The discussion of AI copyright infringement or fair use often skips over all the required steps of the infringement analysis in order to focus on the most intriguing question, “Could a visual generative AI generate a work that potentially infringes a preexisting copyrighted work?” and then the discussion skips further ahead to, “Would the AI have a fair use defense, most likely under the transformative test?” These are relevant questions, but without considering the actual steps of the copyright infringement analysis, the discussion is misleading or even irrelevant. This neglecting of topics and stages of the infringement analysis fails to direct our attention to a properly accused party or entity whose actions prompt the question. Making a sudden transition from a question of infringement in the creation of training datasets to the creation of foundation models that draw from the training data to the actual operation of the generative AI system to produce images makes a false equivalency regarding the processes themselves and the persons responsible for them. The questions ought to shift focus from the persons compiling the training dataset used to train the AI system and the designers and creators of the AI system itself to the end users of the AI system who conceive of and cause the creation of images. The analysis of infringement or fair use in the generative AI context has suffered from widespread misunderstanding concerning the generative AI processes and the control and authorship of the end-user. Claimants, commentators, and regulators have made incorrect assumptions and inaccurate simplifications concerning the process, which I refer to as the Magic File Drawer theory, the Magic Copy Machine theory, and the Magic Box Artist theory. These theories, if they were true, would be much easier to envision and understand than the actual science and technology that goes into the creation and operation of a contemporary visual generative AI system. Throughout this Article, I will attempt to clarify and correct the understanding of the science and technology of the generative AI processes and explain the different roles of the training dataset designers, the generative AI system designers, and the end-users in the rendering of visual works by a generative AI system. Part II will discuss the requirements of a claim of copyright infringement including each step from the copyrightability of the claimant’s work, the doctrines that limit copyrightability, the requirement of an act of copying, and the infringement elements. Part III will summarize the copyright fair use test paying particular attention to the purpose and character of the use analysis, 17 U.S.C. § 107(1), and the current interpretation of the “transformative” test after Andy Warhol Foundation v. Goldsmith, particularly in circumstances relating to technology and the use of copyrighted or copyrightable data sources. Part IV will analyze potential infringement or fair use by the creators of generative AI training datasets. Part V will analyze potential infringement or fair use by the creators of visual generative AI systems. Part VI will analyze potential infringement or fair use by the end-users of visual generative AI systems. For all their complexity, visual generative AI systems are tools that depend on an end-user who conceives of and designs the image and provides the system with a prompt to set the generative process in motion. The end-users are responsible for crafting the prompt or series of prompts used, for evaluating the outputs of the generative AI, for adjusting and editing the iterations of images offered by the AI system, and ultimately for selecting and adopting one of the images generated by the AI as the final image. The end-users then make further decisions about the actual use and its function and purpose for the images the end-users selected and adopted from the outputs of the AI. While working with the AI tool to try to produce a certain image, an end-user might steer the system to produce a work that could, under an infringement analysis, be regarded as potentially infringing, which would lead us again to the fair use analysis based on the end-user’s use of the image.
Read full abstract