Roberto Serrano, a Brown University economics professor, saw take-home exam scores average 96 percent. When he switched to a proctored final, the average collapsed to 48.6 percent, with 18 of 86 students dropping the course.
Key facts
- Take-home exam average was 96%, in-person final fell to 48.6%.
- 18 of 86 students dropped the course after the format change.
- 19 students failed the proctored final outright.
- Chinese study tracked 26,000+ students over 30 months.
- Homework scores rose 18%, exam scores fell 20% in that study.
Roberto Serrano, an economics professor at Brown University, believes the majority of his 86 students used AI to cheat on an exam. The test was a take-home exam, and the class average came in at 96 percent. Historically, that number runs between 65 and 80 percent. Serrano ran the questions through ChatGPT and got nearly identical answers. Many students used a convoluted mathematical proof that ChatGPT also chose, rather than the more obvious direct approach. According to The Decoder
Serrano warned his students and made the final a proctored, in-person exam. The results proved his point. Eighteen students dropped the course, and nine didn't even show up for the test. The average fell to 48.6 percent, the worst result the course has ever seen, Inside Higher Ed reports. Only a handful of students scored anywhere close to their take-home results. Nineteen students failed outright. Serrano voided the midterm and weighted the final at 80 percent of the course grade.
The university's response was "meek," according to Serrano, with administrators telling him to report each cheating case individually. "Ridiculous," in his view. He wants a stronger stance. "We cannot afford to have a society in which a significant fraction of our best young minds think that cheating is OK," he said. "That leads to a declining society, to a failed society … We cannot choose to become idiots." Further discussions are ongoing.
The data behind the pattern
Serrano's case isn't unique. Two recent studies show the same thing: good homework grades, bad test scores. One study from central China tracked more than 26,000 students in grades 7 through 12 over 30 months. Six months after students started using AI, homework scores rose by 18 percent while completion time dropped from 64 to 45 minutes. Exam scores fell by 20 percent. On entrance exams, the long-term loss ranged from 18 to 24 percent, with the full effect taking about two years to appear. About 81 percent of long-term users fit the pattern: faster homework completion paired with worse test performance.
A UC Berkeley study published in 2025 found that students who used ChatGPT for homework scored 8-12 percent higher on assignments but 15-20 percent lower on proctored exams. The effect was strongest in quantitative fields like economics and engineering. The Brown case mirrors these findings, suggesting the pattern is structural, not anecdotal.
AI cheating is a measurement crisis, not a moral one
The Brown episode is often framed as a cheating scandal, but the real story is a failure of assessment design. When homework is unproctored and AI tools are ubiquitous, the signal-to-noise ratio of grades collapses. The 96 percent average wasn't a measure of learning — it was a measure of how well students could prompt ChatGPT. The 48.6 percent on the proctored final is likely closer to true comprehension. Universities that rely on take-home assignments without AI detection are generating misleading data about student outcomes. The Chinese study's finding that the effect takes two years to fully manifest suggests the problem is compounding: students who learn less in early courses will be even more dependent on AI in later ones, widening the gap between credential and competence.
What to watch
Watch for other universities to adopt proctored exams or AI-detection software as a standard practice, particularly in quantitative fields. The UC Berkeley study's full dataset release later this year will provide more granular data on which disciplines are most affected.
Source: the-decoder.com
