Who Owns The Copyright For AI-Generated Thanksgiving Recipes?
New York Times food writer Priya Krishna used OpenAI products to generate new Thanksgiving recipes and images, prompting the question: who owns the copyright for these recipes? According to OpenAI’s terms of use, Ms. Krishna owns them, but in reality, copyright for machine generated content is more complicated than that.
Attribution
© Famartin, CC-SA-4.0, via Wikimedia Commons
Enforcing copyright protection for recipes is notoriously difficult. While U.S. copyright law strives to protect “original works of authorship” by barring unauthorized reproduction of works of authorship including literary works, pictorial and graphic works, dramatic works, etc., this protection does not extend to ideas, procedures, or processes described or explained in the work. 17 U.S.C. § 102. When copyrighting a recipe, the copyright only applies to the verbatim text of the recipe. Thus the recipe can be reproduced with minor changes without infringing on its copyright protections. To avoid recipe theft, recipe developers and authors have developed permissions with instructions on how to reuse their work. Cookbook authors have started adding more easily protected elements to cookbooks including narrative essays and photography, both of which can be copyrighted and less easily reproduced without infringing. Even though recipe copyright protection can be somewhat ineffective, AI-generated recipes still pose an interesting question - who owns these recipes? The AI? The AI developers? Or the user inputting the recipe prompts?
In anticipation for Thanksgiving, New York Times food writer Priya Krishna experimented with artificial intelligence to create new Thanksgiving recipes. She used OpenAI products GPT-3 and DALL-E to create and write the new Thanksgiving recipes and generate images of the recipes. To write the recipes she used OpenAI’s language prediction model GPT-3, which stands for Generative Pre-trained Transformer 3. GTP-3 is a neural network that is designed to receive and understand language inputs and transform the input to what it predicts is the most useful language output for the user. It has been fed around 570gb of text information from multiple publicly available sources including Wikipedia and it is the largest neural network ever created. DALL-E works similarly to GPT-3, with an input and output model. For both, language commands and reference materials, input into the service (GPT-3 or DALL-E) are referred to as “inputs.” The service then generates and returns content based on the inputs received called “outputs.”. In the video, Ms. Krishna inputs her culinary background and preference for recipe development. GTP-3 used these inputs to create recipes for new Thanksgiving dishes including pumpkin spice chaat and naan stuffing. Krishna then used the GPT-3 generated recipes and inputs for the image generator DALL-E, which generated eerily realistic renditions of the recipes.
OpenAI’s Terms of Use, last updated November 3, 2022, states that users own their “input” for both DALL-E and GTP-3. In the Terms of Use, Open AI assigns the user OpenAI’s right, title, and interest in and to the user’s generated output. Thus, according to OpenAI, Priya Krishna “owns” the recipes and photos that GPT-3 and DALL-E generated by her inputs. However, Open AI is quick to note that GPT-3 and DALL-E, while generating new and original content, may still generate the same or similar outputs for OpenAI users or third parties.
Further, OpenAI’s Terms of Use states that Priya Krishna owns the recipes GPT-3 and photos DALL-E generated under her inputs; however, the United States Copyright Office (USCO) has denied copyright protection toAI generated works. On February 14, 2022, USCO denied copyright of an artistic work titled “A Recent Entrance to Paradise” generated by AI. In their reasoning, the board held that creative works generated autonomously without creative contributions from a human actor does not meet the standard for copyright protection. USCO stated that copyright law only protects “’the fruits of intellectual labor’ that ‘are founded in the creative powers of the [human] mind.’ COMPENDIUM (THIRD) § 306 (quoting Trade-Mark Cases, 100 U.S. 82, 94 (1879)).”
In this case, the applicant intended to fully credit AI for the creation of “A Recent Entrance to Paradise”, so the applicant did not include their human inputs as a basis for copyright in the registration application. USCO declined to comment on what level of human involvement in the creation of machine-generated works would meet the statutory criteria for copyright protection. The Supreme Court and lower courts have rejected attempts to extend copyright protection to non-human creations. Federal agencies have followed the courts and have held that while devices may be used in the creation of creative works there must be “at least minimal human creative effort at the time the work is produced.” In the “A Recent Entrance to Paradise” case, if the applicant instead included their contributions to the creative work, perhaps the USCO would view the creative work as a result of human creativity.
In the video, Krishna used very general commands so it is unlikely that USCO would find that her inputs generated outputs that were eligible for copyright protection. But what would happen if she uploaded her entire cookbook as an input? Would that meet the statutory requirements for human involvement in machine-generated work? USCO has yet to confront this issue. With the growing prevalence of AI across industries, it is only a matter of time before the USCO is forced to confront more of these pressing questions for the protection of AI-generated creative works.