Brown University take-home exam exposes how far AI cheating has spread

AI cheating just met a control group.

Brown University professor Roberto Serrano gave his welfare economics class its first take-home midterm this spring. The average came back 96 out of 100, with 40 of 86 students scoring a perfect 100. For context, the historical average in this course runs 65 to 80 percent, on easier exams.

The experiment

Numbers like that are not a smarter cohort; they are a signal. Serrano’s graders ran the test through ChatGPT and got answers mirroring what students had written.

So, with his dean’s approval, Serrano did what most institutions have not: he changed one variable. The final exam moved in-person, same course, same students, no unsupervised access to AI.

The average fell to 48.

The numbers that tell the story

The before-and-after gap is stark on its own, but the behavioral data around it is what makes this a case study, as reported by Fortune and Inside Higher Ed:

• Eighteen students dropped the course before the final
• Nine more skipped the final entirely
• Of those 27 students, 22 had scored a perfect 100 on the take-home midterm

The students who vanished when supervision arrived were overwhelmingly the ones who had aced the unsupervised exam. It is about as close to a controlled measurement of AI-assisted grade inflation as any classroom has produced.

A wake-up call for universities

Serrano calls it a wake-up call for universities still grading unsupervised work. His accidental experiment suggests the gap between what unsupervised grades say and what students can do on their own may be far wider than institutions assume.

The uncomfortable question is what follows. If a single course at Brown shows a 48-point swing between take-home and in-person conditions, unsupervised assessment anywhere becomes hard to trust at face value. Should take-home exams survive the AI era at all? After this semester, that debate has data.