When Brown University economics professor Roberto Serrano moved his final exam from take-home to in-person, the class average plummeted from 96 out of 100 to 48, exposing what may be the clearest evidence yet of widespread ChatGPT use in unmonitored academic work. The shift triggered an immediate exodus: 18 students dropped the course, and 9 more failed to show up for the exam. Among those 27 departures, 22 had earned perfect scores on the midterm.

What Happened at Brown's Economics Class?

Serrano introduced take-home exams in spring 2026 after a December 2025 campus shooting left him shaken. The policy change had an unexpected consequence: enrollment in ECON 1170 jumped from a historical cap of 30 students (sometimes as few as 8) to 86. The March 5 midterm produced 40 perfect scores, far exceeding the historical average band of 65 to 80 percent.

The professor grew suspicious not just of the scores, but of the writing itself. Correct answers contained convoluted phrasing that felt off. When Serrano and his graduate students ran the exam questions through ChatGPT, the model returned strikingly similar answers. Rather than void the midterm outright, he announced the final would be proctored and emailed students that he would compare the two distributions before deciding whether to count the take-home scores.

"Historically the average grade in the midterm of this course has ranged between 65 and 80 percent, and this exam was harder than the exams I wrote in the past, because take-home is an opportunity to challenge the class a little bit more," said Roberto Serrano, Brown University economics professor.

Roberto Serrano, Brown University economics professor

The response was swift. Eighteen students dropped after the announcement. Nine enrolled students did not attend the final. Among those who sat for the exam, the average score collapsed by half. The composition of departures is difficult to explain on any grounds other than cheating: 22 of the 27 students who left had scored a perfect 100 on the midterm.

How Widespread Is AI Cheating in Higher Education?

The Brown case is not an isolated incident. Survey data from peer institutions suggests systematic patterns. A Princeton survey found that 29.9 percent of students admitted to using AI to cheat on at least one exam or assignment. At Brown itself, a provost-led report documented that 56 percent of undergraduate respondents and 67 percent of graduate and medical students said they use generative AI tools daily or weekly, with large majorities expressing concern about the effect on their own learning.

Serrano has refused to let the episode fade quietly. He has taken the story to El País and Inside Higher Ed and argued that Brown's administrative response has been muted. The 50 percent decline in average score is not a marginal cheating signal; it represents a controlled experiment in what happens when proctoring is removed and then restored within a single semester.

How Universities Are Responding to AI Cheating

• Shift to Proctored Assessments: Universities are accelerating the move back toward in-person, proctored, and oral exams as a direct response to AI-enabled cheating in take-home formats.
• Bifurcation of AI Tools: The education market is splitting into tools that help students learn (tutoring, spaced repetition, feedback) and tools that help students submit work (drafting, solving, summarizing), with the latter now under institutional scrutiny.
• Regulatory Pressure on AI Companies: OpenAI, Anthropic, and Google all market education-tuned versions of their assistants and have published guidance on responsible academic use, but these guidelines have limited force against students optimizing for grades under deadline pressure.

For AI companies, the ECON 1170 case is the kind of data point that regulators and university administrators will cite for years. OpenAI, Anthropic, and Google all market education-tuned versions of their assistants, and each has published guidance on responsible academic use. Those guidelines have limited force against a student optimizing for a grade under deadline. The Brown incident suggests that the second category of AI tools, those that help students submit rather than learn, faces an institutional reckoning.

Serrano's numbers put the question of AI in education on trial in a way that surveys and anecdotes cannot. The same cohort, the same material, the same semester, the only variable being proctoring. A 50 percent decline in average score is not a marginal cheating signal; it is evidence of a structural shift in how students approach unmonitored work when AI assistance is available.