Grades dropped from 96 to 48 percent when a Brown professor made students take the exam without AI

Roberto Serrano, an economics professor at Brown University, became suspicious when his 86 students averaged 96% on a take-home midterm, well above the historical 65-80% range. He ran the questions through ChatGPT and received nearly identical answers, including a convoluted mathematical proof that many students had used.

After warning the class, Serrano made the final a proctored in-person exam. The results were dramatic: 18 students dropped the course, 9 did not show up, and the average fell to 48.6% — the worst result in the course's history. Only a handful of students scored near their take-home performance, while 19 failed outright. Serrano voided the midterm and weighted the final at 80% of the course grade.

Serrano described the university's response as 'meek,' telling him to report each cheating case individually. He argued for a stronger stance, warning that tolerating cheating leads to a 'failed society.'

Two recent studies back up Serrano's experience. One study from central China tracked over 26,000 students in grades 7-12 over 30 months. Six months after students started using AI, homework scores rose by 18% while completion time dropped from 64 to 45 minutes. However, exam scores fell by 20%, with long-term losses of 18-24% on entrance exams. About 81% of long-term users fit the pattern of faster homework, higher homework grades, and poor exam scores, with top students losing 24% of performance.

A UC Berkeley study of more than 500,000 grades at a large Texas research university showed that in courses heavy on writing and programming assignments, the share of A grades jumped 13 percentage points after ChatGPT launched. The increase was concentrated in unsupervised homework, with courses relying on homework seeing a 16 percentage point higher increase than those relying on proctored exams.

The convergence of Serrano's case with large-scale studies suggests that AI-assisted cheating on unsupervised assignments is widespread and measurably undermining the validity of grades. However, it remains unclear what proportion of students are using AI, and whether the pattern holds across disciplines and educational levels.

The professor's call for a stronger institutional response highlights a tension: universities must balance trust in students with the need for academic integrity. Individual reporting of cheating is resource-intensive and may be ineffective if the problem is systemic. The findings also raise questions about the long-term impact on learning, as students who rely on AI for homework may fail to develop essential skills, a concern echoed by the China study's finding of significant performance drops on entrance exams.

Uncertainties remain: the studies are observational and do not prove causation; the China study is limited to one region; and the UC Berkeley data may not reflect the latest AI models. Nonetheless, the evidence points to an urgent need for educational institutions to rethink assessment design and integrity policies.