The Real University Crisis Is Not AI. It Is a Failure to Build Agency

What if the most important thing a university teaches is the one thing an AI tutor cannot do: make a person responsible for what happens next?

That question changes the argument about higher education. The familiar debate asks whether universities are too political, too expensive, too bureaucratic, or too vulnerable to artificial intelligence. Those concerns are real, but they are symptoms of a deeper confusion about the product universities provide.

For centuries, education was organized around scarce knowledge. Professors possessed it, libraries stored it, and institutions distributed it. Now a student can ask an AI system to explain calculus, generate a study plan, critique an essay, write code, monitor a software project, and remind them when a task needs attention. In technical work, an agent can even inspect a repository and repeatedly check whether a change passes its tests.

When knowledge and routine execution become abundant, the value of education moves elsewhere. It lies in judgment, agency, trust, and the practiced ability to act with other people under conditions of uncertainty.

The university that understands this shift will become more valuable. The university that continues selling access to information, credentials, and administrative machinery will become harder to justify.

From knowledge transfer to agency formation

A personal AI tutor can explain almost anything at the speed and level a learner requests. That is not a minor improvement to education. It attacks the historical reason for the classroom.

But explanation was never the whole educational experience. A person may understand a principle and still fail to use it. They may know the facts of an argument and still be unable to defend a position in public. They may produce an elegant plan and still lack the courage, discipline, or social skill to carry it through.

This is the difference between information acquisition and agency formation.

Information acquisition asks: Can you retrieve the answer?

Agency formation asks:

• Can you decide which question matters?
• Can you distinguish a plausible answer from a trustworthy one?
• Can you act when nobody has provided a complete script?
• Can you absorb criticism without treating it as an attack?
• Can you coordinate with people who do not share your assumptions?
• Can you take responsibility for consequences that an automated system cannot own?

A residential university can be valuable because it places a person inside a dense network of obligations, encounters, deadlines, friendships, conflicts, experiments, and opportunities. Its transformative power is not primarily that a professor transmits a fact. It is that a young adult repeatedly encounters situations in which knowledge must become conduct.

Consider a student who joins a bipartisan political organization. The valuable lesson is not simply learning where two parties stand on immigration or public spending. It is discovering that the person across the table is neither a caricature nor an abstract threat. The student must listen, formulate a question, tolerate discomfort, and remain in the room after a conversation becomes difficult.

That is a kind of learning an AI can simulate but cannot fully substitute for. A system can generate the strongest argument for an opposing view. It cannot make the student feel the social and moral weight of disagreeing with an actual person, nor can it create the mutual trust that makes future cooperation possible.

The scarce resource in an AI rich world will not be access to answers. It will be the capacity to choose, act, and remain accountable when answers conflict.

This also explains why civil discourse cannot be taught as a one time lecture. Conversation is a muscle. It develops through repetition, resistance, and recovery. The same is true of initiative. A student learns agency by making decisions, seeing them fail, revising them, and trying again while other people are watching.

The surprising connection: education and software agents are both systems of feedback

The technical details of modern AI tools reveal something important about human development. A scheduled software agent can repeatedly check a code project, observe whether the tests pass, and determine what to do next. The remarkable feature is not merely that it can perform a task. It can operate inside a feedback loop.

A feedback loop has four parts:

1. A goal.
2. An action.
3. Evidence about the result.
4. A revised action.

Much of education fails because it supplies goals and information without enough meaningful loops. Students listen, submit an assignment, receive a grade, and move on. The system may measure performance, but it does not necessarily develop the learner's ability to notice reality, interpret feedback, and alter behavior.

The best parts of university life are loop rich. A laboratory experiment produces a surprising result. A debate exposes a hidden assumption. A team project reveals that a technically correct plan cannot survive poor coordination. A failed internship application forces a student to reconsider how they communicate their abilities. Each episode connects action to consequence.

AI can increase the number and speed of these loops, but only if people use it as a coach, simulator, critic, or tool for iteration. Used poorly, it removes the loop. A student asks for an essay, receives polished prose, and submits it without encountering the friction that would have revealed their confusion. A programmer accepts generated code without understanding its failure modes. An employee asks an agent to monitor a task but never defines what success actually means.

The same technology can either deepen agency or conceal its absence.

This gives us a useful test for any educational practice: Does it make the learner more capable of running their own feedback loops?

A strong assignment does not merely ask for a correct answer. It asks the student to state a hypothesis, test it, explain what changed their mind, and identify the next experiment. A strong seminar does not merely reward agreement with the instructor. It makes students revise their views in response to serious objections. A strong advising system does not simply tell students which course to take. It helps them understand the tradeoffs and make a decision they can defend.

The technical world is building machines that can loop. Education must build humans who can set the direction of those loops.

Trust is the hidden infrastructure of learning

This is why institutional neutrality and public trust are not cosmetic concerns. They determine whether people believe a university is helping them think or trying to recruit them into an approved identity.

A university can hold values. It should care about truth, human dignity, intellectual honesty, and democratic self government. But there is a crucial distinction between values that govern inquiry and political positions that govern belonging.

When students believe that certain conclusions are socially safer than others, the feedback loop breaks. They stop exposing uncertainty. They ask what answer will be rewarded rather than what evidence supports. Faculty may become less willing to challenge fashionable assumptions. Viewpoint diversity shrinks, not necessarily because anyone has formally banned dissent, but because the cost of dissent becomes too high.

This is an educational failure before it is a political one. Learning requires the ability to encounter information that threatens one's existing map of the world. A community that cannot distinguish disagreement from hostility deprives its members of the very friction that improves judgment.

Institutional neutrality, properly understood, does not mean moral emptiness. It means that the university does not use its authority to settle questions that its members should be free to investigate. It means conservatives, progressives, religious believers, skeptics, and people who do not fit neatly into any category can enter the conversation without first performing ideological compliance.

The practical consequence is significant. If AI makes argument generation cheap, then the credibility of the setting in which arguments are tested becomes more valuable. Anyone can ask a system to produce an argument for a position. Fewer institutions can create an environment where that argument is challenged by intelligent people with different commitments, where evidence is examined publicly, and where participants must continue working together afterward.

Trust is therefore not a public relations asset added after the academic mission. It is the operating system of the academic mission.

The university must accept more responsibility because students have less margin for error

The agency thesis also changes the economics of higher education.

A young person deciding whether to borrow tens of thousands of dollars is not making an ordinary consumer choice. They are often making it before their interests, abilities, and long term plans have fully developed. The institution has better data, greater expertise, and a direct financial stake in enrollment. It is unreasonable to place nearly all of the risk on the least informed participant.

If a program routinely leaves graduates with heavy debt and weak employment prospects, the university should bear some consequence. That might include income based repayment commitments, partial loan underwriting, transparent outcome reporting, or limits on expansion in programs whose costs cannot be justified by their results.

This is not an argument that education should be reduced to salary. Human beings need history, art, philosophy, scientific understanding, and civic formation even when those subjects do not produce the highest immediate income. It is an argument against pretending that financial consequences are irrelevant to the student who must pay the bill.

The principle is simple: the party best positioned to understand and influence an outcome should share responsibility for its risk.

The same principle applies to technology. If an institution requires students to use AI, it should teach them how to verify its output, protect private information, recognize fabricated citations, and remain accountable for submitted work. If it bans AI without redesigning assessment, it merely creates a game of concealment. If it embraces AI as a shortcut, it may graduate students who can produce artifacts but cannot explain, test, or defend them.

Responsibility cannot be outsourced to either debt contracts or software.

This is also where access and merit become more complicated than their usual slogans suggest. Standardized measures can sometimes reveal talent that wealthier applicants have been able to hide behind polished recommendations, expensive activities, and carefully managed experiences. At the same time, a test score alone cannot measure every form of potential. The serious question is not whether one metric is pure. No metric is. The question is whether the overall system expands opportunity while making its criteria visible and contestable.

A trustworthy institution must explain how it identifies talent, how it supports students after admission, and what it knows about the outcomes of different paths. Endowments and philanthropy can make that possible by reducing debt, funding research, and supporting infrastructure. But the existence of resources creates a higher obligation to demonstrate public value, not a lower one.

A new design brief for higher education

The emerging model should not be “university versus AI.” It should be AI for knowledge, institutions for formation, and people for judgment.

That model implies several changes.

First, universities should move basic knowledge transfer toward flexible, AI assisted formats. Students should be able to receive explanations at their own pace, practice privately, and obtain immediate feedback. Classroom time should be reserved for what benefits from human presence: debate, experiments, collaboration, coaching, performance, and difficult decisions.

Second, every student should practice disagreement as a form of civic and professional competence. This means structured dialogues, adversarial collaboration, public reasoning, and assignments that require a fair account of an opposing view before criticism begins. The goal is not artificial balance. It is the ability to understand why an intelligent person might reject one's own conclusion.

Third, institutions should measure agency rather than merely completion. Useful indicators might include whether students can define a problem, make a justified choice under uncertainty, respond to feedback, coordinate a group, and explain the limits of an automated tool. These are harder to grade than factual recall, but they are closer to what employers, communities, and democratic societies actually need.

Fourth, universities should reduce internal friction without confusing simplification with abandonment. Administrative systems often grow because institutions are responding to genuine obligations, including mental health care, accessibility, safety, compliance, and complex advising. The goal is not to return to a romantic past in which those needs were ignored. It is to eliminate duplicated approvals and slow processes so that resources reach students and researchers faster.

Finally, institutions should make their economic contract legible. Publish program level completion rates, debt burdens, earnings ranges, and transfer pathways. Offer alternatives when a four year degree is not the best fit. A society that respects education should not need to pretend that every learner requires the same institution.

Key Takeaways

• Use AI to increase feedback, not eliminate effort. Ask it to critique, simulate objections, generate practice problems, or identify weaknesses. Do not let it perform the part of the work that develops your judgment.
• Practice disagreement deliberately. Once a week, explain the strongest version of a position you reject. Then identify the evidence that could change your mind.
• Choose educational paths by risk adjusted value. Compare total cost, likely debt, completion rates, career outcomes, and the specific experiences the institution provides.
• Demand visible institutional responsibility. Universities should disclose outcomes, share downside risk, and explain how admissions and financial aid decisions are made.
• Build agency through short loops. Set a goal, act, inspect the result, and revise. This pattern applies to studying, coding, career planning, and civic participation.

The central mistake in the AI era is to define education by the activities machines are beginning to perform. If education means delivering explanations, then AI will indeed replace much of it. If education means producing credentials, employers and learners will increasingly ask whether cheaper signals can do the same job.

But if education means forming people who can choose worthwhile goals, challenge their own assumptions, work with disagreement, learn from consequences, and accept responsibility for action, then the need for serious institutions becomes clearer, not weaker.

The university's future will not be secured by defending every old lecture, office, or procedure. Nor will it be secured by adding AI to the existing system and calling that innovation. It will be secured by becoming the place where human beings learn to direct powerful tools without becoming tools themselves.

The question is no longer whether a machine can teach us the answer. The question is whether we are building people capable of deciding what deserves to be asked, tested, and done.