How universities are affected by AI moving knowledge outside their walls

A student today can ask an artificial intelligence (AI) model anything without embarrassment: Whether it’s a request for an explanation of how compound interest works or of Pareto efficiency, each question gets a patient, direct answer at any time and at whatever level of sophistication the asker needs.

For most of the university’s 900-year history, the core repository of teachable knowledge has been inside the walls of the institution—in its faculty, libraries, laboratories, and classrooms. But now, the center of accessible explanation has moved outside the university’s walls to AI models. And the implications are larger than the current debate about AI in the classroom suggests, rather, the shift affects the underlying value proposition of higher education itself.

For much of its history, the university has been organized around the production and management of scarce knowledge. The medieval university held the books, trained the clergy and lawyers, and certified mastery of a fixed body of authoritative texts. The industrial-era mass university expanded access dramatically but kept the same architecture: Authoritative knowledge lived inside the institution, delivered through standardized curriculum, fixed time-to-degree, and a credential at the end. The internet later weakened this monopoly but did not break it. Raw information became abundant, but interpretation, explanation, and feedback still required faculty.

The mass university could be compared, in its essential design, to a modern industrial factory. Students are the raw material, faculty and curriculum are the production process, and the credentialed graduate is the output. Instruction is standardized so the line can run at scale, time-to-degree fixes the throughput, and the diploma certifies that the product met specification. This is not a hostile description—the industrial-age model democratized higher education and did enormous good. But a factory is built to turn standardized inputs into standardized outputs efficiently, and standardization is precisely the kind of work intelligent automation performs exceptionally well.

What has changed in the past few years is not simply access to information but the cognitive layer itself. AI does not merely provide information. It explains, customizes, gives feedback, and synthesizes the type of work that justified the lecture hall and office hours. This is in line with what Benjamin Bloom identified in 1984 as the “two-sigma problem”: one-on-one tutoring produced learning outcomes two standard deviations better than classroom instruction but tutoring at scale was economically impossible. For 40 years the problem was unsolved, blocked not by pedagogy but by economics. AI has now removed the economic barrier: Every student can have a patient, responsive tutor on demand. Whether these systems reliably reproduce Bloom’s full two-sigma gain remains an open empirical question. But it is clear the constraint that made one-on-one explanation a luxury has collapsed.

This does not mean AI is correct, complete, or unbiased. In fact, it is none of those things consistently. But universities no longer have a near-monopoly on knowledge production and dissemination. The student no longer needs to come to campus to receive an explanation of the invisible hand, the second law of thermodynamics, or the Federalist Papers. The explanation comes to each of them, on demand, in whatever form and at whatever time they request. The boundary that defined the institution for nine centuries has, in a single technological generation, become quite porous if not disappearing in total.

To see why this matters, it helps to be precise about what the modern university bundles together. Three products are sold under a single tuition price.

The first is content delivery, including lectures, readings, and problem sets, that transmit established knowledge from those who hold it to those who do not. This is a student’s daily experience, and it has been the primary justification for the university’s cost and time since the rise of the industrial mass university in the late 19th century.

The second is credentialing—the degree as a signal to the labor market that its holder has completed a recognized program at a recognized institution. The credential is what employers actually purchase when they require a bachelor’s or master’s degree, and it is what makes the university economically essential for most students even when the content could be acquired elsewhere.

The third is networking and formation. Networking is the slower, harder-to-measure work of meeting people, establishing connections, joining an intellectual community, and entering jobs, communities, and professional circles that follow from them. Formation is the cultivation of judgment, ethical sensibility, and the capacity to think seriously under uncertainty. Both happen through sustained relationships with faculty and fellow students, through encounters with hard cases, through participation in traditions of inquiry, and through the experience of being taken seriously as a thinking person by people who have lived a life of inquiry themselves.

For roughly a century this bundle held together. Content delivery was a visible product. Credentialing followed automatically from completing the content. Networking and formation came along almost as a byproduct of the time students spent on campus with faculty and peers.

AI breaks this bundle in multiple ways. Content delivery is the factory’s production line, and it is the most substitutable component. A capable AI tutor, integrated with course materials, can increasingly deliver explanation and feedback at a quality that exceeds the average large-lecture experience, meaning the part of the university that most resembles a factory is the part most exposed to automation. Credentialing remains sticky but is under pressure. Employers across technology, finance, and government have announced moves toward skills-based hiring, though research suggests practice still lags the proclamations. Networking and formation are the hardest for AI to replicate, because they are fundamentally relational and embodied in personal networks. They require human judgment encountering other human judgment over time, in conditions of mutual stakes.

It is difficult to predict what the university will become, but the direction of change is already visible, and it is running along several distinct tracks at once.

The first track is unbundling and speed. Western Governors University has built a competency-based model in which students advance by demonstrating mastery rather than by seat time. Southern New Hampshire University has also developed competency-based, direct-assessment models that prioritize demonstrated competence over time spent in courses. Google’s career certificates sit just outside the university system but show the same pressure: short, job-oriented credentials designed to connect learners with employers without requiring a full degree. Arizona State’s partnership with Starbucks offers eligible employees 100% upfront tuition coverage for an online bachelor’s degree through the university. Abroad, South Korean universities have developed company-linked contract departments with firms such as Samsung and SK Hynix, tying curriculum, scholarships, and employment pathways more tightly to corporate workforce needs. This track trades away the residential experience and most of what came with it, but it protects access for working adults, career-changers, and students who could never afford four years of foregone income. A fast and legible credential is thus not a lesser choice but the only workable one.

The second track is AI-embedded instruction. Arizona State again offers a leading example: Its institution-wide partnership with OpenAI is framed around using ChatGPT Enterprise to enhance teaching, learning, and discovery across the university. Georgia Tech‘s online master’s in computer science—low-cost, high-volume, and fully online—showed a decade ago that elite instruction could scale; AI can now be layered onto that earlier model of scaled digital instruction. Here, the flipped classroom becomes the natural architecture; content is absorbed at home through an AI tutor, and class time is reserved for discussion, critique, and application. This track forgoes the traditional lecture and much of the routine explanation faculty once provided but protects the classroom itself, repurposed for the work that still requires a room full of people.

The third track is a quieter shift in assessment. If AI can produce a polished paper, a working proof, or clean code, the finished artifact stops being what a university can meaningfully grade on its own. Some programs have responded by lowering the stakes attached to any single piece of work. In one emerging model, students choose assignments from a pool larger than any one student is expected to complete. Most of these earn credit and detailed faculty feedback but no evaluative grade; only a few are graded outright. Students decide how much evaluative weight to take on, and faculty spend their time responding to student work rather than scoring it. This is early and uneven, but the direction is significant: less weight on any single output an AI could have produced, more attention on the ongoing exchange between student and instructor.

The fourth track runs in the opposite direction from the first: a smaller set of institutions doubling down on intensive human formation. St. John’s College conducts its curriculum largely through small seminars, tutorials, and laboratories centered on primary texts rather than conventional lectures. The service academies treat the formation of judgment and character as their explicit institutional purpose. Medical residencies and law school clinics have always worked this way. What this track trades away is scale and cost-efficiency; what it protects is the one product in the bundle that AI cannot manufacture.

These tracks are not mutually exclusive, and most institutions will likely combine elements of several. But they are not equivalent either. The first two compete on cost, speed, and delivery—the terrain where AI’s advantage is largest and still growing. That is a wager that the credential can outrun the substitution. It may hold for some institutions but asks a great deal of luck for others.

This is why the fourth track deserves more attention than its current scale suggests. The small seminars, capstones, and clinical placements do more than transmit knowledge—they build sustained relationships through which both networking and formation happen. A student who spends years in the same seminar or a rotation in the same clinic forms working relationships with faculty and peers that outlast the degree itself. AI can supply the relevant knowledge, but it cannot decide what to do with it in a particular case, under uncertainty, with real stakes. It cannot be held responsible for the choice. It cannot reliably tell a serious argument from a merely fluent one. And it cannot make a person a participant in a living tradition of practice—a physician, an engineer, a public servant, or a scholar—rather than a consumer of its outputs. Judgment under uncertainty, responsibility for outcomes, discernment, and membership in a community of inquiry are what formation produces, and networking follows naturally from the same relationships, once formation has given students something worth connecting over.

This is less a leap forward than a return. The university did not begin as a factory but rather a guild. Arguably the first universities—Bologna in the late 11th century and Paris soon after—were a universitas, or a sworn corporation of masters and students organized to pursue a craft. Their characteristic exercise was not the lecture alone but the disputatio: the formal disputation in which a student took a position and defended it under challenge. To learn was to be initiated into a community of argument, conducted face to face, under masters, judged by the standards of the craft. The industrial-era university preserved apprenticeship at the research level but expanded undergraduate education through standardized curricula and mass credentialing—a real democratic achievement that nonetheless traded the guild’s formative core for the factory’s throughput. AI is now dissolving that factory layer and thus exposing the older logic beneath: education as apprenticeship into a community of practice.

None of this requires abandoning AI. AI can take over much of the explanation and feedback that currently consumes faculty time, freeing faculty for the work that only humans can do. A faculty member whose AI handles initial drafts of problem-set feedback has more time for the seminar discussion, the one-on-one meeting, and the careful reading of a difficult capstone. AI used this way enables formation by removing the routine work that has always crowded it out.

There is no avoiding the question of cost. Formation-centered education is more expensive per student, not less. It requires smaller classes, more faculty time per student, deeper community infrastructure, and stronger ties to external practitioners. The change will be painful, and the pain will be uneven: Selective universities with strong endowments will find this path far easier than tuition-dependent regional ones, which is itself a problem that deserves serious attention from policymakers and accreditors.

The first instinct of many institutions will be to bolt AI onto the existing assembly line. Courses will become more efficient, advising more automated, grading more scalable, and content delivery more personalized. Some of this will be useful, and much of it may be necessary, but it will not answer the deeper question AI has placed before higher education.

The industrial university was built for a world in which explanation was scarce, expertise was localized, and access to knowledge required institutional mediation, but that world is disappearing.

The question, then, is not how universities can preserve the old model with better tools. It is what remains worth preserving when the old model no longer has the same economic logic. The answer is not content delivery. Nor is it credentialing alone, though credentials will remain important for some time. The answer may be formation: the slow work of bringing students into communities where judgment is practiced, arguments are tested, ethical responsibility is learned, and intellectual seriousness is modeled by other human beings.

AI will do what factories do best: standardize, scale, and deliver. The university should not try to out-factory the machine. Its reason for being is older and harder. It is to form people capable of knowing what to do with knowledge once knowledge is everywhere. If knowledge has moved outside the walls, the university’s task is not to pull it back in. It is to form the people who know how to use it well.