Elements of Style: Copyright, Similarity, and Generative AI

73 Pages Posted: 22 May 2024

See all articles by Benjamin Sobel

Benjamin Sobel

Cornell University - Cornell Tech NYC

Date Written: May 18, 2024

Abstract

“You can’t copyright style” is a shibboleth in today’s debate over generative AI. This slogan is, at best, meaningless. More likely, it’s wrong. Sometimes, what we call “style” is copyrightable. “Substantial similarity” is the doctrine that assesses when stylistic copying becomes infringement, but it is notoriously erratic, and judges find it especially hard to apply to images. Current law obfuscates artists’ rights to control their works and the public’s rights to use generative AI.

Part I explains how image-generating AI works and debunks the prominent metaphor that it is a “collage machine.” The metaphor erroneously posits that it is possible to differentiate “mechanical” reproductions of works of visual art from “intellectual” reproductions, and it erroneously implies that the distinction has legal significance. Generative AI is clearly learning to reproduce something from its training data: what matters is what that something is.

Part II defines style as a holistic attribute of a work, or a group of works, that comprises a constellation of expressive choices. These expressive choices might be unprotectable individually, but in combination, they may constitute protectable expression. Part II documents courts’ struggles to assess similarity in visual art and attributes these struggles to the substantial similarity test’s near-irreconcilable demands: courts must simultaneously dissect images into their constituent elements—a task judges claim they are unable to do—while also assessing works’ aesthetic appeal holistically and intuitively. Style has always been a challenge for substantial similarity because it is the form of expression least susceptible to analytical dissection and most likely to elicit inarticulate aesthetic intuitions. Generative AI models’ replication of style is a hard problem for copyright law because the models are purpose-built to identify and reproduce precisely the forms of similarity that are hardest to analyze legally.

Keywords: AI, generative AI, copyright, style, image-generating AI, art, visual art, similarity, intellectual property, cyberlaw, copyright law

JEL Classification: K23, O34

Suggested Citation

Sobel, Benjamin, Elements of Style: Copyright, Similarity, and Generative AI (May 18, 2024). Harvard Journal of Law & Technology, Forthcoming Vol. 38, Cornell Legal Studies Research Paper Forthcoming, Available at SSRN: https://ssrn.com/abstract=4832872 or http://dx.doi.org/10.2139/ssrn.4832872

Benjamin Sobel (Contact Author)

Cornell University - Cornell Tech NYC ( email )

2 West Loop Rd.
New York, NY 10044
United States

Do you have a job opening that you would like to promote on SSRN?

Paper statistics

Downloads
144
Abstract Views
393
Rank
375,113
PlumX Metrics