An economics professor at Brown University uncovered stark evidence of AI-enabled cheating in his classroom. On a take-home exam, his 86 students averaged 96 percent. The professor grew suspicious of the uniformly high scores and switched the final exam to an in-person, proctored format.
The results were dramatic. Eighteen students dropped the course entirely. Nine others failed to show up. Among those who took the proctored exam, the average score plummeted to 48.6 percent. A 47.4-point collapse reveals the performance gap between AI-assisted work and genuine student knowledge.
The professor's experience aligns with findings from two major studies. Researchers at UC Berkeley and universities in China analyzed student performance across both unproctored and proctored assessments. Their data confirmed what Brown observed: students who rely on AI for homework and take-home assignments show dramatic score declines when facing supervised exams.
The pattern suggests students are using AI tools like ChatGPT not to learn material but to bypass the learning process entirely. When forced to demonstrate knowledge without AI assistance, many students lack the foundational understanding needed to pass.
This creates a measurement problem for educators. Take-home exams no longer reflect student competency. They measure access to AI tools. Universities face a choice: abandon take-home assessments, explicitly allow AI use and grade accordingly, or redesign coursework to make cheating harder and learning deeper.
Brown's professor exposed a systemic issue across higher education. Many instructors don't catch AI cheating because they lack tools to detect it reliably. Others haven't adjusted assessment methods since generative AI became mainstream. The 96-to-48 percent drop is not an anomaly. It's evidence that AI adoption in education has outpaced institutional responses, creating assessment validity problems that grades alone can no longer hide.