A Brown University professor discovered that his economics class average soared to 96 percent on a take-home exam, but collapsed to 48.6 percent when he switched to a proctored, in-person final—a gap he attributes to widespread AI cheating. Two recent large-scale studies confirm the pattern: students show higher homework and assignment grades after adopting AI tools, yet their exam scores decline by 18–24 percent. The finding raises concerns that reliance on AI for coursework may mask learning gaps that surface under supervised testing.
Summaries like this, in your inbox every morning.
Sign up free →What happened
A Brown University economics professor observed his class average jump to 96 percent on a take-home exam—well above the historical 65–80 percent range—then plummet to 48.6 percent on a proctored in-person final. He traced the take-home scores to ChatGPT by running the questions through the tool and finding nearly identical answers; 18 students dropped the course after the proctored exam was announced, and nine did not show up at all.
Why it matters
Two large studies confirm the pattern. A Chinese study tracking over 26,000 students in grades 7–12 found that six months after AI use began, homework scores rose by 18 percent while exam scores fell by 20 percent; entrance exam losses ranged from 18 to 24 percent over time. A UC Berkeley study of over 500,000 grades at a large Texas research university showed A grades jumped 13 percentage points in courses with heavy unsupervised homework after ChatGPT launched. The implication is that students may rely on AI for assignments but struggle on supervised tests, raising questions about the durability of their learning.
What to watch
The professor asked the university for a stronger institutional response, but administrators told him to report each cheating case individually—a response he called 'meek' and 'ridiculous.' He emphasized the societal risk: 'We cannot afford to have a society in which a significant fraction of our best young minds think that cheating is OK.' Further discussions are ongoing.
The Brown case illustrates a stark divergence between unsupervised homework performance and supervised exam results when AI tools are available. Serrano's observation—that nearly every student scored near-perfect marks on the take-home midterm, only to see grades collapse on the proctored final—points to a structural problem: AI can enable homework completion without corresponding knowledge retention. Two independent studies, one from central China and one from a large Texas research university, validate this pattern at scale. The Chinese study, which followed 26,000 students over 30 months, showed that the negative effect on entrance exams took about two years to fully manifest, with 81 percent of long-term AI users fitting the pattern of faster homework but weaker exam performance.
The UC Berkeley study adds nuance: the 13 percentage point jump in A grades after ChatGPT launched was concentrated in unsupervised homework, and the effect was more pronounced in courses heavy on writing and programming assignments. This suggests the benefit accrues specifically where human oversight is absent. Serrano's institutional frustration—that his university asked him to report cheating cases one at a time rather than adopting a systemic response—underscores a mismatch between the scale of the problem and the response mechanisms in place. His concern that normalized cheating among top students poses a societal risk reflects a broader anxiety: that AI-enabled shortcuts may erode the credential value of academic achievement itself if left unaddressed.
No discussion yet for this article
Get curated AI news from 200+ sources delivered daily to your inbox. Free to use.
Get Started FreeFree · takes 30 seconds · unsubscribe anytime
1 minute a day. The AI essentials.
200+ sources · Email / LINE / Slack
