That Chart From Brown Isn’t Really About Cheating (opinion)

By now you may have seen the chart: two grade distributions from a single economics course at Brown University, shown side by side. On the take-home midterm, nearly the entire class is stacked at the top of the scale, with an average of 96 percent. On the in-person final, the distribution collapses. The average was 48.6 percent, the lowest in the course’s history. Between the two exams, 18 students dropped the class, nine more stayed enrolled but never sat the final and 19 ultimately failed the class.

The professor, Roberto Serrano, has taught the class on welfare economics and social choice theory at Brown for nearly two decades. This spring, for the first time, he gave a take-home midterm, an accommodation for students anxious about sitting in classrooms after December’s shooting on campus. When the scores came back implausibly high, he ran the exam through ChatGPT and found convoluted proofs similar to those his students had submitted. He told the class what he suspected, gave them the chance to prove him wrong and made the final exam in person. The chart is what happened next. Emma Whitford told the full story in Inside Higher Ed earlier this month.

The chart is already doing what charts like this do: It’s becoming shorthand for an epidemic. Students today will cheat whenever they can. This generation is lazy, doesn’t love learning, will outsource anything that isn’t policed.

I’ve spent the last 15 years studying motivation and engagement, and I want to offer a different reading. Not because the cheating didn’t happen; Serrano’s suspicion looks well-founded. But “did students cheat” is the least interesting question this chart raises. This is one course, in one semester, under one assessment format. What we’re looking at isn’t a portrait of a generation. It’s a single, unusually clean picture of what happens when a decades-old incentive structure meets a technology that removes all the friction. And on the question of what that incentive structure does to students, we have data.

What Students Say About AI

At the University of Pittsburgh, where I work, we’ve been asking students about how they’re using AI. When Pitt fielded the Student Experience in the Research Institution survey in spring 2024, 2,251 undergraduates responded, and the numbers told a less dramatic story than the headlines.

Among students who answered the question about the frequency of their AI use, 38 percent said they never used AI at all that academic year, and only 15 percent used it several times a week or daily. And when students did use it, the most common purposes weren’t drafting essays or completing assignments. They were using it for brainstorming, or for research, or for studying. Generating practice questions. Making flash cards. Checking their understanding. The same patterns held across more than 45,000 students at 11 peer research universities. Students may underreport, and AI adoption has grown since. But this is not a portrait of a generation itching to cheat.

Talking to students is where it gets interesting. In spring 2025, 13 Pitt faculty researchers sat down with 95 students across four campuses. The conversations confirmed what the researchers already knew: Most of those students were using generative AI. Most had an internal sense of which uses were helping them learn and which weren’t. Then the students revealed why they were reaching for AI in ways they knew weren’t helping: The deadline was tomorrow. The assignment felt like busywork. Or they’d hit the limit of their own understanding and had no idea where to turn next.

When explaining why they used AI, one student put it plainly: “I have a grade that I need to accomplish at the end of the day … If it’s either I do it … versus fail? I’d rather do it.”

She isn’t making a moral argument. She’s describing the game accurately. She hands in a product, receives a grade and the grade determines her scholarship, her graduate school application, her career. The learning, in that arrangement, is more or less beside the point. Eighty-two percent of the Pitt respondents agreed that AI can be detrimental to their own learning. They know. Pitt isn’t unique here, of course: Students across the country are saying the same thing. According to a survey from Student Voice and Inside Higher Ed, the top reason students use generative AI in ways that violate academic integrity rules is the pressure to get good grades. Students are making trade-offs inside a system that incentivizes them to prioritize good grades over learning.

And then there’s the detail I’d put next to the Brown chart in every faculty meeting in the country.

Some of the students in our focus groups asked their professors to bring back blue-book exams. Not because they wanted to be policed. Because the temptation to use AI, knowing their peers were using it, was so hard to resist that they wanted it removed. Students are asking us to take the temptation away. They aren’t defending a right to cheat. They’re telling us, in the plainest language they can find, that the game we built is one that almost no rational person can refuse to play. Indeed, Student 22 from the chart quickly achieved near-mythical status on social media on account of their exceptionality, their apparent refusal to join their classmates in playing the game.

Why the Shortcut Feels Rational

None of this should surprise anyone who knows the motivation literature. Edward Deci and Richard Ryan’s self-determination theory has documented, across hundreds of studies since the 1970s, that when external rewards become the primary reason to do something, the internal reasons tend to wither. For decades the contradiction held because the workarounds were costly and risky. AI removed the friction, and the incentive structure we’d been quietly running on stopped being self-enforcing.

There’s a second mechanism, and it’s the one that made me put my coffee down when I first encountered it. My colleague Scott Fraundorf, a cognitive psychologist at Pitt, ran a series of experiments with Afton Kirk-Johnson and Brian Galla in which students tried two study strategies and chose one going forward. Across experiments, the strategy that required more mental effort consistently produced better learning—and just as consistently, students rated it as worse. They read the effort itself as evidence the strategy was failing. Students can’t reliably tell productive struggle from failure, so the very thing that was helping them felt like proof they couldn’t succeed.

That finding sits at the center of everything I study. The most consequential moment in learning is what I’ve come to call the space between struggle and surrender, the moment when a student has hit difficulty and hasn’t yet decided what to do about it. In that moment they ask one of two questions: Can I do this? or How can I do this? They look nearly identical. They aren’t. The first seeks a verdict about the self, and it triggers self-protection. The second seeks a strategy, and it sustains effort.

Which question a student asks depends far less on the character of a student than on the signals her environment is sending. When the environment says the grade is what matters and the deadline is immovable, the signal is clear: Prove you can do this, or fail. AI then offers something uniquely corrosive, a shortcut that feels like competence. The polished output makes the messy paragraph the student was wrestling with look, by comparison, like evidence of failure. The tragedy is that the messy paragraph was where the learning was happening.

Now look at the chart again. The gap between those two distributions is being read as a measure of dishonesty. I think it’s better read as a measure of how completely the grade has replaced the learning as the point of the exercise for enough of the students that the difference shows up at the level of a histogram. That gap didn’t open this spring. AI just made it more visible.

Every System Gets the Outcomes It Designs For

There’s a maxim from health-care quality improvement, usually attributed to Paul Batalden: Every system is perfectly designed to get the results it gets. What I appreciate about that sentence is that it has no villains in it.

Consider one more number from the Brown story. Serrano’s course typically enrolled around 30 students. This spring, with take-home exams promised, 86 signed up. You could read that cynically. Or you could notice that students were choosing a course based on its evaluative architecture, which is exactly what the system has trained them to optimize, in a labor market where grades in certain courses function as tickets to certain careers. Students who treat a grade as a credential to be secured at minimum cost aren’t confused about what school is. They’ve read the design correctly.

And nobody in this story behaved unreasonably. The professor made a humane call after a campus tragedy, checked his evidence, told his students the truth and gave them a chance to prove him wrong. The students responded to the incentives in front of them the way rational people do. The university’s own committee on generative AI, reporting this month, urged faculty to de-emphasize punishment and acknowledged there’s no way to detect AI use with certainty.

Everyone is doing their job inside a design that none of them individually chose and none of them individually can change.

That’s why the cheating question is the wrong place to park our attention. Reducing cheating has less to do with individual morality than with the design of the learning environment: whether assignments have earned the student’s investment, whether struggling feels safe and purposeful rather than arbitrary, whether assessment makes thinking visible instead of just making products gradeable. That work isn’t cheap. It takes time faculty don’t have and training institutions mostly don’t provide. And even the professor who redesigns every assignment she controls still hands her students a grade at the end, because the grade isn’t hers to abolish. The evaluative architecture is built above the classroom, in transcripts and credit hours and the hiring systems that consume them. The classroom is just where the damage shows up.

But the signals inside a classroom are still ours to send, and everything I’ve learned about motivation says they matter more than any policy. Start by telling students the truth about effort, early and explicitly: This course is designed to make you struggle, and the struggle is the mechanism, not the verdict. Then lower the stakes of being wrong. Frequent, smaller assessments don’t just show us what students are thinking week to week; they remove the single high-pressure performance that makes the shortcut feel like survival. And ask to see the thinking, not just the product: a draft, a revision, a few minutes of conversation about how an answer came to be.

None of this AI-proofs a course. It does something better. It changes the question students hear from prove you can do this to how will you do this. And students, in my experience, tend to answer the question we actually ask them.

The chart from Brown will keep circulating, and it will keep being offered as evidence about the character of this generation. When I look at it, I think about the students in our focus groups, the ones who admitted everything, who could see the game and see themselves playing it and didn’t much like what they saw. The ones who asked us to bring back the blue books. They’ve already told us what the system is doing to them. The chart just drew it.