Hyper Text

---

Synthetic text extrusion. Virtual teaching assistants. Illusions of mastery. Silicon Socrates. Four years after the debut of ChatGPT, higher education is starting to look different.

By Trey Popp

---

---

On September 15, 2025, a dozen freshmen from the College of Arts and Sciences gathered in the Neural and Behavioral Sciences Building for the second meeting of a first-year seminar offered through Penn’s undergraduate program in Science, Technology & Society. By a consensus they’d reached the week before, the students piled their silenced cellphones in the back of the room before casually sorting themselves around three circular tables. Laptops remained tucked in their bags as they produced pens and pencils to take notes by hand. At the head of the classroom, Rob Nelson drew a long line across a wall-mounted markerboard. At one end he wrote the year 300,000 BC. Near the other end he wrote 2022.

The idea was to mark some precursors to the awesome and anxiety-stricken moment of Right Now. For most of our species’ existence, communication was a strictly oral affair, noted Nelson, a lecturer in Penn’s Graduate School of Education (GSE) who spent 18 years as a higher ed administrator overseeing academic technologies and program development. The first writing systems sprang up roughly 5,000 years ago, though more as a way of counting livestock than refining ideas. About 2,500 years later, an Athenian philosopher named Socrates criticized writing for implanting “forgetfulness” in the souls of men to whom it offered not “true wisdom” but “only its semblance.” Men who outsourced their memories to scratch marks would “appear to be omniscient” but “generally know nothing,” he argued—not least because inscribed letters are inert and unchanging, thus forever “silent” in the face of any questions a reader might wish to ask them.

We know about Socrates’s dim opinion of the written word, of course, because Plato wrote it down. Though good luck finding a copy for the next millennium or two. Johannes Gutenberg’s invention of the moveable-type printing press in 1440 changed that. Humanity’s long epoch of “primary orality” gave way to what some scholars call the “Gutenberg Parenthesis,” as printed matter began to spread ideas far wider than word of mouth could hope to do. That in turn begat the defining feature of life for every freshman in this room, because “in order to participate in culture during the Gutenberg Parenthesis,” Nelson observed, “mass education became the enabler.”

So it has remained even after the rise of radio and television propelled us into an era of “secondary orality,” broadcasting speech and images around the globe to enlighten (or enfeeble) the masses irrespective of their ability or inclination to read.  Yet even all that—including its 21st-century juicing by broadband internet—has acquired the sepia stain of ancient history since 2022, when OpenAI released ChatGPT [“Alien Minds, Immaculate Bullshit, Outstanding Questions,” May|Jun 2023].

Most of these seminar students were high school sophomores at that time. They entered Penn as the fourth undergraduate cohort to take college classes in the era of generative AI chatbots; the soon-to-graduate Class of 2026 has been swimming in this water for four years. So in a sense, these students constituted Exhibit A in Nelson’s freshman seminar, which was titled simply: “How Is AI Changing Higher Education?”

That was the same question I wanted to answer, along with its implicit corollary: “And what should be done about it?”

It seemed like a tough assignment to teach a seminar tackling questions that have largely flummoxed higher ed institutions themselves. But Nelson and his students were game for me to sit in on their three-hour sessions—as well as the two-hour meetings of a concurrent section being offered to GSE master’s students in higher education administration. I did so over the course of the fall 2025 semester, during which I also interviewed 16 Penn students, a handful of faculty members, and sampled three other courses that grappled in one way or another with the educational challenges and opportunities presented by AI.

I set out with little more than curiosity, a hunch that AI is changing more about higher education than just how to cheat, and a gut feeling that I’d gain more insights from students than their teachers.

MARKET RESEARCH

The first thing I learned was that hundreds of Penn undergraduates will gladly wait in line for an hour or more to score a complimentary Owala water bottle. On September 8, Google representatives pitched a tent on campus to offer students a free year of access to Gemini Pro. At the time, this large language model (LLM) was widely regarded as a laggard in the generative AI space. Presumably that’s why every student I asked told me that the main reason they’d joined the queue was for the Owala—except one who had her eye on free ice cream.

About LLM chatbots more generally, opinions and experience varied.

Amelia Carroll, a freshman in the School of Engineering and Applied Science (SEAS) from Cleveland, had tried ChatGPT just enough to lose faith in it. “I would use it for research, and it gave me bad information and bad links,” she told me. “Most people I know are generally pretty skeptical.” But not everyone, she quickly added. “There are a lot that use it pretty consistently for everything. I’m in a scientific computing class, and somebody didn’t know how to do an assignment, and so they just asked ChatGPT to do the entire thing.”

This split perspective turned out to be a theme. When I asked College freshman Denisia George how she planned to use Gemini, the aspiring neuroscience major doubted she’d spend much time trying. “Most of them are stupid,” she said, “and it’s probably going to be stupid, too.” Mariah Lewis, a freshman political science major standing nearby, chimed in with agreement. “I try not to use AI,” she said. “I feel like it’s definitely dwindling how much we actually have to think in life.”

But if chatbots are stupid—or even just mediocre—why would it take any effort to resist using them? And why were “some people” turning to them for help?

Saving time was a common answer. “It’s club application season,” observed Emily Leung-Kaplan, “so a lot of us are in a time crunch, and I know most people use AI to save time.” Yet Leung-Kaplan, a Wharton freshman from Seattle, professed a different motive. “I use it to just help me understand problems a little better. Sometimes, if I don’t have other resources readily available, Gemini is just easy to access, and I can ask it to explain the concept to me.”

For Emilia Cropf, a College freshman from Philadelphia, saving time and soliciting explanations were two sides of the same coin. “If I’m researching a particular topic, or I just need clarification on something I’m learning in school, it’s really helpful to have a concise and exact explanation behind whatever I’m learning—instead of watching, like, an hour’s worth of YouTube videos or just individually googling and searching for information.” The pre-veterinary student didn’t trust Chat (as students often call it) to summarize a book; it had burned her in the past. But it seemed on firmer ground with STEM concepts. “Sometimes I’ll use it to explain something like a process that I already vaguely know,” she said, “and it helps to cement and solidify an idea.”

Leung-Kaplan and Cropf were getting at two issues that now shape virtually every student’s college experience. Chatbots aren’t 100 percent reliable—and the only way to competently assess their outputs is to know a fair bit about a topic already. But they are available 24 hours a day, answer questions with superhuman patience, and never pass judgment.

“I’m gonna be so honest,” SEAS freshman Joy Wong told me. “The TAs here are very specific about what you can and cannot ask them. So if you have to turn to ChatGPT to understand something— because it’s on the homework, and you’re not allowed to ask about the homework—you might as well.”

Wong, who is majoring in computer science with a concentration in AI, had high praise for a specific bot called CalcStar Blue. It is one of several “custom GPTs” developed by Robert Ghrist, the Andrea Mitchell University Professor of Mathematics and Electrical & Systems Engineering, to mesh precisely with his mathematics classes, right down to week-by-week learning goals. Indeed so many students mentioned CalcStar during my reporting that I reached out to Ghrist toward the end of the semester. (More later about that fascinating conversation.) But despite the excellence of that specialized tool and the accessibility of off-the-shelf LLMs—or perhaps because of those things—Wong worried about the ways they might undermine her education.

“I think people get too reliant on it. Like, even now, I’m feeling a little bit too reliant for basic concepts,” she confessed. She worried that “people will stop thinking critically—or asking questions directly to the professor, who knows a lot.”

Oresta Hewryk, a fourth-year PhD candidate in biology, had witnessed that dynamic firsthand as a teaching assistant. Hewryk’s doctoral research focuses on using “machine learning and AI pipelines” to probe data on chemical compounds from traditional medicinal plants in search of new therapeutics. She planned to test Gemini Pro’s ability to automate the labor-intensive process of combing through scientific papers. So her work epitomizes the excitement many scientists and scholars have about AI’s potential to accelerate research productivity. Yet she fretted about the “stark difference” separating her own undergraduate experience from that of today’s students.

“There’s more of an expectation that everything is so readily available, so you don’t have to go out and do the research yourself,” she told me, referring to how easy it is to hit up ChatGPT for explanations. “That’s a tool, if you know how to use it. But it’s also an issue when you’re asked to come to your own conclusions, and you are instead asking an LLM to come to those conclusions for you.

“You can’t get an LLM to do your wet lab work for you,” she explained, “but you can easily say, ‘This reaction caused bubbling and turned green. What does this mean if these are all the chemicals I used?’ And that kind of distorts the best part of research, which is having to come to the conclusions yourself. … And I think that’s where the gap is starting to form—at least what I see as an educator—where students are not relying on their own thoughts. They don’t trust themselves to come to that conclusion, but they trust an LLM to come to that conclusion for them.”

Hewryk allowed that an LLM might serve up trustworthy answers in the context of a first-year biology lab. But she wondered whether LLMs might be eroding critical skills that they are fundamentally unable to replace. “If you’re doing cutting-edge research as a PhD or a PI [primary investigator],” she said, “the whole point is coming to a novel conclusion or a novel prediction. And if AI relies so much on preexisting research and methods, is it going to be able to make that novel prediction?

“So I’m seeing this now in undergrads,” she concluded, “but I’m kind of curious to see how that will affect graduate students and new PIs in 30 years.”

Some emerging research gives cause for concern about the prospect of AI-driven “de-skilling.” A recent study in The Lancet Gastroenterology & Hepatology, for example, found that physicians who used AI for three months to help them identify precancerous polyps during colonoscopies were able to detect significantly fewer of these lesions once the tool was taken away. Yet a previous meta-analysis in Gastrointestinal Endoscopy found that AI assistance raises overall detection rates by about 20 percent—which is, after all, the whole point of the procedure. There is little question that AI tools can be effective. In some realms they may prove indispensable. The challenge facing universities is how to keep them from short-circuiting the educational process. In a 2025 study, researchers from MIT compared the brain activity of subjects who wrote essays in response to SAT prompts with and without the aid of LLMs. LLM users “struggled to accurately quote their own work” just minutes after completing it, and EEG recordings of brain activity showed that they “consistently underperformed at neural, linguistic, and behavioral levels. These results raise concerns about the long-term educational implications of LLM reliance,” the authors concluded.

Many undergraduates I talked to were alive to this concern—but caught in a bind. Sensing that AI tools are likely to shape their professional prospects, they want to become proficient users. But many worried that the better the tools get, the riskier they become for students to rely on. And many feared becoming dependent. “I try to use any sort of AI as a last resort,” said Ryan Steinfurth, a College freshman planning to major in math. “Just because it kind of does it all for you. Which is nice—but if you’re trying to learn something, it’s not the best way.”

---

DEMYSTIFYING AI

Rob Nelson kicked off the September 29 meeting of his freshman seminar by writing four statements on the markerboards lining the room:

All models are wrong, but some are useful.

Prediction is hard, especially about the future.

Correlation is not causation.

The map is not the territory.

Splitting into groups, the students illustrated each aphorism with case studies drawn from the book AI Snake Oil: What AI Can Do, What It Can’t, and How to Tell the Difference, by Arvind Narayanan and Sayash Kapoor, of Princeton University’s Center for Information Technology Policy. The examples ranged from weather forecasting to earthquake prediction to criminal-justice risk assessment algorithms [“Black Box Justice,” Sep|Oct 2017]. The ensuing discussion touched on a seminal study known as the Fragile Families Challenge. This 2017 project invited 160 research teams to build machine-learning models to predict six life outcomes (such as a child’s grade point average or likelihood of household eviction) for children whose families had been part of a 15-year longitudinal study. Given nine years of detailed survey results that encompassed millions of data points across hundreds of variables—including height and weight, family composition and income, cognitive assessments and linguistic fluency, and household discipline practices—the teams were challenged to predict outcomes that had been recorded in year 15. All this academic and computational firepower produced a decidedly underwhelming result: even the very best models performed only slightly better than a coin flip.

One of Nelson’s goals was to demystify AI for his students. AI Snake Oil’s useful categorization of its three main flavors—generative AI, predictive AI, and automated decision-making AI—gave them a lens to peer past the breathless branding and recognize these products as tools in a more prosaic sense, with particular strengths and weaknesses that suit them better to some tasks than others. (LLMs are just one type of AI, but they drive the chatbots that have become synonymous with the term for most undergraduates.) Nelson had also charged each of his students with identifying a thinker with a compelling perspective on AI and sharing their ideas with the class.

The diversity of their choices gelled well with the seminar format. One chose Emily Bender, a linguistics professor at the University of Washington who has belittled chatbots as “synthetic text extruders” and “plagiarism machines” that are incapable of understanding their own output—but lull us into believing otherwise by dint of our “human tendency to attribute meaning to text.” Another gravitated toward Alison Gopnik, a University of California psychology and philosophy professor who views LLMs as a cultural technology akin to books, library catalogues, and the internet; and just as we don’t blame a library catalogue for a book that contains an erroneous claim, we shouldn’t let the weaknesses of LLMs distract us from their strength: rendering “complex, large, and inchoate” bodies of knowledge “more visible and tractable than they were in the past.”

Several students focused more narrowly on the implications for higher education. One cited a recent New Yorker essay by Ted Chiang, who wrote, “Using ChatGPT to complete assignments is like bringing a forklift into the weight room; you will never improve your cognitive fitness that way.” A couple others offered a partial rebuttal by quoting Wharton professor Ethan Mollick, a chatbot enthusiast who has incorporated AI into his classes more swiftly than probably any other professor at Penn. “Our fear of AI ‘damaging our brains,’” Mollick has written, “is actually a fear of our own laziness.” College freshman Luke Castelli, who struck me as the seminar’s most eager and discerning adopter of AI tools, offered a reflection based on a summer internship with an AI consultancy in his native Utah. Its founder, he noted, liked to say that “AI readiness isn’t about the right tools, it’s about the right judgment.” So for universities, Castelli suggested, “one solution is to create more courses on AI. If we understand the risks of AI, we can exercise our judgment.”

A number of Penn faculty members have answered that call. One is astrophysicist Bhuvnesh Jain, the Walter H. and Leonore C. Annenberg Professor in the Natural Sciences. This past fall, approximately 60 undergraduates took his “Introduction to AI: Concepts, Applications, and Impact.” By providing an “overview of how AI works, how it is applied, its limitations and where it might be headed,” according to its course description, the class aimed to convey “everything you need to know about AI before stepping into the real world.” (Peter Struck, the Stephen A. Levin Family Dean of the College of Arts and Sciences, sat in frequently for the weekly sessions.)

Michelle Zhang, a freshman from Los Angeles intending to major in classics, gave Jain’s class high marks. “It’s very mathy, a lot of it,” she told me midway through the semester. “We talk about anything from the technical side—like here we’re talking about squash functions” she said, pointing to a page of exquisitely handwritten notes, “and also the social side, like what do we think is going to happen in the future?

“It’s been really helpful,” she reflected, “in terms of allowing me to see it as less of this, like, mysterious force—and more as just what it is, which is a tool that’s been built to do a specific task.” Some students overestimate what it can do, she reckoned, and others are “incredibly scared of it.” Both attitudes could hold people back from discovering good uses. For instance, she routinely snapped photos of her inked notes and uploaded them to an LLM for digitization. In line with research indicating that hand-writing notes improves knowledge retention, Jain prohibited laptops in his classroom. “But he actively encourages all of us to use AI to study,” Zhang said, for example by prompting it to generate practice quizzes based on notes and course materials.

Rob Nelson took a similar tack, pairing a decidedly low-tech classroom approach—“I’m a big believer that PowerPoint has destroyed class presentations,” he told his students—with assignments involving substantial AI experimentation.

For their September 29 meeting, he tasked the freshmen with using generative AI to produce an automated acceptance or rejection letter from an official undergraduate club. One prompted ChatGPT to reject “Rose” from the Wharton Pickleball Club in “a mean and pretentious manner.” In the hard-hearted Hunger Games of competitive-entry Ivy League extracurriculars, this kind of thing apparently happens all the time—and Chat passed the test with flying colors. “That’s pretty pretentious,” the student laughed ruefully after reading the letter aloud. But the previous week, Nelson’s master’s students had made a less captivating discovery. When prompting an LLM for editorial advice on short essays, the students were initially impressed by the tailored feedback it provided—only to discover that everyone else had received pretty much the same generic suggestions. That’s not to say that it wasn’t potentially useful; much writing suffers from common ailments. But it was a reality check on the trustworthiness of tools that are expressly trained in the art of impersonation.

Nelson’s freshmen also created “virtual teaching assistants.” Using an educational platform called BoodleBox, they essentially layered specific bodies of knowledge atop an LLM (they could choose between several well-known models) to design a customized chatbot whose goal was to help them succeed in the class. The results were by turns funny and fascinating. One student tried to infuse her bot with the personality of a Japanese anime character (or that’s how I took it, anyway). Its initial amusement wore off with rigid repetition, but that in itself was a useful lesson. Another tried to give his a penchant for edgy sarcasm (perhaps to counter the cloying sycophancy that many chatbots employ to flatter and hook their users). He lamented that it was too tame—except with Nelson, whom it insulted freely when he graded the assignment. Everyone uploaded the class syllabus and readings into their bot’s knowledge base, but one student went a step further and included a bevy of articles bearing Nelson’s byline. He figured that aping the instructor’s writing style could hardly hurt his grade. (Nelson, surprise surprise, thought this bot was pretty slick.)

Toward the end of their meeting, I posed two questions to the seminar group and asked for a show of hands. “Who thinks artificial intelligence is a good name for the tools we’ve been discussing?” I asked. To my mild surprise, only one person thought so. Some discussion ensued about the nature of intelligence and the difficulty of defining it. Then I asked my second question: “Who is glad that they are in college at a moment when these so-called AI tools exist?” Every single hand went up.

ENDLESS FORMS, MOST BEWILDERING TO BEHOLD

Robert Ghrist agreed that intelligence is difficult to define. But the professor of math and electrical and systems engineering, who also serves as Penn Engineering’s associate dean for undergraduate education, was not ready to dismiss the term “artificial intelligence” as an exercise in branding hype.

“I do find these models of AI to be doing much more than just searching things on the internet and reporting—doing much more than just trying to predict the next thing it should say,” he told me. “They are reasoning. They’re reasoning at a level of complexity that exceeds that which I see from most humans.”

“Most humans” is a broad and occasionally disappointing category. I asked Ghrist if he’d ever sensed an AI model outstripping his own intellectual capacities.

“In certain aspects, yes,” he replied. “Good AI” had “stunned” him with its reasoning in two realms. One was “research-grade mathematics”—no small feat for a man who has led over a dozen projects in algebraic and computational topology for the likes of DARPA and the NSF. The other related to Ghrist’s abiding love for English literature, which for a time had been his intended major as an undergraduate. “Having conversations with my favorite LLM about Blake, about Joyce, about Milton, about synthesis ideas, I’ve gone to places that were stunningly creative,” he said. And they were not just “being ripped out of something from the internet,” he professed, “because I know the literature.”

For Ghrist, determining what qualified as “good AI” had much to do with the context. He told me he used Anthropic’s Claude for writing but often turned to Elon Musk C’97 W’97’s Grok for creative idea generation. “Grok is just bonkers. It’s just nuts—in the good and the bad way—and I love it for that.” Google’s new Gemini 3 was “outrageously good” at technical problem solving, he said, but “I would never use it for writing.” He suggested that their differences from one another, and from human minds, added a new layer of complexity to a very old challenge.

“Aquinas dealt with non-human intelligences quite a long time ago and actually tried to categorize them,” he said, surprising me with a reference to the 13th-century Catholic theologian and philosopher. “I don’t think he finished the job.”

“The job got harder as of late,” I suggested.

“Much harder,” Ghrist concurred. An historical analogy occurred to him. “Imagine living back in the time when microscopes came into being,” he mused. “We could grind the lenses and really zoom in. What did we discover? Oh my God, there’s life everywhere—at microscopic scales, a multitude of forms that is unimaginable! They’re crawling all over you. They’re responsible for you getting sick. They’re responsible for you maintaining health. They’re floating in the air—you’re sucking them in your nose right now. And then, below that, there’s viruses. Which, are they even alive?

“I believe we are at a similar moment—not for categories of life, but for categories of intelligence, maybe categories of consciousness,” he said. But “it won’t be the end of the world. It’ll be really confusing, and people will argue and fight about it, and we’ll discover just weird stuff—and then we’ll get used to it.”

---

COEFFICIENTS OF FRICTION

However you rate the current intelligence of Claude or Grok or Gemini, enthusiasts are wont to respond with a widespread article of faith. “AI today is the worst AI will ever be,” as Luke Castelli put it, echoing the common refrain. “It’s only going to get better.” If so, the choices facing educational institutions will only grow harder.

Josh Miller, a freshman in Nelson’s seminar, hit the nail on the head by citing Marc Watkins, who directs the AI Institute for Teachers at the University of Mississippi, where he is a lecturer in writing and rhetoric. “What generative AI gives to the user is a frictionless experience,” Watkins has written. “Regardless of what you ask a chatbot, the response is always instantaneous, confident, and reasonable sounding. Users trade speed over accuracy and cede their critical thinking to a technology they likely don’t understand, just that it quickly gives them a response.” But “learning is friction,” Watkins stresses, which “puts education at odds with the current wave of frictionless chatbot interfaces dominating the market.”

In a mid-October interview, Miller told me he had grown wary of one of the most pervasive uses of LLMs among students: summarizing assigned readings. By senior year of high school he was doing this a fair bit. “But in college, I’ve generally been avoiding it,” he said. A summary was just too easy to read quickly before moving on. “I don’t remember any of the information,” he said, “versus when I take the time to actually sit down and read, like, a chapter in a book—I’m actually going to remember.”

But here’s the thing: virtually everyone seems to be using Chat to summarize text. One upper-level philosophy seminar I attended epitomized how normal this has become. A student called upon to explicate the thesis of an assigned essay during a heady discussion interrupted herself midway through to say, “Just for full disclosure, I’m referring to a ChatGPT summary right now.” To which her instructor replied, “That’s OK. So am I.”

The most intriguing perspective I encountered on this issue belonged to Michelle Zhang, the aspiring classicist. Among her peers, she said, “one of the most common uses of AI is putting in a 15-page reading and saying, ‘OK, give me five pages, max.’” Insofar as a skillful summary dispenses with “fluff,” she speculated that this practice could ultimately change academic writing for the better, by bending scholars toward concision from the get-go. No one will mourn the death of jargon-larded flatulence. “But is it good,” she continued, “if I essentially dumb down reading that my professor assigned because he or she thought it was valuable to read?”

Fealty to a professor’s judgment is one question—and students who alter readings are accountable for the consequences. But Zhang also raised a more insidious prospect. Perhaps the most pernicious means of social control in George Orwell’s 1984, she noted, was the pared-down language of Newspeak. “One of the ways the authoritarian regime inserts itself into the lives of everyday people is through the adoption of speech that is incredibly shortened,” she said. “And that’s dangerous, because when we shorten the number of ways that we can express ourselves—or the ways that other things can be expressed to us—we give more power to the people who are shortening these words.

“So then you think, well, who is doing the summary?” she went on. “Who owns the machine that’s doing the summary? And can I 100 percent trust it?” And even if our tech overlords prove less malign than Orwell’s totalitarian superstate, shouldn’t college students be able to “read a complex sentence and do their own summary in their own heads”?

Zhang had been honing such skills in a first-year seminar at Kelly Writers House taught by Julia Bloch Gr’11. AI was only tangential to this class, but the session I sat in on showed how LLMs can be deployed to increase friction for educational purposes. For a homework assignment, Bloch’s students had prompted chatbots to produce a haiku from a news story, iterating and refining the output until they were satisfied. After a lively debate about whether this counted as authorship, an in-class exercise presented them with a series of paired poems or prose excerpts. Each pair comprised one authentic text (Robert Frost, Emily Dickinson, Ernest Heminigway, etc.) alongside a version rewritten by ChatGPT. The challenge was to distinguish the genuine article from the ersatz one, explain your reasoning, and (though time limited this part) express an aesthetic preference between the two.

As a critical-reading exercise, this was as engaging as it was cheap to produce. And I got the sense that some students had been doing something similar on their own. ChatGPT by this time had earned a reputation for certain stylistic tics, like its liberal use of em dashes and the word delve. (The verb is rare in American English but comparatively common among the low-wage African English speakers who have helped train OpenAI’s models.) One of Bloch’s students said that chatbots had awakened him to rhetorical idiosyncrasies of his own, pushing him to expand his linguistic palette. It struck me that even the act of camouflaging an LLM’s output to claim it as one’s own is a sort of exercise in refining prose, albeit a dodgy one.

Later in the semester I interviewed Aliza Jankowsky, a junior majoring in Philosophy, Politics & Economics and minoring in data science. She said that she mainly used LLMs—ChatGPT, Claude, and Perplexity—for “high level overviews and summaries,” but also used them as writing assistants. She had a dim view of using chatbots in the early stages of a draft. “I lose a lot of my voice when I ask it to create it fully from white paper,” she explained. But she does use them to “structure” her ideas, “tell me which ones are more persuasive than others,” and “poke holes in an argument.” Soliciting this sort of “pushback” is “not necessarily timesaving,” she said, but “that is where I’ve also found it to be useful.” Meanwhile, she was content to save time by offloading lower-level tasks. “It’s great at rewording things,” she said. “I send a lot of emails, and I like that I can draft my email and then put it into an LLM and ask it to rewrite it with a specific tone, or a specific voice for a specific person or purpose.” (Though not, so far as I know, on behalf of a University pickleball club.)

The freshmen in Nelson’s seminar seemed to regard LLMs as trustworthy authorities on grammar, syntax, and the “correct” feel of formal prose. Many of them, perhaps all, were in the habit of running any writing assignment through a chatbot for copy editing prior to turning it in. Nelson devoted a great deal of the semester to a sort of compositional counterprogramming. He assigned exercises inspired by the longtime college writing instructor John Warner, a trenchant critic of the “five-paragraph essay” format whose deadening conventions have come to blight US high school instruction (and, by extension, much chatbot output) with formulaic rigidity and “pseudo-academic BS.” By emphasizing writing as thinking, and pushing the students to embrace their own intellectual and rhetorical idiosyncrasies, Nelson encouraged reflection about what distinguishes genuine human expression from the probabilistic derivations LLMs intuit from a given body of training data.

---

EFFORTS AND ILLUSIONS

My broad sense from speaking with undergraduates (to whom I offered anonymity to discuss uses of AI that they would not care to acknowledge publicly) is that outright cheating is not a major problem at Penn. Some of that comes down to the limitations they perceive. “I find that it’s really unable to truly philosophically engage with ideas,” said David Kerendian, a College junior majoring in economics and minoring in philosophy. “In our classes,” said SEAS sophomore computer science major Teyo Agoyo, “we’re not able to just chuck a homework assignment in there. It’s not really able to answer at that high of a level.” Michelle Zhang lamented its abuse by “students who just have no desire to do their own work. But I think that happens less often than people think,” she said. “It’s not great at generating creative work—or work that actually involves people understanding the words that they’re saying.”

A Penn education is a huge investment, and most students seemed genuinely intent on learning. To me, their most compelling uses of AI were often the most mundane: using an LLM to digitize handwritten notes, or convert readings into audio files to listen to on the move, or create practice quizzes from course materials. This echoes my own experience with these tools. I still cannot get an LLM to convert my reporting into a satisfying paragraph—at least not without spending more time prompting than it would take me to compose it on my own. But their ability to transcribe dialogue has supercharged my productivity and also made me a more attentive interviewer.

As for Robert Ghrist’s glimpses of AI intelligence(s) verging on consciousness, I can’t help but suspect that these are substantially a function of his own. Without the deep well of knowledge he possesses in certain domains, he wouldn’t be able to assess LLM outputs—or push back against the mistaken ones to drive the dialectic forward. Inexpert users, by contrast, are liable to misjudge flawed arguments for valid ones and suffer the consequences. “This is the perennial problem,” Ghrist allowed. Every human sooner or later reaches his limit of knowledge. But a chatbot just keeps going—whether its reasoning stays on or off the rails.

For students who found it risky to rely on AI in the early phases of an assignment, or were underwhelmed by their attempts to push LLMs into higher levels of intellectual discourse, the sweet spot often seemed to be somewhere in the middle. “The way that it really helps is when you get to that point when you’re stuck,” Josh Miller explained. Stumped by a homework problem late at night, when a TA can’t be expected to respond to an email and the professor’s office hours are two days away, “I can just take a picture of my work and ask ChatGPT, What am I doing wrong here? And I can move forward without hitting that hiccup.”

Almost every undergraduate I spoke with described something like this. ChatGPT now has a “study mode” that offers “step by step guidance instead of quick answers.” It initially struck me—as it has struck many professors—as a terrific use. But I came to see it as a slippery slope. So had SEAS sophomore Elias Chavez. As a freshman facing a daunting workload, the mechanical engineering major had routinely turned to ChatGPT in this manner for a physics class. “I wish I hadn’t done that,” he told me this year. “It didn’t help me on the midterms, that’s for sure.”

Chavez had plenty of company. Computer science professor Chris Callison-Burch teaches a graduate-level introductory AI course to about 650 students. After sensing that chatbots were leading many to complete assignments “faster and maybe without critically engaging it to the degree that I was seeing from students in the past,” he seeded his exam with several questions that should have been gimmes for anyone who’d done their homework the old-fashioned way. Two-thirds of the class got the first question wrong. Eighty percent flubbed another one. 

“I don’t necessarily attribute that to malice or cheating, or anything like that,” Callison-Burch told me. “I think there’s an illusion of understanding that can happen by virtue of using these chatbots. Even if you are doing it conscientiously, it can be tricky.”

Chavez corrected course. Now he’s more likely to consult his textbook and “sit there for however long it takes me to actually figure it out.” He also discovered the wealth of high-quality tutorials and explanations on YouTube. “I don’t know why I just started with ChatGPT,” he said. His physics class this fall was harder, but he did better in it. But by and large, undergraduates receive little formal guidance, so they are feeling their way forward through trial and error. 

Designing a new introductory AI class for freshmen and sophomores has led Callison-Burch to “rethink a lot of what I’m teaching and how I want to assess the students—and also how I want them to figure out how to engage with the tools that are going to be part of their professional lives.” He’s eager to use AI to his advantage, for instance by experimenting with automated question generators and “second-chance testing.”

“You can write a template that allows you to have many, many variations and combinations that test the same concept,” he explained. “So that allows you to easily reuse [exam] questions in a way that they don’t leak. And therefore it also allows you to let the students try to do things on their own, using that question generator as a form of practice.” This dovetails with Callison-Burch’s teaching philosophy, which he sums up as “A’s for all, as time and interest allow.”

“I want the students to achieve mastery,” he told me, “and I’m happy to award as many people an A as have achieved that mastery.”

Ghrist’s custom GPTs seem like another promising path forward. Since 2023 he has been making them for every course he teaches. Using ChatGPT as a foundation—because virtually all students are familiar with it—he seeds each one with detailed course materials and week-by-week learning objectives. He instructs them to assume that students only know material covered in prior weeks. “That is why just going to generic ChatGPT, or your favorite AI, and asking questions about what you’re working with in class, doesn’t work so well. It draws from everything that’s available everywhere on the internet and starts feeding you information, and then all of a sudden, you’re like, Oh my gosh, what is this talking about? I’ve never heard of this stuff. My professor didn’t say anything about this.”

When students get lost or stuck within Ghrist’s GPTs—or think it may have made a mistake—they can type /think to launch a “dialectical examiner,” which acts like a silicon Socrates. “It starts off really simple, low-level,” Ghrist explained, showing me on a browser. “As you go back and forth, it gets harder and harder until it’s asking really deep, conceptual questions. And I’ve heard back from students that this has been super useful in testing whether they understand something—much more so than just grinding through sample test problems.”

There are additional advantages. CalcStar Blue (like its brethren) invites users to begin by introducing themselves and stating their major. “From then on, it is instructed to try to frame things in a way that is going to click with you,” Ghrist said. “Of course, any good instructor would do this as well,” he added. “But if I get a student who says they’re a finance major with a minor in neuroscience, and that they’re talking to me about Lagrange multipliers and constrained optimization—it’s a little difficult for me to make a connection to that thing that they’re interested in studying. Because I know a lot of applications, but I don’t know everything about finance. I don’t know everything about neuroscience. But LLMs kind of do,” he said. “So these custom GPTs are really good at explaining things with connections to what students are really interested in, no matter how arcane.”

Many undergraduates told me they’d used one of Ghrist’s CalcStar GPTs. Those who hadn’t heard of such tools often cottoned onto the idea of an AI assistant designed and approved by their instructor. “I think professors should be incentivized to create custom GPTs,” said David Kerendian, the economics and philosophy student. “For students to really develop the skills professors want them to learn,” he said, they seem “more refined and more specific to what the professor wants.”

“I could teach anybody how to do it in 10 minutes,” Ghrist told me—though building and refining them takes more time. “My biggest frustration,” he said, “is the models are updating and getting more powerful so quickly, that every time a new model rolls out, I kind of have to retest everything to make sure it’s not going off script.”

“These tools are not meant to replace the human element,” Ghrist emphasized. “Because they can’t.” But he considers them an important complement to more traditional teaching modalities—from office hours and traditional textbooks, to his cartoon-strewn FLCT: Funny Little Calculus Text and sprawling library of YouTube videos—and confessed to feeling frustrated by students who are unwilling to try them out.

“It’s imperative that while students are here focusing on their education, they pick up skills that are going to be broadly useful outside of this place,” he said. “If all they do is interact with AIs and learn that way, they’re missing out on some really useful skills.” But “if they’re not using AI for anything,” he added, “they’re missing out on some really useful skills.”

---

HYPER TEXT

I wrapped up the fall 2025 semester by taking in the final group projects of Rob Nelson’s first-year seminar and engaging the freshmen in a roundtable discussion the following week. It was a curious experience. One group effectively demonstrated how easy it was to create a custom GPT that answered queries in biased and deceptively non-transparent ways. Another led the class through a “vibe coding” exercise that chained a few AI tools together to create personal websites based on students’ resumés or LinkedIn pages. Many of the results were hilariously inaccurate. Though he’d hoped for a better outcome, Luke Castelli offered an unimpeachable takeaway: It’s easy to identify howlers in your own resumé, “but when you look at an essay you prompted, it might not be so apparent.”

In our final discussion, it emerged that ChatGPT, which retains memories of past conversations unless users disable that feature, had apparently concluded that one white student in the seminar was African American. Every month or so, this mystified student related, Chat would respond to a query with the phrase, “As a Black student, you…” Another classmate reflected that he’d started his freshman year assuming that LLM explanations of economic principles were reliable, “and I’m sort of wondering if I shouldn’t.”

But as a whole, the freshmen seemed broadly tolerant of the weaknesses that go hand in hand with these tools’ perceived strengths, which many expect will only grow more powerful. (Had the vibe coding group attempted their exercise after the January 2026 debut of Anthropic’s Claude Cowork, for instance, the outcome may have been much more impressive. The sheer speed of change within the AI marketplace is yet another dynamic to which current undergraduates are adjusting on the fly.) When I repeated my questions from earlier in the semester, the students answered in the same fashion. All but one respondent declined to credit LLMs as possessing “artificial intelligence.” But everyone counted themselves glad to be attending college at a moment when these tools exist. And although one freshman only raised his hand halfway up, it was obvious that everyone expected to be using AI regularly, even daily, over the next four years.

“We encountered a decent amount of different perspectives in this class,” one student reflected. “It’s very easy as a student to not critically think about AI, and just use it to help you with whatever. But this class kind of forces you to think deeper than that.”

Another singled out the usefulness of disentangling AI into discrete categories, as they’d learned from reading AI Snake Oil. “Even just that,” he said, “100 percent helps your judgment.” 

They will need it. Contemporary anxieties about LLMs often echo Socrates’ critique of writing—that they can cloak ignorance with the semblance of wisdom. But the Athenian’s qualms were equally rooted in the inability of text to talk back. Now that it can, these students face the question that confronts us all: what to ask it.

---