AI cheating signals drop 50% when Brown professor switches to proctored exams

Roberto Serrano, an economics professor at Brown University, stumbled onto compelling evidence that his students were using generative AI to cheat on exams—and the data tells a striking story. When he allowed take-home tests, performance soared. When he switched to proctored exams, scores plummeted.

How the pattern emerged#

Serrano's discovery began with a painful context. After a shooting on Brown's campus in December 2025 that left two people dead, including someone who had recently met him, the professor made what seemed like a compassionate decision: allow his ECON 1170 course to use take-home exams for both midterm and final assessments.

The policy change had an unexpected effect. Enrollment in the course—which had historically capped at around 30 students, sometimes as low as eight—suddenly jumped to 86 students. More striking still was the performance spike. When students took the midterm exam at home on March 5, the average score reached 96 out of 100, with 40 students achieving perfect marks.

The telltale reversal#

For context, Serrano noted that the historical midterm average in this course had ranged between 65 and 80 percent. The dramatic improvement seemed even more improbable because, as he explained to Inside Higher Ed, he had intentionally made the exam harder than versions from previous years, reasoning that take-home conditions warranted more challenging material.

When Serrano required the final exam to be completed in person under standard proctoring conditions, the results flipped. Student performance dropped approximately 50 percent compared to the take-home exam, suggesting that the earlier stellar results were not genuine reflections of learning.

Broader context on academic integrity#

Research into AI cheating at elite universities supports Serrano's observation. A recent survey at Princeton found that roughly 30 percent of students admitted to using AI to cheat on at least one exam or assignment. Serrano has gone public with his findings, sharing detailed accounts with publications including El País and Inside Higher Ed, underscoring growing alarm over how AI tools are reshaping academic evaluation at competitive institutions.