Brown University Professor Says Exam Data Proves Mass AI Cheating

Roberto Serrano, an economist and professor at Brown University, has published figures that, in his words, prove mass use of AI by students on exams: the average score on the take-home midterm was 96 out of 100, but after he moved the final exam to an in-person format with no access to neural networks, it dropped to 48 out of 100.

What the professor noticed

Serrano taught an economics course and, like many instructors, gave students a take-home midterm that could be written outside the classroom. The class average held at 96 out of 100. That is a near-perfect result for the whole class at once, which looked suspiciously uniform for a difficult economics course.

Why the in-person exam changed everything

To test his hypothesis, Serrano moved the final exam to an in-person format: students wrote it in a classroom, without computers or access to chatbots. The result was striking: the average score fell to 48 out of 100, roughly half what it was on the take-home midterm.

• Take-home midterm: average score of 96 out of 100
• In-person final exam without AI: average score of 48 out of 100
• The difference: almost a factor of two
• Source of the observation: Roberto Serrano, professor of economics at Brown University

Serrano made the story public, presenting the score gap as quantitative evidence that a significant share of students used AI tools on homework and on exams that are not supervised in person.

"We can't afford to get dumber,"

Serrano warns, commenting on the consequences of students losing the ability to solve problems on their own.

Why this is hard to prove directly

Accusing an individual student of using AI on homework is almost impossible to prove formally: there is no camera, no proctoring, and no technical way to capture the moment of cheating. That is why Serrano's approach looks more convincing than many other instructor complaints about AI cheating. Instead of catching specific students, he built a controlled comparison of the same class under two different conditions, with and without access to the tools. The score gap speaks for itself, with no need to prove any one person's guilt.

What this means

Professor Serrano's case is one of the clearest public examples of how the gap between remote and in-person assessment results exposes the scale of AI use at universities. For instructors, it is one more argument for bringing back in-person exams and rethinking assessment formats in the era of accessible chatbots: as long as an assignment can be completed at home without supervision, the result no longer reliably reflects the student's own knowledge.