The EDU Zombie Cure

There is an enormous amount of public concern today about the role of Artificial Intelligence in education, at all levels — rightly so. What are we going to do about it? I ran across two recent essays from students that clang the alarm bells in very thoughtful, detailed, personal ways. They are on the front lines today.

I’ve spoken with both techno optimists and AI alarmists of many stripes — with alarmists inside and far outside of tech. Given my personal experience on the inside of an ethical AI company who has seen great AND awful results, I end up in the middle looking for practical solutions, recognizing that AI is: a) unstoppable, b) amazing and potentially very beneficial, and c) an existential fast-moving, scary technology with many ways to cause harm to people despite human-inventor origins. While I was once working in the center of Education technology, I am no longer an EdTech insider — so am writing this as a distant EdTech observer, who was really triggered by these two recent student essays.

On the one hand it is tempting, when reading these essays from Theo Baker and Owen Yingling on the current state of the AI-era university, to retreat into familiar techno-optimism: that every technological shift has triggered moral panic, that education has never been especially perfect, but that the kids will be alright. Tempting, and wrong. Baker, writing as a Stanford senior, and Yingling, writing as a University of Chicago senior, are not pearl-clutching outsiders. They are participant-observers describing what they have watched happen to their classmates, their classrooms, and themselves. Their accounts converge on the same uncomfortable picture: that a generation of students is voluntarily handing over not only their problem sets but their emails, their gym routines, their texts to romantic partners, their student newspapers, and — most consequentially — the cognitive labor that constitutes learning itself.

Baker calls it addiction. Yingling calls it cancer, then zombification. Whatever metaphor one prefers, both writers are pointing at something real: the gap between what universities say they are doing about AI and what is actually happening in their classrooms has become, in Yingling’s phrase, “an impassable chasm.” Princeton’s cheating cases nearly doubled in a single year while the administration encouraged faculty to “experiment” with generative tools. Stanford finally reinstated proctored blue-book exams in April 2026, a century after banning them on principle. Forty-nine percent of Stanford computer science majors recently told surveyors they would rather cheat than fail.

Anyone who wants to offer constructive advice has to begin by conceding that the critics are substantively correct about the diagnosis. What follows is an attempt to take that diagnosis seriously — not by promising that AI integration will magically solve what it has helped cause, but by proposing a framework and a set of concrete responsibilities for each actor in the system.

Three Things They Are Right About — and One They May Underestimate

First, the official response from elite universities has so far been a predictable error. Press releases about “AI literacy initiatives,” fifty-million-dollar gifts to “support pedagogical innovation,” and faculty symposiums on “teaching in the age of generative AI” have done almost nothing to address the actual behavior of students sitting in lecture halls with Claude or ChatGPT open in the next tab. The rhetoric of “integration” has functioned, in practice, as permission to disengage from the harder question of what students must still do for themselves.

Second, the harm is not primarily about cheating in the narrow sense. Yingling’s most important observation is that AI use spreads. It begins as a shortcut for a dry problem set and ends as a substitute for emailing a professor, composing a message to a romantic interest, summarizing a book read for pleasure. The risk to students is not that they will be caught and punished. The risk is that they will succeed. I have succeeded in my own work in this way, developing research and content better than I could on my own — you can judge this piece, which was informed by AI, whether I succeed or not.

The big risk is that many students will graduate having outsourced the formation of judgment, voice, and taste, and will not be able to retrieve those capacities later because they were never built, when the expensive learning door was especially open.

Third, this is happening fastest precisely where it should be happening least: at the institutions that claim to spare no expense on undergraduate education and that, by their own self-description, exist to cultivate intellect rather than to deliver job training.

What the critics may underestimate is harder to say without sounding like an apologist. Frontier AI is not going away. A graduate who has never thought carefully about how to use these systems — when to defer to them, when to override them, how to verify their outputs, how to preserve one’s own judgment in their presence — is not better prepared for the world. A purely prohibitionist response would produce its own kind of zombification: graduates who are helpless before tools their colleagues use fluently.

Yingling is right that “integration” has been used as a slogan to avoid thinking. But an opposite slogan — that AI must be kept out of the classroom entirely — also avoids reality. The serious work is in the middle, and it requires drawing a distinction that is difficult to execute.

Formation Versus Performance

The distinction that does the most work here is between formation and performance.

A great deal of the confusion in current AI-in-education debates could dissolve once this distinction is made explicit. The problem with a student “Chatting” every essay in a humanities core class is not that AI is bad. It is that the student has confused a formation activity for a performance activity. The student is being asked to write the essay in order to become someone who can think — not in order to produce a document the professor needs. AI use in that context does not improve performance; it cancels formation.

This framework gives every actor in the system a clearer set of obligations. Formation activities must be ringfenced against AI substitution — not by encouragement or neglect, but by design. Performance activities can and should incorporate AI, with explicit instruction in how to do it well.

What Each Party Must Do

Universities

The structural change universities need is not another center, another task force, or another press release. It is a redesign of assessment. Every university that continues to grade students primarily on artifacts a model can produce in thirty seconds is, whatever its public posture, avoiding reality. The choice is not between “trusting students” and “policing them.” It is between assessment instruments that measure formation and assessment instruments that measure willingness to copy and paste.

Concretely, universities should commit to a clear principle: in any course that purports to develop independent judgment and capacity, key assessments must occur under conditions that exclude AI assistance. The form can vary — proctored written exams, oral examinations, in-class essays, public defenses of work, supervised lab practicals, traditional viva voces. The specific instrument matters less than the principle that no student should graduate from a serious program without ever having demonstrated, unassisted, that the formation actually occurred. Stanford’s quiet return to blue books is not a retreat to 1925; it is an honest assessment approach that students must face.

Universities should also stop pretending that grade inflation and AI cheating are unrelated problems. When the modal grade in a humanities course is an A-minus, the marginal cost of using AI to secure that grade is negligible and the marginal benefit of doing the work yourself is invisible.

Finally, universities should be honest with applicants and current students about what is changing. Prospective students considering a computer science degree in 2026 deserve to know that the entry-level job market they imagined when they applied no longer exists in the same form. This is not an argument against studying computer science; it is an argument for telling the truth.

Professors

The most important thing professors can do is the thing that is hardest to scale: be present in the classroom in a way that an LLM cannot be. Yingling’s invocation of Goethe — “a teacher who can arouse a feeling for one single good action, for one single good poem, accomplishes more than he who fills our memory with rows on rows of natural objects” — is not nostalgia. It is a description of the durable competitive advantage human teaching should always have.

Professors should design assignments where the artifact is not the point — where the work is in the process, and the process is witnessed: reading discussions where students are called on by name and expected to defend a position they have actually formed; drafts shared in class and revised in front of others; short, frequent in-class writing exercises; oral exams that no model can take for the student.

A clearer approach to AI policy is to specify, assignment by assignment, what tools are permitted and why. “This problem set is a formation exercise; AI use is academic misconduct.” “This final paper is a performance exercise; you may use AI for research and editing, but you must disclose how, and the argument must be defensibly yours in an oral followup.” The transparency is itself instructive.

Professors who use AI to write their own lectures or grade their students’ work should expect, and accept, that students will respond in kind. If teaching is a relationship between humans, both sides need to show up and be honest.

Students

Treat your undergraduate years as the last extended period in your life when you will be paid, in effect, to build cognitive capacities you may not easily build later. The firms that will hire you may not teach you how to digest a difficult text, construct an argument that convinces someone who disagrees, or sit with a problem long enough to discover how to solve it. They will assume you arrived already able to do those things. School is a precious window of opportunity to open yourself, to challenge yourself, and to figure out directionally where your mission lies.

This does not mean refusing to use AI. It means being honest with yourself about when you are using it as a tool and when you are using it as a substitute. A useful test: after you have produced something with AI assistance, can you defend every claim in it without the tool? Can you reconstruct the reasoning? If the answer is no, you have not produced work; you have produced a costume.

Finally, choose your courses, when you can, for friction. The professor with the strange enthusiasms, the small seminar at an inconvenient time, the course that requires oral examinations or hand-written work — these are now the rare goods. Treat them as such. I started out taking an abstract Mathematics course from a European professor I couldn’t understand, then took a punch-card computer course, tried to learn German in a German-immersion class, ended up as an English major, then Teacher, Diplomat, Tech leader and all sorts of odds and ends. Those challenging courses helped shape the best of me, and helped me find my way.

Frontier AI Companies

The companies building these systems — Anthropic, OpenAI, Google DeepMind, Microsoft, and their peers — bear a real responsibility. They have published research and rhetoric about the educational potential of their products. They should be held to it.

Concretely, frontier labs should ship and maintain genuine study modes — not marketing reskins, but interaction modes designed by people who understand pedagogy and cognitive science, in which the model refuses to produce finished work on request and instead works as a Socratic interlocutor: asking questions, surfacing the student’s own reasoning, pointing at gaps, declining to summarize a text the student has not read. These modes should be the default in education-licensed deployments.

Labs should also be honest in their educational marketing about the difference between formation and performance, and about the cases in which their products are unambiguously bad for users — including the cases involving emotional offloading and parasocial attachment that Baker describes.

And labs should fund — through unrestricted grants, not through products with their names on the buildings — independent research that can actually answer questions about how their products affect cognition over time. The early studies suggesting that reliance on AI for cognitive tasks reduces independent capacity are alarming, underpowered, and politically inconvenient for the companies whose products they study. That is exactly why the companies should pay for more and better research, with no editorial control over the findings.

EdTech Companies

Much of what passes for EdTech in the AI era is a wrapper around a frontier model with a school-friendly logo. Genuine educational technology in this moment would help institutions enforce the formation/performance distinction and measure the results: secure proctoring tools that actually work and respect student privacy; authorship-attestation systems that verify whether a piece of writing was produced under specified conditions; authentic-assessment platforms that make oral examinations and process-based grading practical at scale.

The EdTech companies that build these things will be less glamorous than the ones promising to “revolutionize learning.” They will also be the ones that matter.

Standards Organizations & Accreditors

Accreditation bodies have so far been almost entirely absent from this conversation. They should not be. An accreditation regime that does not ask whether a degree-granting institution can demonstrate that its graduates have actually developed the capacities the degree claims to certify is an accreditation regime that no longer works.

Would it be reasonable to require, as a condition of accreditation, that institutions document at least one substantial AI-excluded assessment in each major degree program and publish aggregate results? It need not be micromanagement — but there must be minimum conditions under which a degree continues to mean anything.

Standards organizations in specific professional fields — bar associations, medical licensing boards, engineering societies — should consider whether their licensing examinations are currently measuring what they were designed to measure, and whether credentials should require demonstrated competence in working with AI as well as without it.

Philanthropies

Philanthropic capital is paradoxically both the most flexible resource in this ecosystem and perhaps the worst-deployed. The fifty-million-dollar gifts to “AI initiatives” are, with a few exceptions, doing nothing the universities would not do anyway. Are the donors just buying naming rights to a transition that is happening regardless?

The highest-leverage philanthropic uses of capital: endowing chairs and fellowships for the kind of eccentric professors Yingling celebrates — the ones who cannot be replaced by a model because they were never replaceable by a syllabus; funding research on cognitive effects of sustained AI use that may produce inconvenient findings; underwriting genuinely independent assessment infrastructure that universities can adopt without paying per-seat fees to a vendor; supporting small liberal arts colleges to maintain their formation-intensive models when dominant institutions cannot.

The Alternative Is Too High a Price

The class entering college this coming fall will be among the first to have used frontier AI through most of high school. By the time they graduate, the models they use will be substantially more capable than the ones their seniors used. If the response of higher education over the next four years looks like the response over the last four — symposiums, statements, an “impassable chasm” between official rhetoric and ground truth — Yingling’s zombification will not be a metaphor. The EDU industry must move apace with AI, which is yes, very difficult and unprecedented.

The response does not have to look like that. It requires conceding what the critics have correctly identified: that something serious will be lost, that the losses will cascade, and that the slogans of “integration” have so far functioned as a refusal to think. It requires a clear distinction between the activities that build minds and the activities that use them. It requires assessment instruments that measure the first kind of activity honestly. It requires teachers willing to be present in ways no model can replicate, and students willing to choose friction when the easier path is one tab away.

Baker, near the end of his essay, watches a classmate work through “Applying the Gale-Shapley Algorithm to The Princess of Clèves” and observes that love was something “to be optimized.” It is a weird image of what is at stake. The work of a university is not to optimize like this. It is to form the kind of person who can recognize, later, when “optimization” is the wrong question. That work is still possible. It is harder than it used to be. It will require more honesty than the current institutional responses may have accomplished. Am I naïve to think that all this adaptation and correction can happen and keep up? The alternative — a generation of credentialed graduates who never quite became themselves — is too high a price to pay by pretending the problem will solve itself.