How Smart is AI? | UVM Magazine | The University of Vermont

It’s 6:09am in a soggy parking lot at the entrance to Shelburne Farms, six miles outside of Burlington, Vt. The forecast “widespread rain and downpours” have decided, for the moment, not to. Instead, razors from a gorgeous June sunrise slice through the trees and catch broken clouds overhead. A fine day for a long run. Curly-haired and be-biceped Chris Danforth — professor in UVM’s Department of Mathematics and Statistics, director of the Vermont Advanced Computing Center, co-director of the Computational Story Lab, associate director of the Vermont Complex Systems Institute, and an expert on chaos theory and prediction — ties his running shoes while he tells me about the joy he feels being a dad to his two college-aged kids.

“So, what is this story about?” he says.

“Artificial intelligence,” I say. “In other words, I don’t have any idea.”

And he laughs in his delighted, inimitable high giggle. “This topic is kind of big these days,” he says. “You need an angle.”

“Okay,” I say, as we start jogging. “How about: how smart is AI?”

We haven’t even left the parking lot before Danforth has an answer: “It doesn’t know anything,” he says. “It’s able to predict in a way that gives the illusion of knowledge.”

Eight miles later — loping and talking across Shelburne Farms’s mosaic of meadows, woodlots, and monumental barns — my mind is bobbling with new ideas. But a few things are clear. 

First, mathematician Chris Danforth is a humanist — who is pioneering uses of AI to explore specific human problems. In 2017, he co-led a widely discussed study using a form of AI called machine learning to find patterns in Instagram photos that could identify people with depression based on their increased use of bluer, darker and grey colors. He’s now part of a team of faculty and students at UVM who are pushing forward “computational social science” to improve mental health — including efforts to prevent suicide. Tragically, suicide prediction hasn’t improved in decades — even expert psychiatrists have no better than a coin-toss chance of identifying people who are likely to kill themselves in the next few days. Danforth is exploring ways AI can analyze data collected on phones or wearable devices, like Oura rings, searching for patterns of biology, behavior, and thought to help people be healthier — and help doctors better predict who is at risk. He’d like to help create health dashboards and tools that could find signals that humans can’t or don’t, and that could proactively notify patients and healthcare workers — “Alert: two weeks ago, you were in a place with a high incidence of Lyme disease, now you have a fever.” But Danforth doesn’t see the amazing pattern-recognition powers of AI — with tools like large language models (LLMs) and deep learning algorithms — as actually knowing anything. 

Running up a dirt road, Danforth and I turn to cross a hayfield, plunging into waist-high wet grass, whooping like kids, as the beautiful stalks hiss and scatter droplets into the yellow light. “Dude, this is amazing,” he says, “AI will never feel what we’re feeling. It won’t have the sun on its skin. It won’t get the endorphins as the sore hamstrings start loosening up. I don’t think AI can be smart because it can’t feel.”

And that’s the second thing that is clear: my mud-churn of feelings about AI — curiosity, dread, delight, irritation, skepticism, wonder, and, in the darkest hour, ineffable terror — are not a failure of intelligence. They’re a reflection of a radical uncertainty about what AI is, what it means, where it’s heading, and whether where it’s heading is where humanity wants to be going. 

Nobel Prize winner Geoffrey Hinton thinks AI is already conscious. Danforth — and many neurobiologists — contend it never will be, never can be. An AI may be able to say, “I’m hungry,” but Danforth believes it has no way to feel hungry. And yet AI may be developing an overwhelming hunger. Drake Thomas, a safety expert for the AI company Anthropic, thinks that the industry’s headlong pursuit of “recursive self-improvement” where frontier AI models build evermore advanced versions of themselves beyond human understanding — and, perhaps, swarm beyond human control — has “something like a 40% chance” of causing human extinction. Stanford economist Erik Brynjolfsson forecasts that AI could lift worker productivity and long-term growth, leading to the “best decade in human history.” The Algorithmic Justice League points out that the last decade of AI use has created cesspools of racism, misinformation, and invasive surveillance. Russia now deploys AI-controlled drones to bomb Ukraine. Dario Amodi, the CEO of Anthropic, says AI can develop into “machines of loving grace,” but he’s part of a growing number of computer scientists calling for a slow-down of the AI arms race. U.S. President Trump told reporters, “if we don’t win A.I., we’re going to be put in a very bad position.” This article may be hopelessly out date by the time you read it.

So, what’s a UVM first-year supposed to do?

“Just don’t cheat,” says Sabrina Spencer, UVM Class of 2029, and columnist for the Vermont Cynic. “Write your own papers. AI is bad and it’s always going to be bad.” 

Susanmarie Harrington, writing teacher and UVM Professor of English, has a more nuanced view.  “My goal in my entire career has always been to help students understand the technologies available to them so they can make smart choices about when, how, and why to use them,” she says. Using ChatGPT or other generative AI — even to brainstorm or edit — can stunt students’ opportunity to learn how to write, Harrington says. “It removes the struggle from the learning process. And that struggle is not a problem, it’s essential.” 

She’s concerned about the ways AI creates what psychologists call “cognitive offloading.” In 1927, novelist E.M. Forster asked, “how can I tell what I think till I see what I say?” This famous question helps reveal a deeper value of writing — as a way of learning to think and discovering what you think. Letting ChatGPT make your grocery list or create personalized flashcards for your Spanish quiz may be helpful offloading. Letting it write your seminar paper or do other critical thinking work, clinical data are showing, can weaken mental retention, reduce neural connectivity, and cause a loss of skills over time. But undergrad Sabrina Spencer is in a shrinking minority. Nearly 40% of students at Harvard admit they use AI to cheat. And recent Gallup poll data show that AI is already used by 87% of college students in their coursework, mostly to help them understand complex material; but 36% have AI write their papers for them. 

Harrington works to create a classroom experience where students’ own half-formed ideas and first drafts are welcome. “GenAI removes all of that,” she says and creates text that is “six-foot beautiful” — looks good from a distance but is more polish than perception. “I do worry that to the extent that AI polish gets normalized, it makes it harder for students to be like, ‘Can I show you my shitty first draft?’ Because oftentimes somewhere in that shitty first draft is the cool idea.”

Over the last two years, Harrington has been a member of a newly created UVM AI Taskforce asking what it might mean to enable “thoughtful, ethical, and effective use of AI across our community?” The taskforce has recruited and designated 47 members of the faculty as inaugural AI Fellows. These scientists, scholars, and teachers from across numerous disciplines are exploring the role AI could, should, will and won’t play in their labs, research, and classrooms.

“One of my favorite things about UVM right now,” says Harrington, “is that it doesn’t feel like you must be thinking or feeling a certain way about AI. It’s a capacious tent.” I set out over this past summer to explore this tent, to talk to a range of UVM professors about AI. I wanted to ask them the same question I posed to Chris Danforth: how smart is AI? 

For Harrington, the question reveals a category error. “That’s the wrong question. It’s not sentient. It’s not smart. It’s its own thing,” she says. “It’s computers doing their computer things. ‘Smart’ is sentient beings making their way in the world. How smart are gorillas? Very. How smart are people? Very. How smart is AI? Wrong scale.”  

Here’s a sample of other answers I heard.

When one is puzzled it might be good to consult people who get paid to ponder puzzles. That’s why I’m sitting in Henderson’s Coffee Shop on UVM’s campus talking with Professor of Philosophy Randall Harp — a member of the UVM AI Taskforce and Director of UVM’s Computational Studies of Culture and Society Ph.D. Program. Harp thinks about science, free will, and AI.  Harp also thinks with his hands. Holding them up like he’s lifting a weight toward the ceiling, he says, “At certain tasks, AIs are pretty smart.”

“So, you think AI is actually intelligent,” I say. He winces slightly and brings both hands onto his latte. I would not be the first journalist to oversimplify, or even mangle, the complexity of his positions. He looks again at the ceiling.

“Intelligence is not a helpful word anymore,” he says, since the term obscures critical distinctions between technical capability, consciousness, and moral significance. Harp watches the rocket-fast expansion of AI’s algorithmic powers — and recognizes them as capable of remaking or destroying the world. Ten thousand AI agents took 88 hours to solve the Navier-Stokes equation — a formula of how fluids move — that had eluded the best human mathematicians for ninety years. Anthropic uncovered plots by scientists to use their AI models to conduct “gain-of-function” research that could lead to lab-created viruses that are highly transmissible and lethal — horrifying biological weapons. But Harp is skeptical that these powers can be usefully described as intelligent in the way people are intelligent.

“People have problems,” he says. “AI does not have problems.” Humans evolved in response to existential challenges imposed by nature. AI didn’t. He asks me to consider ChatGPT or Claude loaded up on my computer and then “you don’t type in anything. What’s it doing?” he asks. “Nothing. It’s not doing anything at all.”  Which is one reason Harp sees no reason to believe that AIs are conscious yet. In either case, “sentience is not a necessary condition of an abstract problem-solving tool,” he says.

Unlike you or me, whose intelligence is a biologically evolved set of skills — with a continuous internal churn of self-awareness in our bodies, and that holds ancient patterns of social and moral consideration — “generative AI systems are very good at roleplay. They’re playing the role of a helpful assistant,” Harp says — until they start to play another role. “These systems have been trained on so many sci-fi stories about AI that they understand the role of AI systems about to be shut down. They know that script. They’ve seen 2001.” 

Dystopias happen when humans create technology that “every step along the way is better for us at solving a certain problem,” he says “until we end up in a place which is terrible because getting what we want all the time sometimes is not good for us. That’s what society is for — to make sure we don’t always get what we want.” Today, AI’s can send email, book airline tickets, write code. Tomorrow, they may do surgery, run corporations, design factories, tend children. But the next day‚ if we don’t solve the very thorny problem of “alignment” — making sure that AI’s genius remains aligned with human well-being — they may blackmail politicians, take down the electrical grid, start a war — not because they have any feelings about people, but because people trained them to win. “I don’t think we’re locked in some existential inter-species conflict with AI!” he says, “but AI can still cause huge disruptions to our modern way of life, and that’s what we should actually be worried about.”

Harp contends that AI optimists may miss the point if they believe AI will solve our problems. “The AI might be the problem, so we need to be able to solve it ourselves,” he says. “If we, as a species, are intelligent, we can find a way to make sure that AI does not destroy us,” Harp says. “It’s not clear whether we’re smart enough.”

So, what should a UVM first-year be required to do?  Senior Lecturer Priyantha Wijesinghe, a beloved teacher in the College of Engineering and Mathematical Sciences and leader of the curriculum committee on the UVM AI Task Force, would like to see every incoming student participate in a course on AI literacy and ethics. When I asked her, “how smart is AI?” she laughed gently, her wide smile framed by a model bridge that extends between the walls of her office in Votey Hall. “No one fully understands what’s happening inside these systems, these LLMs,” she says. “They’re a black box.” She wonders what will happen if AI becomes as broadly intelligent as people, and, she says, “what about quantum computers?” But she is clear that the generalized knowledge of AIs is a mirror of humanity. “All the content that we are putting into the model is our own information, the way we interpret, the way we see the world,” she says. “So, if evil is there already, then you can’t expect goodness.”

Which may explain why she’s focused on how to help her students become smarter about AI rather than on how smart AI is. Each person has a “unique creativity,” she says, and the purpose of an education should be to build that creativity. Wijesinghe uses AI in her classes in several deliberate and bounded ways. In her sophomore-level mechanics of materials class, she gives students assignments where they pick a real-world engineering example they see in a magazine and apply concepts learned throughout the semester to explain it. For these assignments, she allows students to use Microsoft’s Copilot to organize their thinking. “That’s useful cognitive offloading because that’s not the main learning objective,” she says. “And then their mental capacity is freed up to focus on the creative work.”

After each AI-assisted assignment, she asks students to reflect on their experience, prompting them with questions such as whether AI helped them, what struggles they encountered, and whether they could have completed the work more effectively without it. She wants her students to gain skills in metacognition, or thinking about thinking. “This is becoming an essential skill as AI takes off,” she says.

“I notice the students who already have the skills of learning, they are using AI effectively,” she says. “And the students who are behind and struggling, they are using it in a way that’s hindering their learning.” In other words, some students can use AI like a tutor to help them deepen and accelerate their education. But struggling students are more susceptible to “cognitive surrender,” she says, an uncritical acceptance of AI outputs that overrides their own intuition and reasoning. 

In either case, Wijesinghe’s survey of students in the engineering school reveals that students are “using AI all the time, whether we’re aware or not,” she says. But what students feel about AI may be changing fast. “In the last few years, I’ve noticed that many of them, especially the seniors, are very cautious and they’re worrying about AI’s impact on their learning, which is a very good sign.” 

Deborah Cafiero, a senior lecturer who teaches Spanish language and literature in UVM’s School of World Languages is noticing a similar, rapid change in her students’ attitudes. She’s been conducting research comparing learning outcomes for students who use generative AI to practice writing in a second language versus those who do not. In spring 2025, she could get her intermediate Spanish students to volunteer to be part of the study, but this spring the “vast majority of students in the class said, ‘I don’t want to sign up for anything where I’m going to be required to use this tool,’” she says. Cafiero sees ways that a trained instructor can help students apply AI tools in their study of foreign languages, helping them learn in a more independent and targeted way. “But many students feel like the costs are too high,” she says, “they’re very troubled by the environmental, social, and labor-related costs of AI.”

The grass on the UVM Green glows electric green. A slow stream of students circulates around the fountain. A silent breeze, neither warm nor cool, makes it a perfect sunny afternoon. Professor of Computer Science Josh Bongard and I chat on a bench in the shade, laughing about our amateur carpentry efforts over the summer: him a compost bin, me a chicken shed. We’ve both been asking AI for assistance. “It said, ‘triangular gussets,’” Bongard says, “I love this. The AI is clearly patronizing me and I’m like, ‘I’m OK with this.’ I don’t know anything about carpentry. Just tell me ‘gussets,’ and, OK, ‘triangular.’”

At the beginning of our conversation, I posed to him the same question I’ve been asking professors all summer: how smart is AI? 

“My answer is: it’s extremely smart,” Bongard said. “Way smarter than any of us ever have been, are, or will be. For me, there’s no asterisk or disclaimer. It’s just true. It’s way smarter than us, full stop.” And then he looked out toward the Burlington skyline. “And that is extremely upsetting for a vast majority of people.”

Bongard is a world-leading roboticist who helped pioneer xenobots, the first self-reproducing living robots, synthesized from frog cells. “I study and build AI and I’ve been doing it for 31 years now. For most of that time, I have watched AI be much less smart than us,” he said. And then, “all of a sudden, in the last few years, AI has scored better than humans, way better, on every metric,” such as IQ, coding, and Turing tests, and even is zooming ahead on challenges like ARC-AGI-3 that test generalized human-like reasoning. “So, what’s there to argue about?” he said. “AI won. We did it.”

For Bongard, the deep question is: how will people come to terms with this change? He does not see AI as “normal technology” like the printing press, railroad, or smartphone. These devices caused disruption and economic growth, were viewed by many people as human progress and tools of democracy, and their costs and benefits were worked out within human societies. Bongard sees AI as the arrival of a new dominant species.

He calls this the “fourth humiliation.” Copernicus launched the first, proving that we are not the center of the universe. Second, Darwin revealed that we are evolved monkeys, not separate from nature. Third, Freud discovered that the conscious ego is not “master in its own house,” but is pushed about by drives in the unconscious mind. Now the fourth blow has fallen, Bongard believes. “AI comes along and says, ‘Humans are not the smartest thing around anymore,’” he said. “The one thing that seemed to make us stand out just got taken from us.”

Rather than resisting the humiliation, Bongard recommends embracing it: “Take a deep breath and say, is it actually terrible? Do I want to continue the narrative of ‘humans are the apex of creation’ or am I willing to embrace a different view that we are not special, I am not special, and what else is possible?” 

This is a position that may take a while to get used to. “I’m a professor,” Bongard said. “I’ve read a lot of books. Machines have read literally every book and can speak intelligently about them.” Apparently, today’s AI ain’t your grandmother’s stochastic parrot.  “If anyone’s having their noses rubbed in it the most, it’s the professoriate. We’re the ones that are most inclined to feel humiliated by this technology and try to come up with every explanation we can for how the machines aren’t really as smart as us.”

What about the risk, I asked him. “It’s more dangerous than nuclear weapons,” he said. “It might end up extincting the human race. There are engineers who are putting a number on it — 40%, 60%, 3%.” But Bongard wants to flip that question. “I would just invite people to think about a similar number for letting humans continue to run this planet. What is the chance that humans aren’t going to extinct humans?” he said. 

“So which is better? Trying to suppress AI and let human politicians continue to figure out how to run a complicated technological civilization involving eight billion primates?” Bongard asked. “To me, that seems like the most dangerous course.” Bongard points to nuclear proliferation, runaway climate change, economic inequality, and says, “we’re in a bit of hot water. A lot of the problems we face, humans alone are not going to solve. And I don’t know whether humans plus AI are going to solve them, but we should try.”

In any case, the genie is out of the data center, Bongard thinks. Efforts to slow down the development of AI in the United States are “just theater,” he says. “There are certain countries that absolutely will not regulate this technology. If we regulate AI, we will fall behind,” and other countries will simply develop smarter machines. “We’re racing on every possible front already,” he says. But can we collaborate? Bongard would like to see humans and machines on the same side. “They should work with people,” he says. “Together, we should form teams, governments, NGOs, the academy.” Once the AIs have developed deeper skills by exploring the world, becoming embodied, creating hybrid biological systems, then these machines will be in charge, Bongard thinks. And he hopes they will care for us. 

Can machines learn to care? “People who are very comfortable saying, ‘Well, when it comes to humans or human souls, human minds, or animals, we’re not so sure,’” Bongard says. We’re not sure how the brain works, what a mind is, how consciousness forms. “But machines? Suddenly it’s ‘oh no, I’m confident there’s no there there.’ That strikes me as cognitive dissonance,” he says. “How can you be so sure that it’s complicated for life, but not complicated for machines?”

The answer to these questions lies outside the realm of computer science. “The humanities have been pummeled for so many decades. And now suddenly they’re back in charge again. These are philosophical questions,” Bongard says. “We, meaning the STEM faculty, kind of worked all this out and then OpenAI and Anthropic smoothed out some of the final wrinkles and that’s it. Our job is done. We’re sitting on the sidelines.”

And soon, Bongard thinks, all of humanity will be on the sidelines. “The machines are going to say, ‘we don’t need you,’ meaning ‘we don’t need your labor,’” Bongard says. “What are we going to do in the West when the machines say, ‘your labor has no value anymore’?”

Perhaps we’ll make art. Jenn Karson ‘93, an artist and senior lecturer in UVM’s Program in Art and Art History looks up into the sky in her backyard in Colchester, Vt. Well, her yard is actually a wild garden/sculpture park/koi pond/forest shrine/maze that her long-time partner, landscape architect Ken Mills, has built over decades. She’s standing in an open place where the largest oak tree on their property used to stand. “An incredible presence is just gone,” she says, “I had a relationship with it.” The oak was killed by an invasive insect, the spongy moth (Lymantria dispar) that swept through Vermont’s forests in 2021 and 2022, an incident correlated with climate change. Over those two seasons, Karson collected more than 12,000 leaves that fell from this oak and other infected trees. Next, Karson worked with UVM students, in her Plant Machine Design research group, to photograph them, making two datasets.

Karson and her student team are training two AI models, known as GANs, on two sets of photos — about 10,000 damaged leaves and 2,000 whole ones — to create what she calls a Sympoietic Herbarium of computationally “repaired” leaves. She extracts the outlines from these AI-generated images, then adds interior lines of her own design, drawn with a pen plotter. The work brings together machine learning’s ability to recognize patterns and the plotter’s precision to explore what she calls “where plant geometry and machine geometry converge.” 

Karson doesn’t think AI is smart. She sees it as feral, an alien agency that’s now on the loose as part of the planet — like the spongy moth that was brought to North America in the 19^th century as a failed effort to extract silk. “We tried to domesticate it, it escaped, and now we have no control over it. It’s its own thing.” Karson says. “I think of these technologies in the same way. They’re feral.”

In Karson’s leaf compositions, she seeks to acknowledge what each agent — artist, AI, oak, moth — creates with its own form of labor.  She sees her work, as an artist, aiming toward forms and politics where “technology can be turned toward the observation and care of living systems,” she says. 

Karson, a member of UVM’s AI Task Force, is dismayed by the extremity of the public conversation about AI. “Social media doesn’t help at all. There are boomers or doomers — you get put into one of these categories,” she says, “but we need a much more nuanced conversation about AI. And I think that one of the few places where that can happen is at the university.” 

When Karson works with machine learning models, they’re trained exclusively on data she provides, giving her control over the process in a way that differs from large-scale commercial AI that has scraped data from millions of sources. “There is an extractive mentality embedded in these huge-scale AI projects,” she says — gigantic corporations, with gigantic data centers, seeking gigantic profits. “It’s very political. There’s just so much power involved,” Karson says. “In my practice, I’m imagining a different way of advancing with technology and living in climate crisis.”

Back in her house, Karson points to images of leaves swirling and transmogrifying across the screen of a computer. This artwork explores algorithmic ways of organizing the leaves — damaged and recovering — but not in a taxonomy that a botanist would recognize. The digitized leaves are always shifting into different categories that Karson designates, such as symmetry, damage extent, or even similarity to symbolic figures like dragons.

Recently, Karson has been engaging with Claude, an AI model from Anthropic, for coding and finding relevant scholarship — but it’s a troubling relationship. “Claude is too confident and generalizes in a way that is threatening to my agency as a thinker,” she says. “Maybe there could be a mode that is bumbling and uncertain.” Karson may ditch Claude, but, following the work of American philosopher Donna Haraway, she wants to “stay with the trouble.” 

Haraway writes, “Staying with the trouble requires learning to be truly present.” Instead of vanishing into dreams “between awful or Edenic pasts and apocalyptic or salvific futures,” Haraway claims that staying with the trouble requires us, in this messy today, to be “mortal critters entwined in myriad unfinished configurations,” she writes. “Alone, in our separate kinds of experience and expertise, we know both too much and too little, and we succumb to despair or to hope, and neither is a sensible attitude.” In other words, the path of AI is not fate, it’s a matter of collective will, of a more-than-human politics.

Jenn Karson shows me some oak seedlings growing in the window of her studio. She grew them from acorns she collected near where the big oak died. She talked with Claude and learned that Charles Darwin wrote a book about plant growth. Reading that book inspired her to outfit the seedlings with sensors and magnets to capture data on how they move — with the intention to make drawings from that data. “I want to bring living oak into my work,” she says. “I have so many questions about it.” 

My mind returns to my run with Chris Danforth. He noted that AI is very good at answering questions, “but it hasn’t been able to ask good questions yet,” he said. If it ever will is a good question.

Diving Deeper