An economics professor at Brown University has provided what many consider to be some of the most compelling evidence yet of widespread AI-assisted cheating in higher education. Roberto Serrano, a veteran instructor of advanced undergraduate economics, watched as his class’s take-home midterm exam produced an astonishing average score of 96 out of 100. Suspicious, he then administered the final exam in person, under normal proctored conditions. The average plummeted to 48—a drop so dramatic that Serrano felt compelled to share the results publicly, first with El País and later with Inside Higher Ed.
The story began last December after a tragic shooting on Brown’s campus left two students dead. Many students expressed anxiety about sitting in crowded exam halls. In response, Serrano offered a fully take-home midterm and final to ease the psychological burden. It was the first time in his decades-long career that he relaxed exam protocols. The irony, he says, is bitter: the one time he tried to be compassionate, a large portion of the class cheated.
The Numbers That Gave It Away
Serrano’s course, ECON 1170, is an advanced undergraduate economics class that typically attracts a small, highly motivated group. Historically, enrollment rarely exceeded 30 students; once he taught just eight. This term, however, 86 students signed up. Serrano believes the take-home format drew many who might otherwise have avoided the challenging course.
The midterm results were, in his word, extraordinary. The class averaged 96, with 40 students earning a perfect score of 100. By contrast, the historical average for this same course hovers between 65 and 80. Moreover, Serrano had intentionally made this midterm harder than usual, reasoning that unlimited time would allow students to tackle more complex problems. Instead, the answers themselves carried a “very convoluted style,” as he described. Many correct responses matched almost verbatim what ChatGPT produced when Serrano and his graders fed the same questions into the AI.
The Trap He Set
To test his suspicions, Serrano designed a straightforward experiment. He announced that the final exam would be held in person, under proctored conditions. He told students he would compare the score distributions of the midterm and final. If they matched, he would keep the midterm grades as originally recorded. If they diverged significantly, he would void the midterm entirely and reweight the final to determine final course grades.
The results were damning. Eighteen students dropped the course before the final. Nine more simply did not show up for the exam. Of those 27 students who disappeared, 22 had scored a perfect 100 on the midterm. Among the students who did sit the final, the average score fell from 96 to 48. Serrano calculates that at least 50 of his original 86 students cheated on the midterm using AI. He calls the evidence “overwhelming.”
A Wider Reckoning in Academia
Brown University is far from alone in facing this crisis. A recent survey of Princeton students found that nearly 30% admitted to cheating on at least one exam or assignment, with the vast majority of cases involving generative AI. Across the United States, universities have spent the past two years scrambling to implement detection tools and redesign assessments. Some have returned to handwritten blue-book exams; others have introduced oral components or project-based evaluations that are harder to fake.
The issue is not confined to economics. Professors in computer science, history, political science, and even creative writing have reported similar patterns. Students often use AI to generate answers they then edit or plagiarize wholesale. Detecting this is difficult because AI-generated text can be paraphrased or mixed with original content. Traditional plagiarism checkers are largely ineffective against sophisticated AI use.
Serrano’s case is notable because it provides a before-and-after snapshot. The same cohort of students, tested on similar material, produced vastly different results when the AI crutch was removed. This natural experiment turns speculation into a concrete number: take away the AI, and half the apparent knowledge disappears.
Student Perspectives and Cognitive Concerns
Students themselves are not oblivious to the problem. Brown’s own provost-led report, released earlier this year, found that most undergraduates use generative AI on a weekly or even daily basis. At the same time, majorities expressed concern about the effect on their learning. They worry about what constant AI reliance does to their “cognitive capacity”—their ability to think critically, solve problems independently, and construct arguments from scratch.
This tension between convenience and education has created a strange paradox. Many students feel pressured to use AI because they believe peers are doing the same, and they fear falling behind on grades. The result is an arms race: professors try to lock down exams, while students find ever more clever ways to bypass restrictions. Some students have even begun using AI to generate essays that are intentionally riddled with subtle errors to evade detection tools, only to then edit them manually.
Why It Matters Beyond the Classroom
Serrano frames the issue in stark terms. “We cannot afford to have a society in which a significant fraction of our best young minds think that cheating is okay,” he told Inside Higher Ed. “That leads to a declining society, to a failed society. We cannot choose to become idiots.”
His concern extends beyond academic integrity. AI is already reshaping hiring practices, as employers look for candidates who can demonstrate genuine skills rather than credentials that may have been earned with AI assistance. In fields like economics, where quantitative reasoning and critical analysis are central, a degree that no longer signals competence loses its value. The broader risk, Serrano warns, is a generation of graduates who cannot think without algorithmic help.
Some universities have begun experimenting with “AI-integrated” pedagogy, teaching students how to use these tools ethically and transparently. Other institutions are moving toward oral exams, portfolios, and in-class writing exercises that require live performance. Yet none of these solutions are perfect, and the pace of AI improvement means detection methods quickly become obsolete.
Serrano’s small experiment—one class, one term—has resonated because it makes abstract worries tangible. When the AI is removed, half the knowledge vanishes. That is the uncomfortable number that universities, professors, and students must now confront if they hope to preserve the meaning of education.