Brown University professor switches exam format after AI cheating suspicions

The average score on Brown University’s take-home midterm was 96 out of 100, with 40 perfect papers. When the same class sat an in-person final weeks later, the average fell to 48, according to an account reported by Ars Technica, citing interviews economics professor Roberto Serrano gave to El País and Inside Higher Ed.

The episode sits at the point where cheap text generation meets expensive credentials. Serrano’s ECON 1170 course had historically enrolled small numbers, but 86 students signed up in spring 2026 after he allowed take-home exams for both midterm and final, Ars Technica reports. A take-home format makes grading easier and can reward careful reasoning; it also turns “show your work” into a prompt. Serrano and his graduate students ran the midterm questions through ChatGPT and received similar results, and he noticed that many student answers shared a convoluted style even when correct.

Instead of launching a formal misconduct process based on writing patterns, Serrano changed the production conditions. He told students he would only count the midterm if the final’s grade distribution looked roughly similar; otherwise he would void the midterm and reweight the final, Ars Technica reports. Eighteen students dropped the course after the announcement, and nine did not attend the final. Of the 27 who dropped or did not show up, 22 had scored a perfect 100 on the midterm.

The numbers are blunt, but they do not resolve what universities actually need to manage: verification at scale. A single in-person exam can expose a mismatch between take-home performance and supervised performance, but it also reintroduces the cost universities have been trying to avoid—rooms, proctors, scheduling, and accommodation disputes—into courses that have grown larger precisely because assessment moved online. Meanwhile, the market for high grades is not an abstraction: students pay tuition for a credential whose value depends on other people believing it reflects skill.

Serrano’s case also shows why institutions are drifting toward procedural fixes rather than technical ones. Detecting AI use after the fact is probabilistic and contestable; changing the exam format is immediate and legible, even if it punishes students who used legitimate help and rewards those who can afford private tutoring. The tool that made the take-home exam attractive—outsourcing the hard parts of writing and reasoning—also made the result difficult to trust.

Brown’s evidence ended up looking like attendance sheets and a bell curve. The midterm produced 40 perfect scores; the in-person final produced an average of 48.