On Tuesday, Jon Ippolito, a professor of new media and digital curation at the University of Maine, gave a presentation on artificial intelligence (AI) and its creative potential. The workshop was part of the Program in Computing Colloquia Series across the spring semester. Participants were encouraged to bring a laptop in order to expand their practice with generative AI. All attendees were anticipated to have a basic understanding of, and experience with, AI because many were coming from the Integrated Program in Humane Studies or Scientific Computing, so Ippolito focused his workshop on theoretical and speculative work as opposed to basics such as prompt engineering or the ethics of AI.
The presentation began with Ippolito asking, “Can [AI] be creative? In what ways can we take advantage of its creativity? And should we be scared about its creativity?” He showed a menu containing four options about various topics on AI, and asked the audience to vote on which they preferred him to spend more time discussing. The menu read: “Unmasking the AI Oracle,” followed by “What we can learn from image generators,” which pertained to AI in media generation, and “Can AI be creative beyond hybrids?” and finally, “When should we use AI?”
After a show of hands, Ippolito focused his lecture primarily on the first option and last two options. Beginning with the AI Oracle, he showed a graphic from a Microsoft internal report about AI, which claimed that it had “Sparks of Creativity,” and then explained the opposing viewpoint of linguist Emily Bender, a professor at the University of Washington, who believes AI is essentially “stochastic parrots,” meaning that it is simply rehashing content already in existence. According to Ippolito, these are the two primary ideas in conflict when it comes to the issue of creativity and AI.
Ippolito then explained that in order to understand AI’s limitations and how it works, it is necessary to look at where it fails. He gave an example of a prompt from a woman trying to plan a hike in the mountains of New Mexico. AI suggested two hikes to her that were real, and then one that didn’t exist. However, the nonexistent hike would have, if it did exist, connected the two other trails together, and the AI had assumed it was there. ChatGPT essentially interpolated between the two real trails in order to make a fake trail, which Ippolito said is “the personality of ChatGPT.”
“[If] we understand this kind of averaging that it does, putting things in places between known things, inventing this other thing, then we have to recognize that this is both a power and a curse,” Ippolito said, as “the average of two facts isn’t necessarily a fact.” A lot of averages do exist, and the average of some facts are facts, but not always, which is where generative AI can run into problems. He proposed the question, “can averages be creative?”
Ippolito then addressed the marketing around AI via the language used when discussing it. Companies use buzzwords to anthropomorphize AI or attribute it more creativity than it has, and Ippolito explained certain words about AI that people should avoid, as they do not accurately depict its capabilities. For example, the word “hallucinate” implies a difference between when AI gets something right or when it gets something wrong — when in reality, either way AI has simply interpolated things in its data set. The same underlying mechanism exists regardless of whether the AI is right or wrong. Ippolito proposed using the word “fabricate” instead, as AI is always building its answers out of something, it’s always “fabricating.”
Ippolito moved to discuss creativity and if AI can be creative beyond interpolation. “The average of two clichés isn’t necessarily a cliché. That’s what I think is actually some of the most interesting ways that generative AI works,” he said. A misconception about AI is that it simply copies content from web pages, but what it’s actually doing is “taking word by word, and even subsets of words, and sort of mashing them up in this big food processor, and then finding connections between the words,” according to Ippolito.
On a three-dimensional scale, if AI can produce a vector sum that is the average of three points, and the date is organized concavely, the new point might not be in the original data set. For example, if there was a hiking trail in the middle of New Mexico where there was a possibility of a path between two existing paths, but no one had ever walked it before, AI has found a new point in the dataset.
Toward the end of his lecture, Ippolito focused on different attitudes toward AI. “There’s a school of thought, sort of postmodern theorists, who say, no, all art is a mashup, and scientific discovery is simply a rehash of what has been before in new combinations. But I think you can say, if you take a spectrum, some combinations are really out there, and some are superficial and closer in that metric of the mathematical space,” Ippolito said. By that logic, AI has the potential to generate brand new connections which have never been previously explored; this is a form of creativity, in its own right.
Ippolito wrapped up the colloquium by explaining his own personal framework on deciding when to use AI. Ippolito believes that a high-stakes versus low-stakes approach is not appropriate. Instead, Ippolito proposes an opportunistic lens versus a prescriptive lens. He used the example of an opportunistic challenge like cracking a safe, where you have many chances to get it correct, versus a prescriptive challenge where you have to disable a bomb, where there are so many wires that if you cut the wrong one, the bomb goes off.
“Classically, we see someone’s trying to crack a safe. Well, let’s say they get it wrong. Who cares? Keep trying, right? Point is, you try a bunch of them, and if you’re good, eventually you crack the safe, right? That is what I would call an opportunistic challenge” Ippolito explained.
He concluded the workshop by thanking everyone for attending, and inviting them to join him for lunch in the Leach Dining Room.
