Everyone agrees it is happening. Below is a handy formula to limit unauthorized AI use in the classroom, for use by faculty, administrators, and legislators alike. Some of the key variables are controlled by the institution, some by the faculty. Everyone needs to work together.
I use the term “unauthorized,” so “cheating” is whatever the faculty member declares to be unauthorized use, understanding that in the AI era, students are fluent in AI and will seek the most efficient way to fulfill course assignments. (Most of the misconduct data is pre-LLM and is not directly applicable to today’s technology but I am assuming the same patterns hold.1 )
The hundreds of conversations I have had in the past year with faculty members across the country make clear that there is almost no agreement about what AI misconduct means. Some faculty members hold that any LLM use at all is cheating; others say “go ahead and use Claude or ChatGPT, but if I can prompt your paper into existence in under 10 minutes, you fail.”2
My proposed rule is flexible by focusing on AI use that is “unauthorized” by the faculty member. The institutional role is ensuring that faculty can enforce their own line.
The formula
P = 1 − (S × L × M × C × A × κ)
P is the probability that a student will use AI in a way the instructor did not authorize. P runs from 0 to 1. Zero means it won’t happen. One means it will.
The six variables are the conditions of teaching. Each runs from 0 to 1. Multiply them. Subtract from 1. That’s P.
S — class size
L — teaching load
M — modality
C — one-to-one contact
A — assessment design
κ — institutional culture
A 1 means the condition is fully present. A 0.1 means it’s gone.
How to score
S — class size. 1 if the class is fifteen or fewer. 0.75 if sixteen to twenty-five. 0.5 if twenty-six to fifty. 0.25 if fifty-one to a hundred. 0.1 if more than a hundred. The class-size literature on cheating is thirty years old and surprisingly consistent.3
L — teaching load. 1 if an instructor teaches one or two courses per term. 0.75 if three. 0.5 if four. 0.25 if five. 0.1 if six or more. The load decides how many students an instructor can know. A 5:5 load with mid-sized sections is 750 papers.4
M — modality. 1 if the class meets in person. 0.75 if hybrid with required attendance. 0.5 if online synchronous. 0.25 if asynchronous with proctoring. 0.1 if asynchronous without proctoring. The systematic review evidence on online exam cheating shows the rate jumping from roughly 30 percent to 55 percent during the pandemic, which is a measurement of structural pre-concession.
C — one-to-one contact. 1 if the instructor has three or more scheduled meetings with each student over the course of the term. 0.75 if two. 0.5 if one. 0.1 if zero. A faculty member who never sits across from a student has no baseline for saying “this is not your work.” The contextual-factors finding replicated since the 1990s, that faculty engagement predicts integrity better than any honor code or detection tool, is the same finding stated in the negative.5
A — assessment design. 1 if the instructor uses an oral defense or in-class component. 0.75 if staged drafts with feedback. 0.5 if take-home with required process artifacts. 0.25 if take-home final only. 0.1 if a single uploaded file graded by rubric. A single upload is the assignment most easily faked and the assignment most institutions are now using.
κ — institutional culture. 1 if the institution adjudicates reports within the term and backs faculty judgment. 0.5 if there’s a written policy that gets enforced sometimes. 0.1 if reports are routinely declined or reversed. κ is the institutional term and it multiplies everything else, because a faculty member who tries to hold the line on authorization in conditions of 5:5 loads and 200-person sections gets nowhere if the institution will not back the report.
Running scenarios
High probability: The conditions that produce the greatest likelihood of unauthorized AI use are large asynchronous courses.
200 students. S = 0.1. A 5:5 load. L = 0.25. Async, unproctored. M = 0.1. No individual meetings. C = 0.1. Single upload. A = 0.1. An institution that won’t enforce. κ = 0.1.
0.1 × 0.25 × 0.1 × 0.1 × 0.1 × 0.1 = 0.0000025.
P = 1 − 0.0000025 ≈ 1.
The probability of unauthorized AI use is 100%. It will happen. Writing AI prohibitions into the syllabus under these conditions is theater.
Low probability: The conditions that produce the least likelihood of unauthorized AI are small, in-person labs and seminars.
15 students. S = 1. A 2:2 load. L = 1. In person. M = 1. Three meetings per student. C = 1. Oral defense. A = 1. An institution that adjudicates and backs you. κ = 1.
1 × 1 × 1 × 1 × 1 × 1 = 1.
P = 1 − 1 = 0.
The probability of unauthorized AI use is zero or close to it. The conditions will prevent it.
Middle scenario 1): These are the conditions that most faculty actually teach in.
30 students. S = 0.5. A 3:3 load. L = 0.75. Hybrid. M = 0.75. One meeting per student. C = 0.5. Drafts and revisions. A = 0.75. Written policy, slow enforcement. κ = 0.5.
0.5 × 0.75 × 0.75 × 0.5 × 0.75 × 0.5 = 0.053.
P = 1 − 0.053 = 0.947.
The probability of unauthorized AI use is almost certain.
Middle scenario 2) Also common conditions for many faculty .
35 students. S = 0.5. A 4:4 load. L = 0.5. In person. M = 1. Zero scheduled meetings. C = 0.1. Take-home final only. A = 0.25. Written policy, slow enforcement. κ = 0.5.
0.5 × 0.5 × 1 × 0.1 × 0.25 × 0.5 = 0.0031.
P = 1 − 0.0031 = 0.997.
Even with the class meeting in person, the absence of one-to-one contact and the take-home final drag the product down to almost nothing. M = 1 cannot save the product when C = 0.1. Faculty members who teach in person but never meet their students one-to-one have the same problem as the faculty member teaching async.
Most classrooms are these middle scenarios, which means that there is a 100% likelihood that there is unauthorized AI use occurring in nearly every college classroom.
Most American universities now run on adjuncts and lecturers carrying 4:4 and 5:5 loads, with sections of 30 to 200, increasingly asynchronous, with no time built in for one-to-one contact, with assessment by uploaded artifact because there is no time to grade anything else, at institutions whose enforcement processes take longer than a semester.6 The rule says those conditions produce P ≈ 1.
Discussion
I decided to create a rule because while everyone agrees that unauthorized AI use is happening,7 I’ve seen no good work toward dealing with it structurally.
According to my rule, the variables compound rather than compete. Some might object that the multiplication is too punishing. A dedicated faculty member with a huge class and a heavy load might still know her students through their writing. Maybe. But the faculty member is also drowning in work, a condition that is in institutional control.
The faculty member who reads this and thinks “I can score 1 on assessment by switching to oral defenses” is right but also defeated by L and S. Oral defenses for 200 students at a 5:5 load are not happening. The rule is multiplicative because the conditions multiply in the faculty member’s work life.
According to my rule, an institution running 5:5 loads, 200-person lectures, asynchronous delivery, no scheduled contact, single-artifact assessment, and a slow enforcement process has created the conditions for unauthorized AI use that a faculty member cannot prevent. State funding that requires institutions to create these conditions should understand that the state is creating the conditions for unauthorized AI use.
Input welcome
The formula will need revision. The variables will need revision. But the multiplication will hold, because the conditions hold, and the conditions are what the formula is for. Note that I have included state legislatures and higher education boards as an audience for this post. They are also responsible for the conditions of unauthorized AI use in the classroom.
The cheating-rate literature is large and the headline numbers vary wildly depending on definition and method. Donald McCabe's multi-decade survey work consistently found roughly two-thirds of students self-reporting some form of academic misconduct (McCabe, Treviño, and Butterfield, “Cheating in Academic Institutions: A Decade of Research,” Ethics & Behavior 11:3, 2001). Wiley’s 2024 update found 96% of instructors reporting that some students had cheated, up from 72% in 2021. Philip Newton and Keioni Essex's 2023 systematic review of online exam cheating found self-reported rates of 29.9% pre-COVID and 54.7% during (“How Common is Cheating in Online Exams and did it Increase During the COVID-19 Pandemic? A Systematic Review,” Journal of Academic Ethics 22, 2024).
Thank you to Keith Hankins for this classroom policy idea.
Carbone (1999), Sorcinelli (1994, 2002), and Weimer (1987) documented the connection between large lectures and academic integrity violations. Joseph Cuseo's literature review on large classes synthesized this evidence (“The Empirical Case Against Large Class Size: Adverse Effects on the Teaching, Learning, and Retention of First-Year Students,” Journal of Faculty Development 21, 2007).
The teaching-load variable has little direct empirical literature, but the closest evidence is Harper and Prentice 2024 on workload and detection, which documents faculty reporting that they cannot substantiate suspicions because they have no time allocated for the work (“Responsible but powerless: staff qualitative perspectives on cheating in higher education,” International Journal for Educational Integrity 20:25, 2024).
McCabe and Treviño's research line over decades repeatedly found that contextual factors — peer norms, faculty engagement, perceived institutional climate — predicted misconduct better than individual factors like age, gender, or GPA (see Decade of Research citation above). Cole and Kiss found students less likely to cheat when teachers were visibly engaged with their learning (“What Can We Do About Student Cheating?” About Campus 5:2, 2000). Broeckelman-Post found that students’ integrity behaviors were heavily shaped by perceived peer behavior and by faculty engagement (“Faculty and Student Classroom Influences on Academic Dishonesty,” IEEE Transactions on Education 51:2, 2008).
The AAUP’s 2025 data snapshot reports that about 68% of faculty members in US colleges and universities held contingent appointments in fall 2023, with 49% employed part time.
See for example Bryan Alexander, “Awareness of the AI cheating crisis grows,” AI and Academia (Substack), May 17, 2025:
Bryan Alexander, “Teaching with AI: strategies against cheating,” AI and Academia (Substack), July 14, 2025:
Bette A. Ludwig, “Purdue’s AI Cheating Scandal in a Computer Science Class, No Less”
James D. Walsh, “Everyone Is Cheating Their Way Through College,” New York Magazine, May 7, 2025.
“Another Undergrad Is Trying to Disrupt College With AI. He Says His Version Isn’t Cheating.”: https://www.chronicle.com/article/another-undergrad-is-trying-to-disrupt-college-with-ai-he-says-his-version-isnt-cheating
Beth McMurtrie, “Cheating Has Become Normal,” Chronicle of Higher Education, April 2025: https://www.chronicle.com/article/cheating-has-become-normal
Brilliant! One might quibble about the assigned weights, but not the process.
Employers should require colleges to include your equation for each course in the course transcript
The revived interest in oral defenses (here and in other discussions of combatting AI) is welcome, but it's far more difficult than just the logistics of scheduling those since there's also the question of grading and assessing them.
If it's simply the word of the instructor (or even their notes) vs. the student, that's going to result in a raft of lawsuits or very strong incentives for instructors to simply award As. Papers and exams at least have concrete evidence that can be quickly shared with others to support an assessment. A recording of these exams might work, but that creates additional privacy and logistical issues.
Oral exams only work if the public (and courts) trust faculty enough to let them assess students that way. Not convinced that this is possible currently.