A Brown professor caught 40 of 86 students cheating with AI. Now he wants take-home exams gone.

A Brown University economist watched a take-home midterm come back almost perfect. The median score was 98%, and 40 of his 86 students got a flawless 100%, according to reporting in El País. In a class where the midterm average usually lands between 65 and 85, that is not a good semester. It is a signal.

The professor, Roberto Serrano, did not need a software flag to know what happened. He saw the same odd reasoning, what he described as a convoluted contradiction argument, surface in answer after answer on the same question, plus near-identical responses from students who had worked together. His response was public and blunt: he denounced the cheating, pushed the case to Brown’s academic code committee, and decided to move his exams back in person. The story hit Hacker News this week and turned into the latest flashpoint in a fight that almost every instructor in the country is now having quietly. The interesting part is not that students cheated. It is what the episode reveals about how little control schools actually have over it.

How routine this has become

Two and a half years after ChatGPT launched, AI on coursework is not an edge case. It is the baseline. Depending on the survey, somewhere between half and the large majority of college students report using AI tools for assignments. A widely cited HEPI and Kortext survey found that about 1 in 5 students admit to putting AI-generated text directly into work they submit as their own. The gap between “used AI somehow” and “handed in AI’s words” is where most of the argument lives.

Here is the finding that should reframe the panic. Stanford researchers who surveyed students across 40 high schools found that the cheating rate barely moved after ChatGPT. As they put it, “the percentage of students who admitted to cheating has remained flat since the advent of ChatGPT,” sitting between 60 and 70% both before and after generative AI became easy to reach. Read that twice. The tool did not create cheaters. It handed existing ones a faster, cleaner, harder-to-spot method, and it lowered the bar for everyone else who was on the fence.

That is why a take-home exam is now a different object than it was in 2022. A prompt that a motivated student once spent three hours on can be answered in 30 seconds, in prose that reads better than most undergraduates write. When the cost of cheating drops to near zero and the odds of getting caught stay low, the behavior spreads from the margins to the median. Serrano’s 98% median is what that looks like in a gradebook.

Why detectors can’t save anyone

The obvious fix, run everything through an AI detector, does not work well enough to lean on. This is the part many administrators still underestimate.

Independent testing keeps landing far below vendor marketing. A 2024 study by Perkins and colleagues ran six major detectors against text from current frontier models and measured baseline accuracy around 39.5%. Turnitin, the most widely deployed tool in higher ed, publishes a variance of roughly plus or minus 15 percentage points on its own scores, which means a 50% reading could honestly be anywhere from 35 to 65. The company suppresses scores under 20% because its internal testing found that range unreliable, a quiet admission that the low end is noise.

The false positives are the dangerous failure. Multiple studies have found that writing by non-native English speakers gets flagged at sharply higher rates, with one review reporting up to roughly a third of non-native essays misclassified as AI. Paraphrasing tanks accuracy across the board. Put plainly: a detector can torch an honest ESL student’s record while waving through a fluent cheater who ran one rewrite pass.

Institutions are acting on this. Curtin University switched off Turnitin’s AI-writing detection across all campuses as of January 1, 2026, citing reliability and equity concerns, specifically the higher false-positive rates for some student groups. Curtin still reviews work for integrity. It just stopped pretending a percentage from a black box counts as evidence.

There is also a legal and reputational tail here that detector vendors don’t advertise. A wrongful accusation built on a detector score is the kind of thing that ends in an appeal, a grievance, or a lawsuit, and the school is the one holding the bag, not the software company whose terms disclaim accuracy. That asymmetry is why cautious general counsels have started treating detection output as a tip at best, never as proof. Notice what Serrano did and did not rely on: not a detector score, but score distributions, repeated reasoning quirks, and matching answers from study partners. Human pattern-matching, backed by evidence a student can be asked to explain. That is currently the only approach that survives contact with an appeals committee.

Assessment is quietly being rebuilt

If you can’t reliably detect the tool, you change the test so the tool can’t take it for the student. That is the shift now underway, and it is the most consequential outcome of this whole mess.

At Brown, Serrano and his colleagues are pulling exams back into the room. As one of them, assistant professor Bobby Pakzad-Hurson, told The Brown Daily Herald, “I don’t see how, especially after seeing this, any faculty member could have any confidence in a take-home exam.” Pakzad-Hurson also cut the grade weight of homework so that assignments AI can finish in seconds count for less. Across the discipline, the same handful of moves keeps recurring: in-person blue-book exams, oral components where a student has to defend their reasoning out loud, in-class writing done under supervision, and a general migration of points away from anything a chatbot can do unattended.

There is a cost to this, and it is worth naming. In-person, handwritten, time-boxed exams favor fast writers and the test-confident, and they penalize the same students that take-home formats were designed to help: people with disabilities, anxiety, jobs, caregiving duties. Two decades of pedagogy pushed assessment toward authentic, untimed, real-world work for good reasons. AI is shoving it back toward the bluebook. Brown’s associate dean Love Wallace offered the humane read on the students themselves, noting that those “who violate the academic code are almost never doing it from a malicious place.” Most are responding to incentives a frictionless tool created. The fix is structural, not moral.

The other half of the rebuild is grading what AI can’t fake: process over product. Drafts, revision history, in-person checkpoints, work that builds on a student’s own earlier in-class output. It is more labor for instructors already stretched thin, which is the unglamorous reason adoption is uneven.

What this means for you

If you are a student, assume the take-home essay as a graded artifact is on its way out in any course that takes integrity seriously, and that your ability to explain your own work in person is becoming the thing that actually counts. If you teach, the lesson from Brown is not “buy a better detector.” It is that detectors are a liability you can be sued over, and that redesigning the assessment beats policing the tool. And if you are anyone who hires based on a transcript, sit with the uncomfortable question underneath all of this: when a take-home grade can be generated in 30 seconds, what is the credential certifying? The schools moving fastest right now are the ones answering that honestly, by making students perform the skill in a room where the chatbot can’t follow.