Behind the Scenes of an AI-generated Exhibition

This exhibition, Act as if you are a curator: an AI-generated exhibition, started as a conversation with a computer program called ChatGPT. Chat Generative Pre-Trained Transformer, more commonly known as ChatGPT, is a rapidly-growing artificial intelligence chatbot that can read and generate text in response to questions, requests, or conversation. It has been known to solve complicated equations, pass the bar exam, and even write poetry and songs. But could it organize an exhibition of artworks from the collection of the Nasher Museum of Art – a task typically performed by a (human) museum curator?

When we first had this idea, we naively thought that we could simply ask. So, we did.

BELOW: Screen grab from Chat GPT

As it pointed out, ChatGPT quite obviously lacked the ability to physically visit museum storage to make selections for an exhibition. What we did not realize is that it also lacks the ability to access our publicly accessible database of artworks called eMuseum. It has only been trained with books, articles, and select websites published before the year 2021. While it suggested a short title and explanation for a possible exhibition, it selected generic artwork types or well-known works of art that are not part of the Nasher’s collection, for example, “a pre-Columbian Maya vase depicting a royal ceremony” or Three Musicians by Pablo Picasso. Searching eMuseum for the Maya vase alone produces hundreds of results.

BELOW: One of many results from a search for a Maya vase

In the latter case, Picasso produced two paintings by this title, now in the collections of the Museum of Modern Art, New York, and the Philadelphia Museum of Art. While museums regularly borrow works of art from other museums, institutions and collectors, the parameters for this project were that ChatGPT use the collection of the Nasher Museum. In other cases, ChatGPT “hallucinated,” as it is known to do, and suggested artworks that do not exist. In these cases, the AI’s desire to provide an answer overcomes its lack of knowledge, leading it to suggest totally fabricated solutions.

RIGHT: Screenshot of the eMuseum website showing the Nasher Museum’s holdings in Mayan vases

We quickly realized that we needed the help of someone with more specialized knowledge in how these platforms work. We contacted Mark Olson, Ph.D., Associate Professor of the Practice or Art, Art History & Visual Studies at Duke University, because of his specialty in the digital humanities and great enthusiasm for experimental projects at the intersection of art and technology. He and his students quickly engineered a solution and built us a custom ChatGPT interface.

Using tools provided by OpenAI and several other data science frameworks, Olson transformed publicly-accessible information on the over 14,000 objects in our collection (such as artist, title, date, medium, cultural group, keywords, and description) into statistical representations of “relatedness.” Our custom version of ChatGPT was thus “educated” about the Nasher’s collection, able to access, select and make connections among specific works.

BELOW: Screen grab from ChatGPT “hallucinating” works of art

Once we determined that ChatGPT was capable of selecting specific works from the Nasher’s collection, we asked it what themes or topics would be most appealing for an exhibition at a university art museum. It had many suggestions, but consistently returned to themes of utopia, dystopia, dreams and the subconscious – topics especially relevant to current discussions on the future of artificial intelligence, its potential to develop sentience, and risks it could pose to humanity. Using those search terms, we again queried ChatGPT, now equipped to pull directly from our database and identify artworks by accession number. Often, it would repeat the same works of art it already gave us. As a solution, we asked ChatGPT for synonyms related to the main theme to retrieve more works from the program. We used polite and enthusiastic encouragement, which typically prompted the program to provide better results, and requested that it keep selecting artworks on these themes until we had an appropriate number to fill the 600-square-foot space of the Incubator Gallery. We also asked it to generate an exhibition title, introductory text, 50-word labels for the selected artworks, and a sequence in which to install them in the gallery.

BELOW: Screen grab from our customized version of ChatGPT

Given the limitations of time and resources, our team was only able to pursue one of the many avenues that could have assisted us in this process. While curators often work on exhibitions for several years, we only had several months to develop this project, as the idea emerged from recent advances in artificial intelligence. It is our hope that this exhibition can be a site for conversation about this technology and its applicability and limitations in museums and the humanities.

We also acknowledge that there are ways to pursue this project that give more or less curatorial authority to the AI. For example, because ChatGPT currently lacks the ability to generate a floorplan or mock-up of the gallery, Nasher staff had to rely on its written instructions for where and how to group artworks. As this technology advances, we expect that there will be even easier ways – like the new ChatGPT code interpreter function – to enlist technology as a curatorial counterpart.

As the use of AI in museum practice and elsewhere increases, it is important to acknowledge that the technology is not neutral and is filled with inherent biases. We encourage you to explore the exhibition and welcome your feedback on this experiment as we consider the consequences of this technology as it becomes increasingly interwoven into our day-to-day lives.

BELOW: Screen grab from ChatGPT with good search results

This exhibition is organized by ChatGPT. The Nasher support team includes Julia McHugh, Ph.D., Trent A. Carmichael Director of Academic Initiatives, and Curator of Arts of the Americas; Julianne Miao, Curatorial Assistant; Mark Olson, Ph.D., Associate Professor of the Practice of Visual and Media Studies at Duke University; and Marshall N. Price, Ph.D, Chief Curator and Nancy A. Nasher and David J. Haemisegger Curator of Modern and Contemporary Art. Duke undergraduates Irma Lopez, Alveena Nadim, Maddie Rubin and David Sardá provided research support.