On Tuesday, May 2, 2023, the U.S. Copyright Office (USCO) held the second of four sessions on the copyright implications of generative artificial intelligence (GAI), titled “Artificial Intelligence and Copyright – Visual Arts.”
The session focused on GAI issues relevant to visual works, and featured two panels with various stakeholders that brought a range of perspectives to the discussion. These panelists included representatives from GAI platform companies, graphic design software companies, think tanks, policy organizations, and law firms, as well as artists concerned by the impact of GAI.
Two issues consistently arose during the course of the three-hour session:
- Authorship in the context of GAI
- Legality of “training” GAI platforms
Regarding authorship, many spoke out in favor of USCO’s recent guidance on copyrightability of GAI works, and agreed with the general proposition that AI cannot be the “author” of a work. Others encouraged USCO to provide additional clarity on GAI works that contain a significant amount of human authorship. Still others questioned USCO’s recent decision to partially rescind registration of Kashtanova’s Zarya of the Dawn, a comic book with Midjourney-generated GAI images. In particular, Kashtanova’s copyright counsel, who participated in the panel, questioned USCO’s reasoning that Kashtanova could not have accurately predicted the visual outcomes of her text prompts, arguing that such “predictability” analysis was absent from USCO’s recent guidance.
Regarding GAI training, panelists took starkly differing positions on the legality of unlicensed use of internet-accessible content to “train” GAI platforms. For example, artist Karla Ortiz characterized this practice as grave and problematic, stating that “anthropomorphizing” these AI tools is “a fool’s errand,” and challenging the notion the GAI models are “trained” in the same way that artists learn from studying other artistic works. Ben B., the representative from Stability AI, on the other hand, discussed how GAI models are “trained,” characterized the use of training materials as “transformative” (a clear allusion to copyright fair use), and pointed out common limitations of GAI (e.g. difficulties producing images of human hands) to demonstrate that GAI learns from exposure to third-party works, and is not merely a “collage machine” that copies portions of those works.
Several participants also called attention to the removal of copyright management information (CMI) (e.g., copyright notices) in the training process. Relatedly, some called for the expansion of Section 1202 of Title 17–which prohibits willful removal of CMI—arguing that it should prohibit non-willful conduct, and reasoning that GAI models have exposed a flaw in Section 1202 through the systematic removal of CMI from billions of images.
The third USCO session on GAI is scheduled for May 17, 2023, and will focus on audiovisual works, and the final session is scheduled for May 31, 2023, and will focus on musical works.