Artificial intelligence platforms justify their unauthorized reproduction of copyrighted training data on the expansive creative capacities that their models allow. They argue that generative AI is a “radically transformational tool for creators of many kinds … which enables new expression and innovation to flourish.” In her new paper, Engineering Semiotic Democracy, Katrina Geddes sets out to hold AI firms’ feet to the fire.
To motivate the intervention, Geddes reports asking ChatGPT to produce an image of Captain America as a queer Black man from Brooklyn. The platform refused the request as a violation of its content policy that prohibits generating certain copyrighted characters. Because the output that her request would produce is likely protected as a transformational fair use, Geddes worries that risk-averse AI firms are blocking the creation of content in a manner that is more restrictive than copyright law requires. Doing so, she argues, undermines people’s capacities to engage and create with AI in precisely the sorts of radically generative ways the platforms tout in their litigation.
For Geddes, generative AI likely has an important role to play in expanding the diversity of creators and creations by lowering the costs of making new works and increasing the capacities of people who want to make them. While AI poses a risk of de-skilling some creators, Geddes also recognizes its potential to up-skill others. Many people who would never contemplate creating visual art if they had to draw or paint will be enabled by AI to express their ideas. This diversity of new voices and new creations represents the semiotic democracy in Geddes’s title.
Of course, some of these new works will incorporate copyrighted works in ways that content owners find objectionable. And content owners will point to them as evidence that generative AI is harming the market for their copyrighted works. To avoid these risks, AI platforms have begun restricting the prompts that they will accept and the outputs that they will generate. It is now much harder to generate copyrighted characters and other protected content, even if the resulting output isn’t unlawful.
Geddes’s intervention is important. One of the fundamentally transformative features of AI is its capacity to encourage broad swaths of cultural production. But if AI users can’t generate the equivalent of 2 Live Crew’s “Pretty Woman” or the Air Pirates version of Mickey Mouse, the value of AI as a creative catalyst is meaningfully less than we otherwise might have expected. Calls to restrict model memorization risk burdening otherwise protected—and important—speech.
Ben Sobel warned us about these issues in his recent Copyright Accelerationism paper (that I JOT’d here). AI firms will claim the benefits that fair use provides, allowing them to use others’ content for free, without extending the same benefits to their users. Read together, these papers should give copyright scholars pause about the realities of generative fair use for all.
Geddes’s concerns are ones we’ve seen before, as she readily acknowledges. Tech firms have long blocked user-generated content that is likely protected by fair use. Studies on copyright take-down requests have found that platforms like YouTube excessively block legal content. Geddes also acknowledges that some platform limitations on generation will arise, not from copyright risk, but from the platforms’ own prohibitions on certain kinds of outputs, including sexually explicit ones. If ChatGPT won’t create a queer Captain America, not because of the copyright but because it blocks prompts relating to sexual material, then copyright isn’t the problem.
In a forthcoming article in the Duke Law Journal, Matt Sag argues that the emerging arrangement of licensing deals between AI platforms and content providers may be the best option open to society. Sure, some stuff will be blocked, but the alternative of massive copyright liability is worse. Sag may turn out to be correct, but Geddes’s article helpfully clarifies what we stand to lose from this scenario. In the coming years, courts, scholars and policymakers must grapple with the tradeoffs that content licensing and output blocking create.
