The Supervised Interval
A term paper never proved that anyone knew anything: it proved they had sat with something hard long enough to be changed by it. That was the product.
There is an argument going around that the college is about to collapse, and the case is stronger than the people making it usually manage to say. A model can now explain a known concept more clearly than most lecturers, write an essay that reads as competent, and pass the greater part of the exams we use to certify that a young person has learned something. If those three acts were the university, the university is finished, and the only open question is how long the buildings stay warm.
Those three acts were never the university. They were its packaging.
I have spent thirty years inside systems sold, each in its turn, as the end of some older way of doing things, and the pattern holds every time: the thing announced as obsolete turns out to have been carrying a second cargo no one had named, and it is the second cargo we lose. The lecture was never only the transfer of a known concept: a book did that more cheaply a century ago. The essay was never only the proof that the concept had landed. The essay was the interval: the days a mind spent inside a difficult question, writing the wrong version and seeing that it was wrong, reaching for a claim and putting it back, until what survived the discarding was a thought the student could not have held before beginning. The discarding was the education. The paper was only its receipt.
AI does not threaten the receipt. It forges it perfectly, which is a different problem, and a smaller one than the one underneath.
Run the distinction I rely on for everything else in this field, because it holds here too: separate what the system does from what the system is. What AI does to education is specific, measurable, already here: it removes the friction between a question and an acceptable answer. That is a does-claim, and we can test it, bound it, respond to it. “The university is obsolete” is an is-claim, and it cannot be tested, only preached, and it is preached now with the same two-faced certainty that attends every other AI story, the prophets of collapse and the prophets of the personalized-learning golden age selling, between them, one product from opposite ends.
While that argument runs at full volume, the other thing proceeds underneath it, unhurried, the way the consequential things usually do. The interval is being closed from a second side, and this one has nothing to do with the technology’s cleverness and everything to do with who holds the instruments. A campus is, among the things it is, a place where power has always wanted to decide what may be said, and the wanting used to run into a friction of its own: syllabi are many, a teacher’s judgment is hard to audit at scale, and what a young person is permitted to think has been, mercifully, expensive to police.
AI is, before it is anything else, the removal of that expense.
The tools arrive wearing the friendly face of assistance, and each of them, turned a quarter-degree, becomes an instrument for deciding the permitted answer. The detector that promises to catch the machine’s writing has been shown, again and again, to flag the honest non-native student’s prose as fraudulent at rates that never touched the native speaker’s: a filter that does not know it is a filter, sorting by a bias no one signed. The curriculum assistant that drafts the reading list drafts the one it was weighted to draft. Aim the sentiment tool at the campus forum to catch “concerning” language, and you have hung a microphone that no longer needs a listener at the far end. None of this requires a conspiracy; it requires the opposite. It requires convenience, adopted a department at a time by people with no agenda more sinister than a full inbox and a mandate to be seen doing something about AI. The agenda, whatever its color this decade, does not need to install itself. It only needs to supply the default.
So the interval is pressed from both sides at once. The model lifts the struggle out of the student’s hands, handing over the finished answer before the reaching could change them, while the same class of tool narrows the range of answers the struggle was ever allowed to reach, quietly, by making the approved one the path of least resistance. A mind handed the answer and a mind permitted only one answer arrive at the same place, which is nowhere: neither has done the descending, and the descending was the education.
I find, and it surprised me to arrive here, that I am not much moved by whether the institution survives in its present shape. Forms die; the cargo is the thing. Let the university fragment into its research institutes and its credentialing counters and its clubs for the young to find one another, and if somewhere in the wreckage a smaller thing survives whose whole business is teaching a person to think against the grade, the essential cargo will have found a new container, and the buildings were never sacred. It is the interval I would not lose, because it is the one thing both forces are reaching for and the one thing neither can be trusted to keep: the collapse treats it as friction to remove, the capture as a range to narrow, and to both of them the young mind learning to refuse the easy answer is not the point of the exercise but the obstacle to it.
Here is the test I would hand a teacher, or a dean, or a parent, in place of a verdict about “AI in the classroom,” a phrase that has already stopped meaning anything. Of any tool proposed for the campus, ask not whether it is efficient, and not whether it is inevitable, but whether it restores the interval or removes it. Does it put the struggle back into the student’s hands, or take it out of them. Does it widen the range of answers a young person may reach, or supply the one they are meant to arrive at. A tool that makes a student sit longer with a hard question and a tool that hands over the sanctioned conclusion do opposite work, and the phrase “AI in education” is built precisely so that one word can cover both before anyone thinks to separate them.
The knowledge was always going to be free, and we should be glad of it. What was never free, what had to be paid for in the one currency that does not deflate, in time and difficulty and the small daily defeat of the wrong sentence written before the right one, was the interval in which a person became someone who could think. That is what is closing now from both ends: a machine that removes the difficulty, a power that removes the choice, quietly, without announcement, on a campus full of the young, who are in no position to miss what they were never handed. Which leaves the noticing to the rest of us. It was always going to be us.
Sources
L. Giray, J. Roe, and J. Diesta Espiritu, “AI writing detectors are ineffective, unreliable and harmful,” English Teaching: Practice & Critique, 2026 (doi:10.1108/ETPC-07-2025-0155). High false-positive rates, the disproportionate flagging of multilingual students, and the paradox that fear of a false accusation pushes students toward the very AI use it means to police.
Weixin Liang et al., “GPT detectors are biased against non-native English writers,” Patterns (Cell Press), 2023 (arXiv:2304.02819). The first measure of the bias, a 61% false-positive rate on Chinese students’ TOEFL essays against 5% for US students, reproduced repeatedly since.