96% on the Midterm, 48% on the Final: How a Professor Trapped 50 AI Cheaters Using Simple Math

Roberto Serrano had graded Brown University economics exams for 34 years, and he had never seen a class this sharp. The March midterm in his advanced mathematical economics course, ECON 1170, came back with an average of 96 out of 100. Roughly 40 of the 86 students earned a flawless 100. In a typical year the class average lands somewhere between 65 and 80 — and Serrano had written this particular exam to be harder than usual.

That was the first tell. The second arrived once his teaching assistants started reading closely. Tucked inside dozens of otherwise-immaculate answers was the same odd detour: a long, convoluted, contradiction-riddled argument for a result that had a short, elegant proof. When the graders typed the questions into ChatGPT, the chatbot produced the identical scenic route. The students hadn't just leaned on AI. They had copied its handwriting, contradictions and all.

What Serrano did next required no forensic software, no proctoring cameras, and no confessions. He used arithmetic. He announced that the final would be taken in person — in a room, on paper — and that if the class's grades no longer tracked the midterm, only the final would count. Then he let the numbers testify. They did. The average fell from 96 to 48, and the episode became what Serrano has described as the largest documented case of AI-enabled cheating in Ivy League history.

The exam that looked too good

The setup was, in hindsight, an accident of good intentions. Serrano had made the midterm a take-home, closed-book, honor-code exam for the first time in his decades at Brown, a move meant to lower the pressure on students during a semester rattled by a traumatic event that shook the campus community.

The result read like a statistical impossibility. A cohort that historically clustered in the 65-to-80 range on a standard exam had produced a 96 average — a median closer to 98, by the campus paper's account — on a test built to be tougher. Nearly half the room turned in perfect scores. No curve, no easier question set, no unusually strong year explains a jump like that. When students outperform the historical baseline by 20 or 30 points on a harder instrument, the exam is no longer measuring their economics. It is measuring something else.

Serrano's suspicion wasn't a hunch about individuals. It was a read on the distribution. Grades in a large class behave predictably from year to year, and a sudden, uniform surge toward the ceiling is the signature of an outside variable. In this room, the outside variable had a name.

The fingerprint nobody could explain away

The clincher wasn't the scores. It was the reasoning. According to Serrano's account, first reported by the Spanish daily El País, some answers contained unusual passages that matched the output produced when the graders ran the same questions through ChatGPT. On a proof that a competent student would finish in a few tidy lines, paper after paper wandered through the same tortured "convoluted contradiction" argument — the kind of over-engineered logic a language model generates when it is pattern-matching rather than reasoning.

Humans arrive at wrong answers in idiosyncratic ways. They make different errors, phrase things differently, and rarely botch a simple proof in exactly the same elaborate manner as the person three seats over. A chatbot does. When a dozen submissions reproduce a machine's specific, non-obvious detour, the shared source is not a study group. It is a shared prompt.

That is the quiet lesson buried in the Brown case. AI cheating tends to leave a fingerprint precisely because generative models are consistent. Ask ChatGPT the same question a hundred times and it gravitates toward the same structure, the same phrasing, the same avoidable overcomplication. The tool that makes cheating effortless also makes it legible to anyone who bothers to run the prompt.

Why the trap was just math

Here is the part worth studying, because it costs nothing and works anywhere. Serrano didn't try to prove that any single student cheated. Proving intent, one paper at a time, is slow, contestable, and easy for a well-lawyered appeal to unravel. Instead he set up a comparison the numbers would resolve on their own.

His announcement did two things at once. First, it moved the final into a proctored room, cutting off the AI entirely. Second, it told students their inflated midterm would only survive if the in-person final roughly matched it. A genuine 96 predicts a strong final. A borrowed 96 predicts a collapse. The design forced every student to, in effect, place a bet on whether their midterm grade was real.

"The empirical evidence of fraud is overwhelming." — Roberto Serrano, describing the case to El País

The response was its own confession. Before the final, 27 students dropped or simply failed to appear — and 22 of them had scored a perfect 100 on the take-home. Read that again: more than half of the class's flawless midterms belonged to students who would not sit the same material under supervision. You do not need a disciplinary hearing to interpret that. A person who earned a 100 does not flee a proctored exam on the same subject. The vanishing act was the verdict.

From 96 to 48

For the 59 students who did show up, the paper delivered the rest of the proof. The class average cratered from 96 to 48 out of 100. Nineteen students failed outright. The same cohort, the same professor, the same branch of mathematical economics — the only variable removed was the internet, and half the class's competence went with it.

A 48-point swing between two exams on the same content, taken weeks apart by the same people, is not a bad day. It is a control group and a treatment group hiding inside one classroom. The take-home measured students plus ChatGPT. The final measured students. The gap between the two numbers is a fairly clean estimate of how much of that dazzling midterm was never theirs to begin with — and it points to at least 50 students who leaned on the machine.

Brown's response was quieter than the scandal

For a case Serrano frames as historic, the institutional reaction has been muted. He has said the administration met his findings with something closer to silence than alarm; a faculty committee reportedly called the episode a "wake-up call," while the provost's office stayed publicly quiet. Cases like these typically route through Brown's standing committee on academic conduct, where an associate dean has framed most violations as products of pressure rather than malice — time crunches, curved grading, the sense that everyone else is doing it.

That framing captures the bind universities are in. Punish 50 students at once and you invite a due-process fight over evidence that is statistical rather than individualized. Punish no one and the honor code becomes a suggestion. Serrano's own answer was blunt and personal: he has abandoned take-home exams entirely. He is not alone. Several Brown economics colleagues have moved back toward in-person testing, with one telling the student paper he could not imagine any faculty member trusting a take-home exam again.

The number every campus is now staring at

Brown is not an outlier so much as an early, well-documented data point. Surveys keep finding that AI use has raced ahead of any policy meant to govern it. In one widely cited BestColleges survey, 22 percent of students admitted using AI tools in ways they themselves considered cheating, even as 51 percent said such use crossed the line. Pew Research has reported that 26 percent of U.S. teens now use ChatGPT for schoolwork, roughly double the share from a year earlier. Broader tallies put the portion of college students who have used AI tools somewhere north of 40 percent.

Detection has not kept pace. A University of Reading experiment that slipped AI-generated answers into a real exam pipeline found that 94 percent of them went completely unflagged by graders. Turnitin advertises high accuracy for its AI writing detector, but independent testing has repeatedly shown these systems miss a meaningful share of machine or lightly reworded text — and their false positives can wrongly brand honest students. Even the classroom itself is porous: an Arizona State University audit of in-person biology courses estimated that 45 percent of a typical course's points were reachable through some form of digital or AI-assisted shortcut.

Add it up and the appeal of Serrano's method becomes obvious. Software detection is an arms race the schools are losing. A proctored room and a before-and-after comparison of scores is not.

Blue books are selling again. Oral exams are back on syllabi. Professors are reweighting grades away from take-home work they no longer trust. None of it is elegant, and all of it costs class time. But the Brown case makes the trade-off legible: when the tool that enables cheating is nearly undetectable in the text, you stop policing the text and start testing the student. Serrano's 48-point drop wasn't a gotcha. It was a measurement — the difference between what his students knew and what a chatbot knew for them.