There is a fact the public conversation about artificial intelligence in education has not yet absorbed, and it is worth saying plainly: AI hasn't cheated on university exams. It has made visible that university exams were rigged.
For years, universities — in Spain, in Argentina, almost everywhere — have been evaluating what is easy to measure: the ability to reproduce content, solve closed exercises, write pieces with predictable structure. That is, strictly speaking, operational work. It is the part of any discipline that can be formalized, automated, and now, delegated to a machine. When a law student memorizes the hierarchy of legal sources, when a philosophy student summarizes the Nicomachean Ethics, when a medical student recites DSM diagnostic criteria, they are doing a task a language model today resolves with more speed, more accuracy, and in some cases better style.
The question is not whether students are cheating with AI. The question is why, for decades, we called "exam" something that a machine without a subject can pass. If a cognitive operation can be automated, it wasn't a cognitive operation: it was a procedure. And procedures aren't examined. They're executed.
What AI is doing — without intending it, as a side effect — is laying bare the difference between the operational and judgment. And that difference is exactly what the university forgot to teach.
What judgment is
By judgment I mean something very concrete, not a humanist abstraction. It is the capacity to do four things that no automated system can do today, and probably won't be able to do tomorrow without a change of nature:
Distinguishing evidence from fluency
When a model writes "Foucault holds that the panopticon is the central metaphor of disciplinary modernity," it does so with impeccable syntax. The operational question — whether the sentence is grammatical — the machine resolves better. The question of judgment is different: in which work, on what page, in which translation is that formulation actually grounded? And if it isn't, what is really there in its place? Judgment begins where fluency ends.
Knowing when to abstain
The machine doesn't know how to abstain. If you ask it something it doesn't know, it answers anyway, with the same confidence. It is structurally incapable of saying I don't know. This isn't a technical defect: it is the direct consequence of its function being to produce the next probable token, not to tell the truth. Knowing when to abstain — knowing where one falls short — is the first intellectual virtue. And it is the opposite of what we evaluate today, where the student who writes ten pages gets a better grade than the one who writes two saying I don't have the elements to conclude here.
Distinguishing types of claim
A verifiable claim ("Lacan published Écrits in 1966"), an interpretive claim ("Lacan considers that language precedes the subject"), an argumentative claim ("Lacanian theory can be applied to LLMs"), and a speculative claim ("Lacan would have recognized in LLMs his own discovery") have completely different epistemic statuses. The machine produces all of them in the same register, without marking the boundary. Judgment is, to a large degree, the capacity to mark that boundary.
Signing
To take responsibility, before a claim, for sustaining it in public — before a tribunal, a community, an editor, a student. The machine does not sign because it has no subject. If a text produced by AI contains a serious error, there is no one to hold accountable. The signature is what turns text into act. And the university, when it works well, is the institution that teaches how to sign.
These four capacities — distinguishing evidence from fluency, knowing when to abstain, marking the status of claims, taking responsibility through signature — are not a humanistic complement added to technical knowledge. They are the knowledge. The rest is procedure. And procedures, today, the machine executes better than we do.
Why the university is lost
I have spoken in recent months with professors at several institutions. The description is repetitive. AI has entered their classrooms without asking permission. Students use it for everything: summaries, essays, translations, exercises, take-home exams. Some professors prohibit it, knowing the prohibition is unverifiable. Others integrate it without quite knowing how. Most don't know what to do.
In March I described in these same pages how two tools that appeared in the same week — OpenMAIC, released by Tsinghua, and Wondering, created by an ex-NotebookLM engineer — were turning any topic into a personalized learning experience for under two dollars. Producing a MOOC costs twenty-five thousand euros and a hundred hours. Generating an AI tutor session costs less than two dollars and thirty minutes. The asymmetry is not a technical problem that will be resolved: it is the new baseline. And the university hasn't yet caught on.
The paralysis is understandable, but diagnosing it requires honesty. The university is lost not because AI arrived too fast. It is lost because AI has made visible that much of what we evaluated was reproducible without thinking. If a machine without a subject can pass an exam, what that exam measured was not thought — it was reproduction. And the institutional response cannot be to ban the machine. It has to be to change the exam.
This demands things the university as institution does not want to hear. It demands recognizing that many courses were measuring the wrong thing for decades. It demands redesigning forms of evaluation that are expensive — because examining judgment is expensive: it requires sustained conversation, argued defense, careful reading of a short text instead of skimming a long one. It demands training professors who can distinguish the fluent output of a model from the sustained thinking of a student. And it demands accepting that during a transition period — five years, ten years — degree programs will have to be reformed at the root or be marked as obsolete.
None of this will happen on its own. And it won't happen from inside faculties, where each department defends its status quo. It will happen, if it happens, from a position that today barely exists: that of the articulator between the technical and the humanistic, someone who knows both languages and can translate between them. In most European and Latin American universities, that position has no title, no budget, no department. And it is exactly what the current situation demands.
The false debate over banning or allowing
The public discussion about AI in the university is still trapped in a false dichotomy: ban it or allow it? It is a useless discussion.
Ban
Technically unverifiable and socially regressive: it punishes the clumsy student who copies and pastes, rewards the skilled one who uses it discreetly.
Allow without thinking
Without redesigning evaluation, allowing it is giving up on evaluating. The machine passes the exam — the exam loses its meaning.
The right question is different: what makes sense to keep teaching, how, and how is it evaluated?
My tentative answer, offered for discussion and not as dogma, has three parts.
The operational is taught and evaluated with AI, not against it. If the task is to summarize, translate, calculate, draft a formal letter — the student uses the model, declares it, and something else is evaluated: the capacity to detect errors in what the model produced, to improve it, to judge whether its output serves the purpose. That is already judgment applied to the operational, and it is teachable.
Judgment is taught in a small, dialogical, sustained, and expensive format. There is no cheap way to teach how to abstain, to mark the status of a claim, to defend a reading. It requires what the medieval university had and the mass university lost: sustained conversation between someone who knows and someone who learns, on a concrete text, for as long as is needed.
That conversation can happen today in different ways — face-to-face in a seminar, in asynchronous recorded defenses with targeted feedback, in extensive written exchanges over a text, in structured peer review between students with protocol — but it has to be real conversation: with response, with a question back, with correction. What cannot be substituted is the presence of someone who sustains a position before the student, because the machine does not sustain: it only produces.
The university needs specific tools, built for this task, not generic products imported from Silicon Valley. ChatGPT is not a university tool — it is a mass-market consumer product trained to please the user. A university tool would be a system that abstains by design when it has no archive to support it, that distinguishes types of claim, that separates literal citation from interpretive reading, that makes visible to the student the path by which an answer comes to be one.
That kind of tool does not exist on the market. It has to be built. And building it requires exactly the translation between the technical and the humanistic that we were talking about. That is what Ateneo, the system I have been working on for a year and a half, is attempting to be a proof of concept of. Not the product. A proof that it can be done.
And education for everyone?
Here an immediate political question arises, which it is worth naming before someone else names it against this text: does this mean giving up on university education for everyone?
The honest answer is that the question is badly posed. The mass university was already stratified before AI. Elite universities always offered individualized tutoring; mass public universities offered classes of two hundred and MOOCs. AI did not create that difference, but it does deepen it: today operational education is virtually free — anyone can learn to program, to translate, to write, with a language model and a connection — while education in judgment remains expensive, slow, and artisanal.
The political question is not how to teach judgment to the whole student body with a zero budget. That can't be done. It is to decide how thin a layer of judgment each student receives, and to build intermediate formats that scale without becoming passive: structured peer review between students with protocol, asynchronous recorded defenses with professor feedback, collective dialogical annotation of common texts, short oral exams alternated with long written ones. It is not the same as one-to-one conversation with a master. But it is much closer to judgment than a passive MOOC. It is the front where the battle is being fought, and where it is lost if no one names it.
A personal digression
A computer science professor, a recent National Prize winner1, told me in a recent conversation something worth holding on to: that the trade of programmer will probably shrink but not disappear, because someone will always have to review; that university professors are lost and come to him to know what to do; that the university itself is lost; and that the place he, from computer science, has occupied for decades — the hinge between the technical and the institution — now needs to be occupied from the humanities, but that this place generates little work. Only a few will be needed.
That last sentence, said with the realism of someone who has spent thirty years in universities, is worth taking seriously. This function is not a new mass profession. It is a narrow, demanding position that requires knowing both languages and being able to speak them with authority before very different interlocutors: rectors, deans, engineers, philosophers, programmers, professors. Few people are needed. But they need to be very good.
And that, which seems discouraging, is in fact the most interesting news. Because it means the work is not to build a mass market of "AI consultants in the humanities" — something that, were it to come, would arrive full of opportunism and noise. The work is to occupy a few key posts well, in some specific universities, during the critical period in which institutions reformulate what they will evaluate and how. It is a narrow, temporary window of opportunity. Whoever occupies it now, seriously, will have a disproportionate weight in what the university will be in ten years.
What I would ask, if I could ask
If I could address a rector, a dean, a university governing council, I would ask three concrete things, not abstractions.
That each faculty designate one person — only one, but specific — responsible for thinking through the integration of AI into evaluation and teaching. With time released from teaching, with a clear interlocutor in the rectorate, with access to engineering and to the humanities. To coordinate between the three shores: rectorate, engineering, humanities.
That a stable fund be assigned to building or acquiring specific tools: configurable, auditable AI systems, with abstention capability, trainable on the institution's own corpora. Not subscriptions to mass-market consumer products.
That evaluation in the humanities, social sciences, and law be redesigned, progressively replacing the operational — summaries, commentaries, long essays without defense — with formats that evaluate judgment: careful reading of short texts, argued defenses (oral or written), error detection in AI-produced output, exercises in motivated abstention, recorded evaluative conversations, structured peer review. This is not an objective for the next mandate. It is an objective for this academic year and the next, if the university wants to arrive in time.
These three asks are not new in this text: they cover exactly the three layers that the March diagnostic left open. Source verification in humanistic domains — that is the tool of the second point. Evaluation of complex thinking — that is the redesign of the third. Curricular reformulation that assumes explaining is no longer the bottleneck, thinking is — that is the coordination of the first. The diagnosis was abstract; the asks are concrete. Going from one to the other is exactly the work of translating between the two languages.
Coda: why it isn't only a technical problem
There is a temptation, when speaking about AI in the university, to slide immediately into the language of "digital competence," "technological literacy," "responsible use." It is the language of policy reports. It does not name the problem.
The problem is not technological. It is that the European university — and to some extent the Latin American one — built during the twentieth century an evaluation model based on the presumption that the student wrote alone, against the clock, with their head. That presumption broke. It is not going to be put back together. And the real debate is not how to restore the conditions that sustained it, but what we evaluate now that the student's head has a computational extension that no one is going to take away.
That is a philosophical question before it is a technical one. It is a question about what counts as thought, what counts as authorship, what counts as formation. And that is why the crossing between humanities and the technical is not optional. Without someone who can think this from both shores, the university will make wrong decisions for years, dictated by technology consultancies that don't understand what is at stake or by humanistic departments hoping the problem will pass.
It will not pass.
The letter has arrived. But it has not arrived at the right address. Someone has to read the envelope before declaring it delivered, and that someone cannot be a machine. It has to be an institution that has decided, seriously, what it wants to keep teaching, and why.
That decision is pending. It is, in fact, empty.
And that is why it is the most urgent work the university has ahead of it.
Part II of a diptych. Part I: From MOOC to AI Tutor (March 2026).