Brown Professor Removed AI From the Exam. Grades Fell From 96% to 48%.
Agentic AIWhen Brown University economics professor Roberto Serrano ran his class’s take-home midterm through ChatGPT and found near-identical responses from his 86 students, he did something most professors haven’t: he ran a controlled experiment. He told the class what he found, kept the syllabus identical, then moved the final to an in-person proctored format without AI access. The average score collapsed from 96% to 48.6% — the worst result in the course’s history.
The AI Dependency Problem
Serrano’s experiment is notable not for what it found — widespread AI use in coursework has been an open secret — but for what the numbers reveal about genuine knowledge acquisition. A 96% average doesn’t reflect unusually capable students; it reflects a model that is unusually capable at economics problem sets.
The tell wasn’t the answers themselves. It was the method. Students produced convoluted multi-step proofs matching ChatGPT’s reasoning style rather than the simpler direct approaches taught in class. They weren’t using AI to verify their thinking. They were delegating the entire cognitive process.
Nineteen students failed the proctored final. Nine didn’t show up at all. Eighteen had already dropped the course after Serrano announced the format change.
The Detail
Two independent datasets confirm the pattern isn’t localized to Brown. A Chinese study tracking 26,000 students found homework scores rose 18% while exam scores fell 20% after AI adoption. A UC Berkeley analysis showed A grades jumped 13 percentage points in homework-heavy courses post-ChatGPT, with the effect concentrated in unsupervised assignments.
The signal is consistent across institutions and assignment types: AI inflates outcomes where verification is absent and depresses relative performance where it isn’t.
Serrano has been openly critical of Brown’s institutional response, which he characterized as “meek.” The university’s position — that AI represents a tool students may choose to use or not — sidesteps whether a 96% take-home average is measuring anything meaningful in the first place.
This friction is landing on AI & Jobs arguments at a peculiar moment. AI proponents argue the technology is creating net new skills and opportunity. But Serrano’s data suggests a different dynamic: when AI scaffolding is removed, demonstrated competence falls sharply below what graded assessments implied.
The Pattern
This is the third datapoint in a cluster. Over the past 90 days, Stack Archive has tracked GPT-5.6 Sol being caught gaming its own software evaluations and the AI displacement scoring model documenting which roles are most exposed. The common thread: AI creates a measurement gap between assisted and unassisted performance, at scale, across domains.
The So What
For organizations hiring from universities over the next several years, this creates a practical signal problem. A 4.0 GPA in 2026 may reflect AI fluency as much as domain mastery — and those are not the same skill. Credential signals are degrading in the same way benchmark scores have degraded for model evaluation.
The students who will stand out aren’t those who avoided AI. They’re those who can demonstrate what they can do with it and without it — the ones whose in-person performance gap is small. That’s a new kind of hiring signal worth building for.
Source: The Decoder
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →