In this article, we apply historical copyright principles to the evolving state of text-to-image generation and explore the implications of emerging technological constructs for copyright’s fair use doctrine. Artificial intelligence (“AI”) is frequently trained on copyrighted works, which usually involves extensive copying without owners’ authorization. Such copying could constitute prima facie copyright infringement, but existing guidance suggests fair use should apply to most machine learning contexts. Mark Lemley and Bryan Casey argue that training machine learning (“ML”) models on copyrighted material should generally be permitted under fair use when the model’s outputs transcends the purpose of its inputs. Their arguments are compelling in the domain of AI, generally. However, contemporary AI’s capacity to generate new works of art (“generative AI”) presents a unique case because it explicitly attempts to emulate the expression copyright intends to protect. Jessica Gillotte concludes that generative AI does not illicit copyright infringement because judicial guidance requires adherence to the constitutional imperative to promote the creation of new works when technological change blurs copyright’s boundaries. Even if infringement does occur, Gillotte finds that fair use would serve as a valid defense because training an AI model transforms the original work and is unlikely to damage the original artist’s market for the copyrighted work. Our paper deviates from prior scholarship by exploring specific generative AI use cases in technological detail. Ultimately, we argue that fair use’s first factor, the purpose of the use, and its fourth factor, the impact on the market for the copyrighted work, both weigh against a finding of fair use in generative AI use cases. However, even if text-to-image models aren’t found to be transformative, we argue that the potential for market usurpation alone sufficiently negates fair use.
There is presently little specific guidance from courts as to whether using copyrighted works to build generative AI models constitutes either infringement or fair use, although several related lawsuits are currently pending. Text-to-art generative AIs present several scenarios that threaten substantial harm to the market for the copyrighted original, which tends to undercut the case for fair use. For example, a generative AI trained on copyrighted works has already enabled users to create works “in the style of” individual artists, which has allegedly caused business and reputational losses for the emulated copyright holder. Furthermore, past analyses have ignored the potential for a model to be non-transformative when its intended output has the same purpose and is of the same nature as its copyrighted inputs.
This article contributes to the discussion by shining a technical light on text-to-art AI use cases to explore whether some uses normatively fail to qualify as fair uses. First, we examine whether text-to-image models present a prima facie infringement claim. We then distinguish text-to-image generative AIs from non-image focused AIs. In doing so, we argue that when the nature of the copyrighted work and the purpose of the infringing use are the same, it is more likely that the original artist will experience market harm. This tilts the overall analysis against a finding of fair use.
Jacob Alhadeff, Cooper Cuene & Max Del Real,
LIMITS OF ALGORITHMIC FAIR USE,
19 Wash. J. L. Tech. & Arts
Available at: https://digitalcommons.law.uw.edu/wjlta/vol19/iss1/1