AI, Art, and Amherst College

On November 17, the Artificial Intelligence in the Liberal Arts Initiative, in collaboration with the Mead Art Museum, hosted an exhibition and panel called “Discussing Dall-E: The Impact of AI on Art.” The panel was moderated by Professor Lee Spector, and it featured Professor Yael Rice from the Art Department and Professor Scott Alfeld, Professor Lillian Pentecost, and Professor Will Rosenbaum from the Computer Science Department of Amherst College. 

Dall-E is a deep learning model, meaning that the computer learns to classify from images or text. The model is able to interpret geometry, which involves organizing images across the internet based on their general appearance. For example, if Dall-E was given the text prompt to draw a “flower,” it would reference the internet for all images that geometrically fit the definition of a flower to inform the image it generates. Using machine learning, it understands what concepts are equivalent to words. It starts with noise (essentially meaningless data), then generates an image from an internal representation based on the text prompt. It forms these representations from collective associations drawn from the internet. 

Professor Spector asked panelists what the prospects are for AI like Dall-E to be considered an artist in the same ways we consider someone a human artist? Alfeld responded that AI “will be able to get a job as a contract graphic designer.” He expanded that AI is quickly able to create many possible versions of an artist’s vision, whereas a human illustrator takes a much longer time to create a single set of options. Rice pointed out that this means AI will be easily exploitable as labor, raising the question “what is this algorithm’s agency?” Rice then went on to explain that the modern debates about what makes someone or something an artist are actually not novel. A similar debate, she explained, arose in the 1850s with the widespread usage of photography. There were many debates about the utility and status of photographs: was it equal to traditional art? Should the ability to reproduce human vision be considered art? In response to these criticisms, early photographers mimicked the subject material of paintings to gain legitimacy. Rice even extended this debate to the 15th century in regards to moveable type, which ushered in the end of illuminated manuscripts. Ultimately, Rice posed Dall-E as an artistic tool, since some contemporary artists are working with AI without deploying the direct replication capabilities of Dall-E. Rather than asking AI to recreate the style of Pablo Picasso or Frida Khalo, these artists experiment with prompts to create new and original works. 

At this point, Pentecost raised the question, does Dall-E have the capacity to produce a style it has never encountered before? Essentially, can it extend beyond the established styles it uses as references to generate a genuinely new depiction of art? Alfeld responded that Dall-E has created a style recognized as novel, which has been dubbed Deep Nightmare. Spector stated that the general conception of Dall-E is that of “glorified predictive typing,” in that it can take a textual prompt and fill in what it thinks the prompt desires as a visual result. But, toying with the more philosophical side of AI, Spector asked whether Dall-E can understand the innovation of a new style. Does it know what it means to be “new” in a meaningful way?

Returning to Pentecost’s original question, Rosenbaum pointed out that a “closer analogy is that Dall-E is an understudy.” It still needs a prompt to generate art, it still needs to be told where to go, but once prompted it can go off and do the artist’s bidding. In this way, Dall-E is “democratizing having an understudy,” he said, as people have increasing open-access to AI software and platforms rather than having to employ people. Signing up for the Dall-E software gives users 50 free credits, and they receive 15 free credits at the beginning of every month. It costs one credit to process one text prompt. Additional credits can be purchased at the rate of $15 for 115. These rates are drastically less expensive than employing an entry-level graphic designer, a position whose minimum average reported salary in the United States is over $52,000.

As a final note on novel style, Spector brought up Alan Turing’s imitation game, a proposed experiment to see whether a computer could imitate human conversation well enough to fool a person asking it questions into thinking it was a human. Spector specifically cited the Lovelace objection, which states that computers cannot generate anything novel. However, he clarified, it is very easy for them to do things that surprise us. So, he asked, is novelty creativity? Or is there a higher threshold, a necessary awareness of being novel?

Spector then pulled up the image of Théåtre D’opéra Spatial, the AI-generated artwork that won in a Colorado State Fair art competition and ignited debate over the medium. Spector used this reference to ask whether AI can be considered just another artistic tool. Rice explained that this artist, Jason Allen, followed the rules of the competition, even disclosing in his submission that he had used the AI software Midjourney. Rice wondered if the level of blowback Allen received reflects a lack of knowledge of art history in general, peddling in the false idea that artists work alone, without help, input, or understudies. A member of the audience then volunteered that Dall-E might be comparable to a collage or sampling in music. In these mediums, the artist drives the direction of their tool, but the materials come from others. Spector pointed out how this raises the question of copyright and the legal principles of what material AI draws from. This connects to the concern of AI companies absorbing artwork generated on their platform, thus streamlining the labor of artists to automatically become property. Who does the finished art belong to – the artist who created the prompt, the person who created the reference material, the company who owns the AI, the engineer who programmed the AI, or even possibly the AI itself?  In the vein of copyright, Alfeld asked how do we qualify an idea as original or as derivative, as much of art is based on inspiration? He said that Dall-E is generative, as it can be trained on our notions of semantics. However, what it creates is “bound by our willingness to interpret it.”

An audience member then posed a question about the potential for misuse of AI in a capitalist society. Rice responded that data breaches are a serious concern, as it raises legal questions over what material is appropriate to be scraped for material. Scraping refers to the process of collecting and analyzing large amounts of data. Rice recollected that one time a woman saw her personal medical records in an AI image database, presumably the result of a data leak,  demonstrating lack of discretion for source material. An audience member asked if any parts of the field focus on preventing these kinds of direct replications, which could also result in forgeries. Alfeld said that it is possible for the training material, whatever was used to inform the newly generated image, to be obliterated. He added this is particularly a concern for military and medical applications of AI. This research is being developed so that if a machine employing AI fell into the wrong hands, it could not be reverse engineered to reveal its original input. Rice concluded the panel by reminding the audience that we need to continue to recognize the specter of bias in meta-data. Algorithms are created by humans, and thus they reproduce human biases.

Looking towards the gallery that accompanied this event, we can see some of the imperfections of AI. The gallery featured half a dozen pairs of images, original pieces from the Mead Museum collections paired with AI-generated recreations. These recreations were made through prompts from the museum’s catalog entries. While none of the recreations were exact replicas, at times it was difficult to tell which was the original. Sometimes certain cues, such as the Dall-E version of an old mosaic fresco not being faded or tinted with age, immediately situated the AI version in a different temporality. The gallery demonstrated Dall-E’s ability to recreate sculptures, paintings, mosaics, cross-stitches, and ceramics. It raises some interesting considerations about accessibility, in that Dall-E prompts resemble alt-text, or short descriptions of images that are read aloud by screen-readers for those with vision problems. Seeing how Dall-E interpreted these descriptions reminds us to consider what images we create for those who rely on narrative descriptions of artwork. How do we help other people see our world?