95,000 Students, One Study, and the Number That Should Change How You Talk About "AI Cheating"
For six months, "AI cheating" has been talked about as if it were a single problem with a single fix: ban it, detect it, or let it through. A study published in Science this past May makes that framing very hard to hold onto.
The research team, led by Cornell's René Kizilcec with colleagues from UC Berkeley and the University of Technology Sydney, surveyed more than 95,000 students across 20 public research universities during the 2023-24 academic year. It is the largest look yet at how students actually use generative AI on their coursework.
The Top Numbers
About a third of students regularly use generative AI on assignments. Nine percent admitted to using it to cheat. Taken alone, those numbers support a year of general anxiety about AI and academic integrity.
Discipline Breakdown Changes the Picture
Regular AI use runs highest in computer science (62%), mathematics (53%), and business (51%). Cheating runs highest in economics (17%) and journalism (16%), and lowest in biology (5%). These are not the same list of subjects. A discipline with heavy AI use is not automatically the discipline with the most cheating. Something other than subject matter is doing the predicting.
The Number that Matters Most
Frequency of use seems to matter the most. Daily users admit to cheating with AI 26% of the time. Drop to monthly use, and the figure falls to 7%. That's nearly a fourfold gap. A student who reaches for AI occasionally, for a specific and bounded purpose, behaves very differently than a student who has folded it into daily habit without much reflection on where the line sits.
What This Means For a Blanket Policy
A campus-wide ban treats a biology lab report and an economics problem set as the same kind of task. The data says they are not. The same goes for a single detection tool applied everywhere: it assumes cheating shows up the same way in every field, when the numbers say cheating tracks discipline and habit, not access to the tool alone.
The study's authors land in the same place. Sometimes the right move is returning to tightly controlled, in-person assessment, when that actually serves the learning goal. Sometimes it's simpler: write clearer guidelines, since ambiguity, not access, was the real problem. And sometimes the fix lives in the assignment itself, redesigned so AI use becomes visible or beside the point. Every one of those suggestions is a judgment call made by someone who knows the specific assignment. None of them is a rule stamped across an entire institution.
The CALM AI Case Supported by Data
This is the exact premise behind the CALM AI Framework: not one answer for every assignment, but a repeatable way to ask, task by task, what belongs to AI, what belongs to the teacher, and what belongs to the student's own thinking. A dataset of 95,000 students now backs up what careful educators have sensed without having the numbers to prove it. There was never going to be one clean verdict on AI and cheating. There were always going to be as many verdicts as there are assignments.
For a structured way to work through that question one assignment at a time, download The CALM AI Toolkit; it’s is free,
If you are looking for a free community built for exactly that kind of assignment redesign, join the AI Faculty Break Room, It’s also free.
Source URL (attribution)
Also see:
AI Disclosure Statement
A note on how this gets made: I use AI to help with research and early drafts, the same kind of grunt work that used to eat my Sunday afternoons. The ideas and judgment calls are mine. I read every piece before it goes out, and if a line doesn't sound like something I'd actually say to you, it gets cut. That's the whole point of the CALM AI Framework: AI carries the volume. I keep the thinking.