyright cases involving copy-reliant technologies, tracing a coherent line from reverse engineering cases through Hathitrust and Google Books to contemporary AI disputes. The paper offers valuable background on how fair use doctrine has evolved from Judge Leval’s transformative use framework through the Supreme Court’s recent refinement in Andy Warhol Foundation v. Goldsmith, explaining why courts have moved beyond viewing fair use as simply a policy balancing tool. It also provides accessible technical explanations of how large language models and text-to-image systems actually work, making clear why these systems typically learn at an abstract level rather than memorizing training data. The paper helpfully situates the generative AI copyright debate within copyright law’s fundamental architecture, particularly the idea-expression distinction and the principle that copyright protects communication of original expression rather than all uses of copyrighted works.
Further Reading
Pierre N. Leval, Toward a Fair Use Standard, 103 Harvard Law Review 1105 (1990)
This seminal article introduced the concept of “transformative use” that was later adopted by the Supreme Court in Campbell, arguing that fair use should focus on whether the secondary work adds value through new insights, understandings, or aesthetic purposes rather than simply serving as a substitute. Leval also emphasized that fair use analysis should derive from copyright’s internal principles rather than extraneous policy considerations.