A Shortcut With a Price Tag
When Roberto Serrano, a blind economics professor at Brown University, switched his final exam to in-person format after suspecting widespread AI use, scores dropped by 50 percent – a number that says more about how students had been studying than any survey ever could.
The Tool Students Were Never Supposed to Need
Ivy League admission is, at its core, a selection process for academically capable people. Students who reach Brown, Princeton, or any comparable institution have already demonstrated they can absorb difficult material, synthesize arguments, and perform under pressure. Generative AI, in that context, is not a crutch for students who can’t keep up – it’s a shortcut chosen by students who almost certainly could keep up, but decided the time wasn’t worth it.
That distinction matters. The concern isn’t that AI is filling a gap where human ability falls short. It’s that AI is replacing effort that students are fully equipped to make. Brown’s students aren’t struggling with economics concepts because the material is beyond them. They are, by most measures, among the more capable undergraduates in the country. What the Serrano situation reveals is a straightforward cost-benefit calculation playing out at scale: when an assignment can be outsourced to a chatbot, and when the risk of getting caught is low enough, a significant portion of students will outsource it.
A recent Princeton survey found that 29.9 percent of students admitted to using AI to cheat on at least one exam or assignment. That number comes from self-reporting, which means the real figure is almost certainly higher – people tend to undercount their own rule-breaking even in anonymous surveys. What Serrano’s classroom experiment adds is something more concrete than a percentage: an actual before-and-after performance gap, measurable in points, tied to a single change in exam format.
The format change was simple. No AI access. In-person. The results were not simple at all. Scores fell 50 percent. That’s not a marginal adjustment or a sign that the exam was harder. It’s a signal that a large share of the class had been submitting AI-generated work and calling it their own – and that without the tool, they couldn’t reproduce anything close to the same output.
What a 50 Percent Drop Actually Measures
Academic performance data is usually murky. Grade inflation, curved scoring, and varying rubrics make it hard to draw clean conclusions from semester-to-semester comparisons. But an exam administered by the same professor, in the same course, covering the same subject matter, with the single variable being AI access – that’s about as controlled as real-world data gets in a university setting. The 50 percent score decline is not an artifact of a harder test. It’s the delta between what students know and what their chatbots knew on their behalf.
Serrano is not letting this go quietly. The professor – who is blind, and whose reliance on adaptive tools of his own makes the contrast with his students’ AI use pointed, if unspoken – has continued to press the issue publicly rather than absorbing it as an institutional inconvenience. That posture matters. Most AI cheating situations at universities get handled through quiet academic integrity proceedings, if they get handled at all. A professor willing to stay visible in the conversation changes the dynamic.
The broader gadget-and-tech angle here is worth being direct about: generative AI has become, for a measurable segment of college students, a standard part of their workflow in a way that was not true three years ago. The tools are fast, accessible, and good enough at producing plausible academic prose that detection is genuinely difficult. AI detectors remain unreliable. Style analysis is contested. Without a format change like Serrano’s – physically removing the device from the equation – there is currently no clean technological solution to a problem that technology created.
Universities are caught in a familiar bind. They adopted digital submission systems, remote learning infrastructure, and cloud-based testing tools because those things were efficient. The same connectivity that made remote finals possible also made AI assistance frictionless. Locking things down now requires either reverting to paper exams and monitored rooms – an operational undertaking most departments aren’t staffed to handle at scale – or developing AI proctoring tools that raise their own serious concerns about surveillance accuracy and bias.
What makes the Brown case different from a general hand-wringing conversation about AI in education is the specificity of the evidence. Serrano didn’t flag suspicious essays or run text through a detector. He changed one variable and watched scores collapse. That’s a methodology any skeptic can follow. It’s also a methodology that other professors can replicate – and some, presumably, will, if they haven’t already started.
Ambition, Scheduling, and the Chatbot Calculation
Ivy League students are, as a population, intensely scheduled. Pre-professional programs, research positions, internships, club leadership roles, and social networks that carry real career weight all compete for the same hours that coursework demands. AI doesn’t just offer a way to cheat – it offers a way to cheat that feels, in the moment, like rational time management. The economics of that decision are almost cleaner than the academic integrity violation itself: if everyone else is doing it, not doing it is a competitive disadvantage, and doing it costs nothing but the ethics.
Serrano’s refusal to treat this as a closed matter leaves a specific question hanging over Brown’s economics department, and probably others like it: if 50 percent of exam performance was effectively being generated by AI, what did those students actually learn? Not what grades suggest – those grades were produced under conditions that no longer apply. What can those students do now, in a job interview, a graduate seminar, a consulting engagement, when there is no chatbot in the room and the work has to come from them?
