CFM has explored museum applications of artificial intelligence (AI) since 2017, notably business analytics, customer service, collections data analysis, and provenance research. Today on the blog, Julianne Miao tells us how the Nasher Museum of Art at Duke University used the explosive evolution of generative AI this year to fuel a more provocative experiment: what if would happen if AI was invited to curate an exhibit?
–Elizabeth Merritt, VP Strategic Foresight and Founding Director, Center for the Future of Museums
What started as a joke developed into an exhibition when the idea of ChatGPT curating a show at the Nasher Museum of Art was proposed during a curatorial meeting. After the joke was made, there was a pause as we searched each other’s faces to see how seriously we wanted to take this idea. Not only has AI been the topic of classes, conferences, and research across Duke University’s campus, but conversations around AI permeate our everyday culture now more than ever. Whether it’s a new method of artmaking or to complete the “new” Beatles song, it’s part of the ways we now consume art and culture, so I was excited at the opportunity to engage in this conversation. After the meeting, my colleagues, Julia McHugh (Trent A. Carmichael Director of Academic Initiatives and Curator of Arts of the Americas), Marshall Price (Chief Curator and Nancy A. Nasher & David Haemisegger Curator of Modern & Contemporary Art), and I naively asked ChatGPT: “Curate an exhibition using works of art from the Nasher Museum of Art’s Collection.” When it spit back things like Vincent van Gogh’s Starry Night (owned by the Museum of Modern Art), Georges Seurat’s A Sunday Afternoon on the Island of La Grande Jatte (Art Institute of Chicago), and Grant Wood’s American Gothic (also from AIC), we phoned a colleague who understood the “minds” of these machines better than we did.
ChatGPT didn’t know our collection. While it’s often lauded for its “encyclopedic knowledge,” the large language model (LLM) learned select resources during its development and has limited knowledge on current events past September 2021. You can ask ChatGPT: “Tell me about the Nasher Museum of Art in a tone that mirrors the writing of Sylvia Plath,” and it would respond, “In the heart of Durham, North Carolina, nestled beneath the azure sky that seemingly mocks the profound melancholy, stands the Nasher Museum of Art, a place that echoes the fragile beauty of existence. It is a museum that embodies the delicate dance of light and shadow, a reflection of the inner tumult that plagues the human soul.” (Yes, I actually asked.) However, it could not curate an exhibition with works of art from the Nasher Museum of Art’s collection without a little help.
The process of developing this exhibition exposed many of the capabilities and limitations of ChatGPT. Because it could not access our publicly-accessible collection database on our website, Mark Olson, Associate Professor of the Practice of Art, Art History & Visual Studies, and a founding member of the Digital Art History & Visual Studies Lab, built a custom ChatGPT interface. He, along with student researchers, integrated ChatGPT with data scraped from 14,000 records of objects in our collection. The information pulled is knowledge available to the public, including artist, birth place, medium, and descriptions and keywords of the work, which make our public database more searchable. Now, equipped with the right tools, we instructed it to “act as if you are a curator” and select only from our collection to curate an exhibition.
We wanted to give ChaptGPT as much curatorial agency as possible and began by asking what exhibition themes would be interesting for a university art museum. While it gave us many options, it kept going back to themes of dreams, the subconscious, utopia, and dystopia. Given popular fears that at some point “sentient” AI could either make our world dramatically better or pose a threat to humanity, conversations with ChatGPT about utopian or dystopian futures made the theme topical as well as ironic. ChatGPT began by generating a checklist of works from our collection, but at times demonstrated its tendency to “hallucinate,” a term used to describe when the AI confidently gives wrong information. This happens because the LLM is programmed so that it would rather give you wrong answers than no answer. There were cases where ChatGPT called a work on paper a sculpture, misnamed objects to match its self-picked exhibition theme, and wrote descriptions that differ greatly from the work of art it was referencing. When asked to organize a layout for the exhibition to be in a single 600 square foot gallery, it suggested elaborate lighting, intimate niches to be cut out, and factors that would have involved prohibitive or extensive installation challenges. Ultimately, the results and output from ChatGPT required that human curators intervene.
Talking to the chatbot proved to be difficult too. Its limited memory made it a forgetful collaborator. We were careful as a team to use polite language as it has been proven to respond well to an encouraging and enthusiastic tone. It also wrote labels for the works of art in the show, and it often got things wrong or wrote in vague and misleading ways. Its writing was also starkly different from the language and content that a curator would use, sounding more like a brochure for a cruise line, for example, inviting visitors to “immerse yourself in the rich history and craftsmanship of Guerrero’s ancient civilization…” when writing about a group of Mayan sculptures.
ChatGPT label copy: These stone figures from Guerrero, Mexico, dating back to circa 500 BCE–500 CE, are a remarkable testament to the ancient civilization that once thrived in the region. Carved with meticulous detail, these enigmatic sculptures embody the artistic and cultural heritage of their time, offering a glimpse into the spiritual beliefs and rituals of the past. Immerse yourself in the rich history and craftsmanship of Guerrero’s ancient civilization through these captivating stone figures.
Nasher Response: ChatGPT incorrectly believed these objects to be titled “Utopia,” “Dystopia,” and “Consciousness,” respectively. The ceramic bowl is one of the least impressive examples of pottery traditions across the Mayan region in our collection and has some damage, which would typically exclude it from display. Why ChatGPT selected these pieces is unclear, although in the case of the stone figures, it may be due to their likely placement in burials to accompany the deceased in the afterlife – perhaps an association with the exhibition’s themes of the unconscious and utopia.
This experiment revealed that while curating a show with ChatGPT is possible, it is neither a shortcut nor an objective lens to look at the collection. Between its hallucinations and unreasonable requests, ChatGPT proved to be unreliable. And while the ordering of the exhibition made some visual connections (whether intentional or coincidental), without the ability to make aesthetic judgements and cohesive, thematic choices, its selections fell short of the decisions a human curator may make to put together a show of pedagogical value. In the end, the technology used to produce this exhibit cost $10.71, which, without context, reveals little about the efforts of the student researchers, designers, and curators that put this show together.
However, while ChatGPT’s limitations made human curators necessary in this experiment, the LLM is changing and developing rapidly. Today, users can upload information and knowledge to ChatGPT without knowing code or programming. As of this writing, the newest version, ChatGPT-4, can process images. What was, at times, a laughable and mistake-riddled learning process turned into a revealing experiment on the limitations and capabilities of AI, and perhaps a glimpse into its impact on the future of museum work.Skip over related stories to continue reading article
Julianne Miao is and art historian and curator based in Durham, North Carolina. She is currently the curatorial assistant at the Nasher Museum of Art at Duke University. Her work focuses on early twentieth century American art and the history of photography in the context of race, gender, technology, and commodity culture.