The University As We Know It Is Finished

 

Any re-imagining of the university in the age of AI must begin with an honest reckoning with what AI cannot do — and what therefore becomes relatively valuable precisely because AI can do everything else. The key distinction is between work that AI does well (such as synthesis of known patterns, argument elaboration, template instantiation, and generating local coherence) and work it structurally cannot do because of the architecture of the technology as such. AI cannot build the trust on which institutional cooperation depends, because trust is not a conclusion reached by processing information about another agent but instead is a relationship constituted over time between persons who have staked something on each other, and who can be betrayed. AI cannot give a person good taste or style, because taste and style are about personal distinctiveness within a community which shares an aesthetic. AI cannot constitute goals, because that act requires a valuing subject. These are not gaps that more compute will close. They are absences that follow from the ontology of the technology itself.

A curriculum designed around AI’s limitations should be seen neither as an exercise in nostalgia nor as a denial of the burgeoning power of these systems. In fact, given the trajectory of AI capabilities, it is the only curriculum with any hope of finding a stable foundation.

What does this mean in practice? Start with the most obvious casualty: the term paper, as an assessment instrument, is dead. Written homework assignments were meant to push (and test) a student’s ability to produce a well-structured, coherently argued text. But this is exactly what Large Language Models (LLMs) do effortlessly and without demanding of the user any of the underlying cognitive work for which the traditional term paper was supposed to be a proxy. A LLM is an artificial intelligence that has been trained to understand and generate text in a human-like fashion. This included sustained argumentative reason: the ability to construct and maintain a complex argument across an extended piece of discourse, distinguishing claims from evidence, handling counterarguments, and reaching a defensible conclusion. Written assignments also demanded epistemic self-regulation, that is, the metacognitive capacity to monitor one’s own understanding, recognize gaps in evidence, revise positions in response to what the evidence shows rather than what one hoped to find. This pedagogically valuable work always operated below the waterline of the actual output of a term paper; what LLMs do is deliver results that simulate these actions without putting the students through their cognitive paces.

The replacement, as many education researchers are arguing, is live assessment and demonstration: real-time diagnosis of novel situations, design critique, structured adversarial debate, and Socratic examination. These formats test the ability to sense-make under pressure, defend a frame against live challenge, revise a model when evidence contradicts rather than confirms it, and recognize when uncertainty is too high to proceed. In practical terms: collaborative student projects will require documented decision logs tracing reasoning behind commitments, the canonical deliverable shifts from polished artifact to demonstrated live reasoning, and oral examinations and hand-written exams will become the primary assessment instruments. But despite this emerging consensus among education researchers, institutional practice has barely moved.

If the post-AI university’s pedagogic value proposition is the formation of cognitive capacity in conditions that cannot be replicated on a screen, then the function and responsibilities of faculty members must also be reconceived. It clearly no longer makes sense for professors to stand in front of a hall full (or, too often, only half full) of students delivering lectures. As a mechanism of information conveyance, AI can now provide the same at near-zero cost, tailor-made to the specific knowledge gaps of individual students. Instead, professors must reconceive of themselves as interlocutors, serving as performative models of how to calibrate uncertainty and revise frames in real time. The classroom experience should focus on helping students to understand how to constitute a goal rather than generate a text in response to a prompt provided by the professor.

This is something closer to the Oxbridge tutorial system, the clinical ward round, or the seminars of many small liberal arts colleges in the United States. These pedagogies were once defended on grounds of tradition or prestige. The post-AI argument is structural: they are the delivery mechanisms for exactly the cognitive capacities that the architecture of AI cannot replicate, because those capacities are developed only by being exercised, not described. Interestingly, this means that the coming of AI is going to mean there will be demand for more professors, rather than fewer.

None of this implies that faculty should pretend AI does not exist, or that the tutorial and seminar should be conducted in proud ignorance of a tool students will be spending the rest of their professional lives using. The opposite is true. Faculty should integrate LLMs directly and deliberately into their instruction as tools that need to be used correctly in order to not be harmful. (The analogy of a blowtorch or a chainsaw comes to mind: these are useful tools, but you need to learn how to use them safely.) Teaching a student to prompt effectively is teaching them to think precisely about what they want to know and why; it is, in this sense, an exercise in goal constitution. Teaching students to evaluate an LLM’s output critically by scrutinizing the machine’s often over-confident syntheses against evidentiary standards defined by the phenomenological reality of the external and material world is teaching them epistemic provenance tracking and calibrated self-assessment. LLMs can also become objects of critical study in their own right: students should be asked to assess why the model did not produce exactly what they had a priori in mind when they initiated the interaction. Handled this way, LLMs can serve as clarifying instruments in the pursuit of the classical objectives of enlightened education: the inculcation of critical thinking and logical reasoning, rhetorical and communicative competence, aesthetic appreciation and the cultivation of taste, moral and ethical reasoning, and ultimately the ideal of self-knowledge.

This brings us to the content of the curriculum itself. As I recently argued, if the goal of a college curriculum is (as it should be) to inculcate oral reasoning and persuasion, ethical analysis and moral judgment, historical and comparative thinking, and the cultivation of taste and discrimination, then we are precisely in the domain of the classical curriculum of the liberal arts. Skills such as goal constitution, situated judgment, and value alignment are exactly the capacities that a serious engagement with history, philosophy, literature, and political theory develops. History trains temporal imagination and frame revision; philosophy trains epistemic precision and the discipline of distinguishing solid argument from vapid sophistry; literature sharpens an appreciation for style and a feeling for hidden meaning; political theory trains the recognition of suppressed goal contestation and the conditions for legitimate alignment. Together they enable students to imagine lives unlike their own, a hugely valuable experience in a world changing as fast as ours.

How to convey the content of these disciplines to students is going to have to change dramatically from the homogenous one-to-many mass-delivery model of the postwar multiversity, but the content is perfectly classical. The university’s present crisis of purpose is, in this light, at least in part a crisis of having abandoned its own best tradition in pursuit of vocational or technical training that AI is now rendering obsolete.

The central challenge for universities will be how to move toward this model at scale. The tutorial and seminar model is labor-intensive by design: a professor working as interlocutor rather than lecturer can engage only a fraction of the students she could previously reach from a podium. The skills required of faculty will also need to change substantially. Under the old model, a brilliant researcher delivered expected value simply by speaking one-to-many; the new model requires someone with the pedagogic sensitivity to calibrate each student’s specific confusions and capacities—qualities that research prowess neither produces nor rewards. Elite universities in particular have built their faculties almost entirely around research achievement, with teaching treated as a secondary obligation. Reconceiving the professoriate will mean altering tenure criteria and promotion incentives, and it will face fierce resistance from scholars whose professional identities are bound up in the research function. None of this is impossible, but none of it will be easy. No doubt some tenured faculty will pour boulders and boiling oil down the side of their ivory towers to prevent these changes from taking place.

Longer term, however, we should expect the disruption caused by AI to be not just pedagogical but to the structure of the university as such. Kerr’s great insight was that the multiversity’s incoherence was not a bug but a feature—that a loosely bundled institution mirrored a loosely bundled society by providing something for everyone, from the Nobel laureate to the newbie grad student, from the NIH grant-seeker to the remedial English student. What held those disparate functions together was a social infrastructure of knowledge transmission: the laboratory, the lecture hall, the examination, the credential. Once AI can provide information delivery at near zero cost there is no longer a compelling reason why research, teaching, and credentialing need be co-located in the same institution. What will replace the multiversity is likely to be not one thing but several: research centers that focus exclusively on the new-knowledge-production business; independent communal residence facilities that know they are in the coming-of-age business; and teaching systems that are honest about what skills they are inculcating. Even credentials from the most exclusive universities may not retain much social signaling value.

Clark Kerr would have recognized this moment. He was no naïf about the multiversity’s contradictions; but he also believed that competent management could hold them in productive tension. What he did not foresee was that the tension would be dissolved not by political upheaval—as it nearly was in 1964, when the student movement that eventually got him fired also signaled the coming fracture of the postwar liberal-technocratic consensus—but by technological rupture. The irony is that the research university, which Kerr celebrated as the engine of American technopolitical supremacy, incubated the very instrument that is now rendering untenable the research university’s inherited form.

What the students who booed the mention of AI at recent commencement ceremonies this spring were registering, in the way that students have always registered institutional failures, is that they were not getting what they came for. But as with more than one student movement before them, just because they rightly identified a structural problem doesn’t mean that they have particularly good ideas about what a better institution would look like. Just as Kerr recast the University of California to match the liberal-technocratic imperatives of the post- war period, so do visionary college leaders today have an opportunity to remake the university to match the requirements of an economy that will be redefined by AI. Achieving this will be a generational project.

by Nils Gilman at persuasion.community on June 17, 2026