Roberto Serrano’s Brown take-home exam collapsed: 96 average became 48 on the final

Roberto Serrano, a Brown University economics professor known for game theory, changed his ECON 1170 midterm to a take-home format after the campus shooting on Dec. 13. He did it for mental-health reasons. And then he discovered what he described as “absolutely ridiculous” AI-assisted cheating on a mass scale.

Here is the part that turns this from campus drama into a systems problem. Of the 86 students who took the March 5 take-home exam, 40 scored a perfect 100 and the class average was 96, despite the exam being designed to be harder than usual. Serrano says that when he later compared that midterm distribution to the in-person final, the final average collapsed to 48. The contrast is not subtle. It is the kind of distributional whiplash that signals the integrity of the whole learning and grading pipeline broke.

Why did Serrano take the risk of a take-home exam in the first place? According to the Fortune reporting, he wanted to remove stress for students who were still too traumatized to sit in a classroom setting after the massacre. Two students in his course had been among the nine wounded in the attack, both survived after weeks fighting for their lives. One of them, Ella Cook, had sat in his office days before the massacre and asked him to be her academic advisor. Serrano described having “a very nice conversation,” then said his shock hit when the names of the victims were released and he realized one was her.

What he got instead was an “AI-assisted cheating scandal,” as previously reported by El Pais. Serrano was tipped off by something “just too smart,” he said: answers included unusual passages that matched results after running the questions through ChatGPT. His graders then used ChatGPT to test the prompt-response pattern and found something crucial. The AI generated a convoluted argument for a problem that has a simpler, more elegant proof, and that same convoluted reasoning appeared across dozens of student exams. In other words, this was not generic “help.” It was pattern evidence.

At that moment Serrano had two options: invalidate grades, or recalibrate the evaluation system without immediately destroying trust. He chose not to void the midterm. Instead, he told students the final exam would be in person, and if the grade distribution did not roughly mirror the take-home midterm, only the final would count. After he returned to class post-grading, he addressed the evidence directly. Fortune reports he told students that pressing a button to ask an AI agent to do the work shows they are “completely irrelevant,” and he asked why they were at a university if they refuse to learn and work hard. He also said that if all they were doing was using a machine to do the work, they would think they need a Brown degree for that.

The reaction, he said, was “silence,” which is its own kind of data. He suspected many of the cheaters were not even in class. He reinforced the honor code and reminded students of the value of their signature. Following this speech, 27 students dropped the course, and 22 of those had scored 100 on the take-home. That is the kind of incentive shift you cannot ignore: the evaluation system became a mirror, and some students chose to step away when the mirror did not flatter them.

Then the final came. Only 59 students showed up for the in-person exam and 19 failed. Serrano reports the in-person class average fell to 48 out of 100, “by far the lowest final exam average in the course’s history.” He characterized the evidence of fraud as overwhelming, saying the combined distributions made it “absolutely clear.” He sent the evidence to Brown’s dean of the college and provost, and neither responded initially. After he escalated the case to the university’s Academic Code Committee, he received a note calling the incident “a wake-up call.” Serrano says the provost has maintained complete silence to this day.

Zoom out, and Brown becomes a case study in how quickly AI changes incentive structures, even for places that rely on honor systems. Serrano, who holds the Harrison S. Kravis University Professorship in Economics and serves as an editor at Games and Economic Behavior, frames the broader trend as a tsunami for institutions. He also ties the issue to a worrying market signal: if elite universities stop producing students who actually learn, the label may lose meaning over time. His frustration is not only about exams; it is about what universities are manufacturing.

Second-order implications for executives and boards are hard to miss, even if your world is not higher ed. When evaluation and credentialing fail, trust becomes the scarce resource. Fortune’s broader reporting puts the cheating trend into context: Princeton’s faculty voted in May to end its 133-year-old honor code tradition of unsupervised exams, mandating proctors starting July 1, the most significant change since students first petitioned for it in 1893. Fortune also reported that 57% of U.S. college students say they use AI tools in their coursework weekly, and a separate Fortune analysis found AI can contribute to “cognitive atrophy,” with educators warning of a “great unwiring” of independent reasoning. And just last week, 47% of surveyed Harvard seniors admitted to cheating.

Serrano has already changed his system for the coming academic year. Weekly homework will carry zero weight toward final grades because those can be completed with AI. Take-home exams are gone, permanently. He acknowledges AI can be useful for learning, but he argues that academic integrity cannot be dropped and that institutions need guardrails and consequences if they fail. In a world where a “press a button” workflow can manufacture outputs that look right, the test stops being the exam. The test becomes whether organizations can align incentives with real learning, real work, and real accountability.