AI in Higher Education Just Got a 3-Year Report Card. 20 Students Wrote It.

Roughly half of the seniors in a 3-year study at Babson College told researchers they worry that AI is eroding their own voice, particularly in the creative and emotional writing they care most about. As first-year students in spring 2023, when ChatGPT was new on campus, the same students worried mostly about cheating. The shortest answer to what 3 years of AI in higher education did to this cohort is that it moved their fear from getting caught to losing something no transcript records.

Kristi Girdharry and Stephen McElroy of Babson College followed the same 20 business students from that first semester all the way through senior year, and the Chronicle of Higher Education's Daily Briefing carried the findings on August 3. The sample is 20 students, and the researchers claim no national sample. It is also the rarest kind of evidence anyone has produced on this subject, because it is the same people answering over 3 years, at an age when 3 years remakes a person. Nearly everything else we have about students and AI is a snapshot taken by someone who never went back. Babson went back.

I spend a fair amount of time in university rooms, and the question that follows me into the hallway has changed in exactly this direction. In 2023 it was some version of "am I allowed to use this?" Now a senior waits for the room to clear and asks something closer to "if I use it the way everyone expects me to, what will be left that sounds like me?" The first question has an answer in a policy document. The second one does not.

What the same 20 students stopped worrying about

In year one, the dominant worry was academic integrity, plain and simple: who actually did the work. It showed up in assignments, grades, honor codes, and it asked a degree program for exactly two things, detection and rules. By senior year the dominant worry was personal voice, whose voice actually comes through, and it showed up hardest in creative and emotional writing. That worry doesn't ask a degree program for detection. It asks for judgment, and a reason to keep writing in the first place.

Most institutional AI policy in the United States was drafted against the first worry. That was the right work at the time, and somebody had to write those rules fast. The students then spent 3 years discovering a problem the rules were never designed to touch. You cannot detect your way out of a voice that thins by consent, one convenient draft at a time.

This is what I mean by the four-year bet. A degree is a wager placed on a student's behalf about which capabilities will still be worth something on the other side, and the institution places that bet with someone else's 4 years and someone else's tuition. Every catalog is a forecast. The Babson cohort is that forecast aging in public: the same 20 people watched what they feared, and what they valued in their own work, move underneath them inside a single degree. The interval between when a course is designed and when its assumptions expire has a name, curriculum half-life, and it is now shorter than the degree it sits inside.

Campus is not the only place where the rules were written for last year's problem. The same lag runs straight through corporate AI, where agent governance is arriving well behind agent deployment. Institutions of every kind are governing a thing that moves faster than their own drafting cycle.

The oldest complaint about a new tool is 2,400 years old

In Plato's Phaedrus, Socrates worries that writing will ruin memory, that students given letters will have "the show of wisdom without the reality." He was wrong about what writing was worth and right about what it cost. Something in the mind did atrophy, and the civilization that followed decided the trade was worth making, with full knowledge of the price.

The Babson seniors are making that trade with better information than Socrates had, and doing the accounting themselves, at 21, unasked. The students most fluent with these tools are the ones raising the most serious question about them, and they are raising it about their own interior life rather than their grades.

Detection protects the assignment. Only practice protects the voice.

What a graduating class carries into the market

Those 20 students walk into a labor market where the entry rungs are being redrawn in real time, and where companies are restructuring around AI without always saying so out loud. The capability they are worried about losing happens to be the exact capability that survives all of this. Judgment and a recognizable point of view are what sit above the algorithm, and they are the last things a hiring manager can rent from a vendor. A student who can tell when the model's draft is competent and hollow is worth more at 22 than one who can produce 10 times the volume of it.

So the practical work in front of every provost and department chair is deciding which assignments exist to be produced and which exist to build a person, then protecting the second category on purpose. My own vantage point comes from the California State University system, where I advise on AI governance, and from Tulane, where I am Innovator-in-Residence. The campuses moving fastest are the ones willing to reopen the catalog.

Two questions I am carrying into my own meetings this month, and you are welcome to both. Which of our requirements exist because a future employer needs them, and which exist because we have always required them? And when a student tells us they are worried about losing their voice, who on our campus is responsible for answering that?

If you are not a provost, the questions still belong to you. If you are paying tuition, ask the admissions office what the program taught 3 years ago that it no longer teaches, and listen for whether anyone can answer it. If you are the student, keep writing one thing by hand, on purpose, badly. That habit will outlast the tool.

20 students at one college in Massachusetts told us what 3 years cost them and what they want back. Every institution making a 4-year bet this fall gets to decide whether that counts as data. Mine says it does. And if you see it differently, I'm easy to find.

Sources: Chronicle of Higher Education, Daily Briefing, August 3, 2026, reporting longitudinal research by Kristi Girdharry and Stephen McElroy of Babson College tracking 20 business students from spring 2023 through senior year in 2026.

Here is what makes Alex a credible voice on this topic: Alex advises the California State University system on AI governance and serves as Tulane University's Innovator-in-Residence, standing that puts him in the room as campuses rewrite policy built for a fear their students have already outgrown.

For more, book a conversation →