A straightforward comparison of exam scores by a Brown University economist has produced some of the most concrete evidence yet of how generative AI is reshaping academic dishonesty. Robert Serrano, an economics professor at the Ivy League school, noticed a suspicious pattern after grading a take-home midterm. Students who had performed poorly on an in-person final exam scored surprisingly high on the midterm. The gap was too large to explain by normal variation.
The Score Gap That Raised Red Flags
Serrano taught an intermediate microeconomics course that included both a take-home midterm and an in-person final exam. The midterm allowed access to course materials but did not explicitly ban AI tools. After grading both exams, Serrano plotted the results on a chart. The visualization revealed a cluster of students who scored in the top quartile on the midterm but fell to the bottom quartile on the final. For Serrano, the pattern pointed to one explanation: widespread reliance on generative AI to answer midterm questions.
Implications for Academic Integrity
The data underscores a growing challenge for universities. Take-home assessments, long valued for testing deeper understanding, have become vulnerable in the generative AI era. Students can input questions into large language models and receive polished answers within seconds. Traditional honor systems or vague AI policies do little to stop this. Serrano's approach, however, offers a practical detection method: compare performance across controlled and uncontrolled settings.
Institutions, however, face a difficult trade-off. Eliminating take-home exams restricts the kind of open-ended problem-solving that many courses aim to teach. Keeping them requires either sophisticated monitoring software or acceptance of AI use, which undermines grade fairness. Serrano's scatter plot demonstrates that the problem is not hypothetical. It is measurable and widespread.
Why This Matters
The scale of AI cheating revealed by Serrano's data pushes education toward a reckoning. If a single midterm in one course can show such a clear signal, the problem across entire universities is likely enormous. This affects not just grade integrity but also the value of degrees. Employers and graduate schools may begin to question transcripts from institutions that have not addressed the issue transparently. The data also shifts the burden of proof. Professors can no longer rely on hunches; they now have a statistical method to verify suspicions. The long-term consequence may be a fundamental redesign of how college students are assessed, with less weight on unsupervised written work and more on oral exams, in-person problem sets or project-based evaluations.
What Comes Next
Serrano has not proposed a specific policy change, but his findings are likely to fuel debates within Brown and beyond. Other faculty members may replicate the analysis in their own courses. Administrators may consider stricter AI prohibitions or invest in proctoring technology. Students, meanwhile, face a clearer choice. The data shows that relying on generative AI for take-home exams creates a measurable risk of failure when the crutch is removed. For educators, the lesson is that the cheating problem is no longer invisible. It can be charted, quantified and confronted directly.



