Current fair use law provides inadequate protection for model owners. Safe Harbors create the space to innovate while also supporting creators

By expanding the capabilities of models, machine learning researchers and practitioners have stumbled into the muddy waters of fair use. As a result, websites like Getty Images have banned AI-generated content (Vincent, 2022), and lawsuits have been filed against products using foundation models, including against GitHub Copilot, Stable Diffusion, OpenAI’s ChatGPT, Meta’s Llama, and more (DOE 1 v. GitHub, Inc., N.D. Cal. 2022; Andersen et al. v. Stability AI et al., N.D. Cal. 2023; Vincent, 2023; Chabon v. OpenAI, Inc.,, 2023; Kadrey v. Meta Platforms, Inc., 2023). In this work, we shed light on this subject matter for machine learning researchers and highlight that significant additional work is required to de-risk foundation model deployments for generative use cases, focusing primarily on U.S. laws. First, we provide an overview of U.S. case law on the fair use doctrine.7 We draw analogies to foundation model use cases. We supplement these with a review of prior experiments, as well as novel experiments, and illustrate that foundation models can produce content that is sufficiently similar to copyrighted material. Furthermore, the case law suggests that even certain types of transformations of the training data would not be considered fair use. Thus, the risk of infringement is real, and fair use will not cover every scenario where a foundation model is created or used. The exact amount of risk is unclear, and the law will evolve with ongoing litigation. Second, we overview technical mitigation strategies that will reduce this risk in accordance with the current state of the fair use doctrine. Grimmelmann (2015) stated that “paying attention to robotic readership refocuses our attention on the really fundamental questions: what is copyright, and what is it for? To say that human readers count and robots don’t is to say something deep about the nature of reading as a social practice, and about what we want robots—and humans—to be.” Lemley and Casey (2020) suggested that humans and AI should be held to similar standards when it comes to copyright. If this is the case, it is the job of machine learning researchers and practitioners, working together with legal practitioners, to ensure that foundation models create transformative content which would pass muster under the same fair use analysis as provided to a human. To get there, new strategies and techniques will need to be developed, taking steps to ensure that foundation models behave in more transformative and novel ways. We call for more research to align technical mitigation strategies with fair use, including better output filtering mechanisms relying on higher-level semantics and new innovation in training-time techniques like extraction-preventative learning from human feedback. Developing these mitigation strategies is an important research challenge for machine learning and natural language processing and would bring practices in the two fields into better alignment with the law. Lastly, we argue that a co-evolution of technical mitigation strategies and law can help establish a middle ground where the positive impact of foundation models is realized while reducing the harms to data creators’ intellectual property rights. With the current uncertainties of fair use doctrine, as Sobel (2017) and others noted, the law may sway to one extreme or another. On one hand it could lead to overly permissive interpretations of fair use that could allow any generative AI use, disregarding the rights of data creators. Or it could lead to overly restrictive interpretations of fair use that could broadly prevent foundation model training and use, concentrating power among entities that have already acquired vast quantities of licensed data. By developing and deploying strong technical mitigation strategies, it may be possible to lessen the risk of such extreme legal outcomes. And the law should take into account the existence and strength of such technical mitigation strategies. This could involve a multi-pronged approach: considering technical mitigations in fair use assessments, clarifying the status of DMCA protections for foundation models, or developing DMCA-like safe harbors for deployments that use strong technical mitigation efforts, pursuing policy strategies for reducing harms to labor, and more. Realizing this middle ground requires the participation of a much broader community including the data creators impacted by foundation models, technologists, legal professionals, among many others. We encourage more multidisciplinary work to further the co-evolution of law, policy, and technical methods for mitigating intellectual property harms. Overall, the goal of this work is to act both as a guide and call-to-action for ML researchers and practitioners to actively pursue technical mitigation strategies. We hope that this guide helps instill a better understanding that fair use is not a panacea, and that a nuanced comprehension of the legal landscape’s intricacies is vital to effectively navigate potential pitfalls and uncertainties. Furthermore, this work may also prove useful to lawyers and policymakers, providing them with more insight into potential technical details of foundation models, including technical mitigation strategies, and how they might play a role in the developing legal best practices and potential reforms.