Artificial Intelligence at the University: Widening the Gap Between Teaching and Learning

What's really happening with artificial intelligence at universities looks less like an earthquake than an X-ray. Language models produce, at almost zero cost, exactly what the institution had been rewarding for decades: fluent, plausible academic prose, with no guarantee of any understanding behind it. That's why they make visible a crisis between teaching and learning that was already there. The public debate is split between those who announce that AI "broke" the university, those who organize the resistance, and those selling salvation. All three camps argue about the tool. They almost never argue about the relationship.

The machine as interpreter

I'll start with a figure that usually gets glossed over. The British survey by HEPI and Kortext, published in March 2026, was carried out in December 2025 among 1,054 full-time undergraduates. According to it, 95% use AI in some form and 94% have used it for assessed work, up from 53% in 2024. The headlines stopped there: almost everyone is "cheating." But the most common use for that work isn't pasting in generated text: according to figure 4 of the report, only 12% admit to doing that, compared with 3% in 2024. The most common, at 61%, is asking the machine to explain concepts; in the report's own words, almost two-thirds of students use AI to have concepts explained to them.

Language models produce, at almost zero cost, exactly what the institution had been rewarding for decades: fluent, plausible academic prose, with no guarantee of any understanding behind it.

The figure deserves a careful reading, because it's self-reported, from a single country, and based on an undergraduate sample. Even so, it says something no moral panic has wanted to hear. A majority of students use AI as an interpreter for what they didn't understand in class, in the readings, or in the assignment prompt. The companies' own reports, which are interested sources and have to be read as such, point in the same direction. Anthropic's first education report analyzed university students' conversations with Claude in April 2025. It found that 39.3% were devoted to creating or improving educational content, and that a large share of the rest consisted of asking for technical explanations.

In 1965, Pierre Bourdieu, Jean-Claude Passeron, and Monique de Saint Martin published Rapport pédagogique et communication, translated into English as Academic Discourse: Linguistic Misunderstanding and Professorial Power. Using vocabulary tests with French university students, they measured how much of the language they were being taught in the students actually understood. Comprehension turned out to be highly imperfect and tied to social background, while the faculty took it for granted; the students, for their part, quickly learned to cultivate an illusion of understanding based on familiarity with the words. The book's conclusion is uncomfortable: the pedagogical relationship rests on a linguistic misunderstanding and on the shared fiction that this misunderstanding doesn't exist. The professor pretends to be understood and the student pretends to understand. Each illusion serves as the other's alibi, and whoever has mastered the register by family inheritance starts with an advantage.

Sixty years later, that misunderstanding has found an industrial middleman. Calling it cheating doesn't cut it. What's happening is more like the outsourcing of a translation the classroom never fully did. The interesting question is no longer whether the machine explains well—sometimes it does, and sometimes it makes things up—but why an interpreter was needed in the first place.

A crisis with a birth date

The genealogy is older than Bourdieu, and it has a Córdoba accent. The Liminar Manifesto of the University Reform, dated June 21, 1918, and written by Deodoro Roca, described a university regime founded on "the divine right of the university professoriate." It set against it another idea of authority, the kind that in a home of students is exercised not by commanding "but by suggesting and loving: by teaching." And it left behind a line that every quality-assurance committee should have hanging on the wall: "If there is no spiritual bond between the one who teaches and the one who learns, all teaching is hostile and consequently barren." It wasn't talking about technology. It was talking about a broken relationship, and about how hierarchy couldn't fix it.

Half a century later, in Deschooling Society (1971), Ivan Illich wrote that school teaches the pupil to confuse teaching with learning, grade advancement with education, a diploma with competence, and fluency with the ability to say something new. Read in 2026, that last pair is almost a technical description of a language model. An LLM is fluency, period: it produces the plausible surface of a text that has understood something. If an institution confused fluency with ability, a cheap fluency machine leaves it without an alibi.

The confusion between fluency and learning has also been measured. In 2019, a Harvard physics team led by Louis Deslauriers compared, with randomly assigned students, lectures by experienced, highly rated instructors against active-learning sessions on the same material (PNAS, 2019). Those who worked actively learned more, but felt they had learned less than those who sat through the lecture. The authors note that cognitive effort was experienced as a sign of learning worse, and that a brilliant lecturer can generate such a feeling of learning that students end up preferring them. Long before chatbots, fluency was already manufacturing the feeling of learning.

This reading isn't new for me, and the chronology matters. In July 2020, in the middle of the pandemic and two years before ChatGPT, I co-signed a manifesto with Edgar Iglesias, José Luis Lalueza, and Moisès Esteban-Guitart in the Revista Internacional de Educación para la Justicia Social. Right from its title, it called for a critical, intergenerational, sustainable, and community-based education, and it aimed at reforming what couldn't be put off any longer. The pandemic revealed it for the first time, and AI is doing it again. The crisis doesn't have a commercial launch date.

Three trenches, one shared assumption

The rupture narrative says AI broke the essay, assessment, even thinking itself. The resistance narrative answers with bans, detectors, open letters, and a return to handwritten exams. The hype narrative promises a personal tutor for every student and institutional licenses by the truckload. The three look like enemies, but they share an assumption: that before AI, assignments and exams certified learning. If that were true, there would be something to break, something to defend, and something to improve. If it wasn't, all three are arguing over an original that never fully existed.

The three look like enemies, but they share an assumption: that before AI, assignments and exams certified learning.

Seymour Papert already had a name for this mistake in 1987, in "Computer Criticism vs. Technocentric Thinking" (Educational Researcher, 16(1)): technocentric thinking. It means asking what the computer does to learning as if the computer were an agent rather than one piece within a culture. Each trench also has its pet paper. The alarmists cite the MIT Media Lab preprint Your Brain on ChatGPT, with 54 participants and not peer-reviewed when it went viral, as if it were an autopsy of thought. The enthusiasts cite meta-analyses showing positive effects on performance, often built on short studies that measure the quality of the output while the tool is still available.

The best-designed studies tell a different story. In a field experiment with high school students, Bastani and her team (PNAS, 2025) compared GPT-4 without safeguards against a version designed as a tutor, which gives hints instead of answers. With the unguarded version, performance improved during practice and got worse once the tool was taken away. With the tutor version, that harm was largely mitigated. A randomized experiment with 117 university students, by Fan, Gašević, and colleagues in the British Journal of Educational Technology (2025), found that those who wrote with ChatGPT turned in better texts without learning more. The authors gave it a name that's already making the rounds: "metacognitive laziness." Here at 421, Tomás Milgron walked through that literature from the angle of cognitive sovereignty. I want to underline what those studies say about the university more than about the brain. What's decisive isn't the tool but the design of the relationship between whoever teaches, whoever learns, and what gets assessed. In other words, exactly what the trenches don't discuss.

The omission is easier to see from the vantage point of advertising. Before the university, I worked in interactive communications, as a creative director but also as a copywriter and a planner, and back then strategic planning held a central place in the communication process. It wasn't enough to define an audience and polish a message: you had to understand what was going on in those people's lives and at what media and technological moment you'd be talking to them. Seen that way, the rhetoric of rupture is instantly recognizable as an advertising genre rather than a diagnosis. It is, in fact, planning done well: it spots a zeitgeist anxiety, the unease about a technology almost nobody fully understands, and hands it back as a story, with a before and an after and, at the center, something to adopt or reject. The industry doing that is just craft. What's odd is that much of the university debate bought the story without asking the question that was never missing in planning. The tool and the message have been discussed ad nauseam; what's happening in the lives of the people who study, and the media and technological moment in which those lives unfold, a lot less so.

The closed loop and the detector that rewards the canon

If you look at the whole system, something stranger than fraud shows up: a closed loop. Students write with AI. Faculty design materials with AI and, sometimes, grade with it. Anthropic's report on how university instructors use Claude analyzed some 74,000 conversations from May and June 2025, from Free and Pro accounts with higher-education email addresses, and supplemented it with a survey of faculty at Northeastern University. Assessing student performance made up 7% of the total, and in those conversations the task was handed off to the machine in an automated way 48.9% of the time. That happened even though the surveyed faculty themselves considered grading the area where AI was least effective.

Research isn't off the hook either. An estimate covering four major AI conferences held after ChatGPT's arrival concludes that between 6.5% and 16.9% of the text in their peer reviews may have been substantially modified by a language model. An analysis of more than fifteen million biomedical abstracts, published in Science Advances, calculates that at least 13.5% of those from 2024 went through one. AI automates a relationship that was already socially and bureaucratically automated: rubrics nobody reads twice, boilerplate comments, research judged by publication count. None of those habits was born in 2022.

The most widespread defensive response, detectors, adds an irony that deserves more attention than it gets. In a study published in Patterns, Liang and colleagues (2023) ran 91 TOEFL essays, written by non-native English speakers, through seven popular detectors. On average, 61.22% were labeled "AI-generated." Essays by American eighth graders were classified almost without error. The mechanism is simple: detectors punish predictable prose, with common vocabulary and little variation. They penalize those who write outside the dominant register and acquit those who have mastered the canon, or who know how to ask the machine to sound less predictable.

Policing through linguistic suspicion reproduces the hierarchies of class, background, and native language that the university claims to want to correct.

Ángel Rama called it the «lettered city»: the apparatus through which Latin American elites turned writing into a border of power. The detector is its statistical version. Policing by way of linguistic suspicion reproduces the very hierarchies of class, origin and mother tongue that the university claims it wants to correct. For those who write in Spanish, a language less represented than English in the data used to train the models and the tools that police them, this is no technical footnote. It's a problem of academic citizenship.

Symbolic capital: who loses value when fluency gets cheap

There's a layer the public debate would rather not touch, and I'll frame it in structural terms, not personal ones. In Homo Academicus (1984), Bourdieu distinguished two kinds of capital within the university. Academic capital is power over the reproduction of the professoriate: hiring panels, committees, positions. Intellectual or scientific capital is the prestige earned through one's work. He also described two strategies: conservation, typical of those already consecrated, and subversion, typical of newcomers. And he spoke of hysteresis: dispositions formed in an earlier state of the field that keep operating after the field has already changed.

Read this way, part of the «resistance» can be understood as a conservation strategy. Those who built their authority under the regime of print culture, through mastery of the written register, citable erudition and the essay as proof of thought, are watching a machine cheapen precisely that fluency. Defending the essay as it was is also defending the value of a form of capital. More recent arrivals, more connected to algorithmic digital culture, tend to have the opposite problem. Their capital in the critical appropriation of technologies, experience design or platform literacy converts poorly into recognized academic capital.

The board repeats itself, with variations, across much of the world's university systems. Faculties are aging, and the top tier concentrates both age and decision-making power: in Spain, for example, more than a third of full professors at public universities are sixty or older. At the base, a layer of precarious teaching staff, temporary or part-time, keeps growing; it carries much of the teaching load without any say in the decisions that shape it. And the systems that evaluate academic careers reward publication far above teaching, so reinventing the pedagogical relationship earns little on the scorecards that decide who gets promoted. I have no solid data cross-referencing attitudes toward AI with age, rank or type of contract. What I'm proposing, then, is a theoretical interpretation, not an empirical finding.

Three caveats keep this from turning into a generational war. First: the difference isn't about age. The «digital natives» myth was debunked years ago, for instance by Kirschner and De Bruyckere in Teaching and Teacher Education (2017). What matters is position in the field and the kind of relationship one has with digital culture, not date of birth. Second: the ambivalence runs both ways. There are consecrated figures who reconvert their capital in no time and rebrand themselves as «AI experts». There are newcomers captured by the attention economy or by the very platforms they claim to critique. The third includes me. I write from inside the field, from an in-between position. For more than a decade I've coordinated a course on digital identity that is now a microcredential, and I know how hard it is to get that work to count on the same scorecards as an indexed journal article. There is no outside to look in from.

Those who built their authority under the regime of print culture, through mastery of the written register, citable erudition and the essay as proof of thought, are watching a machine cheapen precisely that fluency.

I teach at an in-person university and advise at a distance-learning one, and what I experience at each has little in common. At the distance institution the alarms went off sooner, and the hunt for possible student cheating started sooner. But it also moved sooner toward a different stance, one where, instead of policing, people were encouraged to think more about process than product. I suspect the explanation has little to do with technology and a lot to do with alibis. In-person teaching keeps one that distance learning never had: the sense that if there's a classroom, there's learning. A university that can't offer the classroom as proof depends heavily on what its students turn in, and a machine that churns out plausible submissions hits it first and dead center. That same exposure also forced it to ask sooner what counts as evidence that someone actually learned. Maybe that's why, wherever the alibi still stands, people can still argue about the machine instead of about the relationship.

The fine print of sovereignty

That leaves political economy, which usually arrives wrapped in the language of opportunity. Big Tech offers educational licenses in contexts of austerity, and efficiency gets sold as pedagogy. European regulation moves at a different pace. The AI Act classifies as high-risk the systems used in education to decide admissions, assess outcomes or proctor exams. But Regulation (EU) 2026/1744, known as the Digital Omnibus on AI and in force since July 27, 2026, pushed those obligations back from August 2, 2026 to December 2, 2027. Meanwhile, decisions are being made in licensing contracts that almost nobody in the classroom has read. The enshittification Cory Doctorow describes for the internet finds fertile ground here. A university that doesn't know what pedagogical relationship it wants to protect will hardly know which clause to negotiate. As Tomás Balmaceda and Tobias J. Schleider warn in their analysis of philowashing, debating how a system should behave doesn't answer the prior question of which decisions we should hand over to it.

That's why the question of cognitive sovereignty, which Juan Ruocco framed in 421 as the capacity of a person, a community or a nation to control its own knowledge, attention and decisions, isn't settled inside each individual head. It's also at stake in institutions that decide what counts as knowledge, what counts as cheating and who gets the last word on both. The Black Mountain College experiment reminds us that there were universities that tried out a different relationship between those who teach and those who learn. The arguments between doomers and hypers remind us that the question about AI is almost always a question about us.

There's one final irony I don't want to waste. This piece will be translated into English with an AI, and a person will review it afterward. If the translation comes out fluent, nobody will know whether the text was understood along the way. It's the same thing that happens, every week, with thousands of assignments uploaded to a grading platform. The difference is that here we're saying so.

More than a century ago, the students of Córdoba demanded a bond between the one who teaches and the one who learns. The machine didn't break it: it found it already made of polite misunderstandings and turned it into a service. What remains to be seen is what we'll do when we can no longer pretend we understood each other.