The Quieter Number

On May 21, the journal Science published the largest study yet of how undergraduates are actually using generative AI. A team led by Igor Chirikov at Berkeley, René Kizilcec at Cornell, and Ivan Smirnov at the University of Technology Sydney surveyed 95,513 students across twenty public research universities, using an indirect-questioning method designed to coax honest answers about behaviors students would rather not put their name to.¹

The number the press picked up was nine. As in: nine percent of students who use AI have used it to cheat.² Among the students who use it daily, the figure climbs to twenty-six, more than three times the rate among monthly users.²

The headline number is the one that runs on a chyron. The numbers that almost nobody printed are the quieter ones underneath.

Who Is Not in the Picture

About a third of the students surveyed reported using generative AI regularly when completing assignments.³ That fact, on its own, is unremarkable. The disaggregation is where the story is.

Forty-five percent of male students reported regular use. Thirty-three percent of female students did.² Students from underrepresented racial minority groups reported lower rates of regular use than white and Asian peers.² Low-income students used the tools less than higher-income classmates.² The discipline gap was equally stark: sixty-two percent of computer science students used AI regularly, fifty-three percent in mathematics, fifty-one percent in business, with the humanities trailing.³

The students least represented in the chart are not abstaining out of principle. They are abstaining because of access, because of cost, because nobody in their department has yet said, in so many words, that learning to use the tool is part of the work. Chirikov and his coauthors are blunt about the implication. These students "may fall behind in college and eventually in the workplace because of unequal access to or practice using AI."²

The cheating number reads as a discipline problem. The disparity number reads as a curriculum problem. Both are real. Only one of them is being legislated.

What Detection Cannot Solve

The instinct, in a year of headlines about academic dishonesty, has been to reach for detection. Catch the work that crossed the line. Sanction the student. Move on. The Science paper closes that door politely. The authors note that "banning GenAI is unlikely to be effective in preventing AI-assisted cheating and may disadvantage students" who will be expected to know how to use the tool the moment they enter a workplace.⁴ Their recommendation is harder. Reform the assessment, discipline by discipline. Not a universal policy. The kind of recommendation that takes a department a year.

What a Future Worth Building Looks Like

The future the Science paper is gesturing at is not a future where AI is everywhere or nowhere in higher education. It is a future where the work students do is shaped, on purpose, so that AI is a co-worker rather than a substitute. Where the final paper is one artifact among many, not the only piece of evidence anyone sees. Where the process of getting to the paper, the questions a student asked, the drafts she discarded, the moment she changed her mind, is treated as part of the work rather than a private prelude to it.

The benefit of that shift is not only about catching the nine percent. It is about the larger group on the other side of the disparity, the students who are not using these tools yet, and whose absence from the chart is its own kind of warning. If process becomes the unit of learning, the playing field tilts back toward effort. A student who does not have a paid subscription but who keeps thinking, revising, asking, and noticing has something to show for it. A student who outsources the work has something missing in the same record.

This is the part a system like Koan is quietly trying to be useful for. Not as a detector. Not as a gate. As a way of keeping the trail. The drafts a student kept. The questions she asked Aidan when she was stuck. The pauses that mean something and the pauses that do not. None of that, by itself, is teaching. It is the kind of evidence the next generation of assessment will need, and which most classrooms have never had the time to collect.

The Sentence Underneath

The reason the disparity finding matters more than the cheating finding is that it is a slower problem. Cheating is a story that resolves in a semester. A widening gap in who learns to use the tools that the next decade of work will assume, that resolves over a generation. The paper is, in the careful language of academic writing, pointing at the second clock.

The institutions that take the paper seriously will not start with a new academic integrity policy. They will start with a question that is harder to answer in a meeting. What does the work in this department look like, if we redesign it so that the students who are not using these tools yet are not, in the act of doing the work, falling behind the ones who are?

If nine percent of users is the number we keep saying out loud, what would change if we started saying the other one too?