Brown Caught 50 Students Using ChatGPT On One Midterm. Here's What Actually Failed.

Key Takeaways

- Roberto Serrano, Brown University economist, flagged at least 50 students using ChatGPT on the midterm - Midterm: 96/100 average with 40 perfect scores out of 89. Final (in-person): 48/100 average - 22 of 27 students who skipped the final had aced the midterm with a perfect 100 - The core issue: take-home, closed-book exams and AI tools fundamentally don't mix

---

Roberto Serrano doesn't sound alarms for fun. He's the Harrison S. Kravis Professor of Economics at Brown. Decades of published research. Serious credentials.

So when he says he has "conclusive evidence", 50 students, ChatGPT, the midterm — I paid attention.

The exam was take-home, closed-book.

ECON 1170. Advanced mathematical economics.

40 out of 89 students hit 100.

Class average: 96 out of 100.

Half the class aced an advanced undergrad math economics exam. Perfect scores everywhere. That's not a bright class. That's a broken assessment.

The graders flagged it first. Irregularities in the answers. Serrano ran them through ChatGPT. Exact matches — "unusual passages" reproduced verbatim. Then he noticed something else: students who supposedly studied together turned in identical exams. Missing the same questions in the same spots. That's not collaboration. That's copy-paste.

Then came the final. In-person. No ChatGPT. No take-home. Average dropped to 48 out of 100. Twenty-seven students who took the midterm didn't show.

Twenty-two of those 27 had perfect 100s on the midterm.

They knew what was coming.

---

Why ChatGPT Cheating at Brown Was Inevitable

Brown's AI policy is unambiguous.

AI-assisted work = academic dishonesty. Penalties range from assignment failure to expulsion.

Clear rules. Completely unenforceable against a student with a laptop and a ChatGPT subscription.

Brown economics professor Rajiv Vohra put it simply: AI is not a problem with in-person exams.

It may be an issue with homeworks or take-home exams.

That's the polite academic way of saying what Serrano's data proves.

Take-home, closed-book exams were designed for an era when the hardest part of cheating was physically getting the answers. That era is over.

Serrano is switching to fully in-person next year. Weekly exercises won't count toward the final grade anymore. They could be done with AI. No more take-home exams.

It's the right call.

It's also a retreat. Honest students get policed harder because of the dishonest ones. Professors spend more time proctoring than teaching. Employers and graduate programs get less signal from transcripts, not more.

Here's the thing that should keep you up at night. One exam question had a "very simple, direct proof." ChatGPT produced a "very convoluted contradiction argument." Still scored points. The grader didn't catch it.

Cheating badly still worked.

---

What ChatGPT Cheating at Brown Means for Hiring

Here's what this means if you're evaluating people.

A student who scores 96 on a take-home math economics exam at Brown is not the same student who scores 48 on an in-person final.

Probably not even close. Both get the same degree. Both show the same GPA. Both land on a resume as "Brown University, Economics."

You already knew transcripts were imperfect signals. Now they're worse. A high GPA means a student performed well on assessments that may have been AI-assisted. In courses where take-home work dominated. In an environment where detection rates are low and penalties are slow.

That doesn't mean every high-GPA student is fraudulent. It means you can't use GPA as a proxy for the skills you actually need.

If you're running a graduate program: you're admitting students based on a signal that's actively degraded.

The students who got into Brown on their own merits still get in.

The ones who figured out how to use ChatGPT to fake their way through? Same credentials. You can't tell the difference from a transcript.

Brown's penalty structure is right. Failure, suspension, expulsion. Enforcement after the fact doesn't fix the credential inflation already shipped. The graduating class already shipped. The damage to the next class is still being calculated.

---

How to Actually Verify Skills in the ChatGPT Era

Serrano landed on in-person exams.

Not elegant. But it works. Vohra confirmed the same pattern across the economics department: AI is not a problem with in-person exams.

Settled.

The harder problem: what about courses where in-person assessment is impractical?

Language courses. Project-based work. Anything requiring creative or extended problem-solving. Those are a different category now. Portfolio assessment. Live demonstration. Oral examination. Not any format where a student can open a browser tab.

For small businesses hiring: you already knew to test people on what they can actually do. This confirms it. A degree is table stakes, not proof of competence. Run your own assessment. Ask for work samples. Talk to references who actually saw the person ship something.

For universities: this is an arms race you're losing. Detection tools are always one step behind generation tools. Serrano caught this since graders were paying attention and the grade distribution was absurdly implausible. That's not scalable.

The structural answer is redesigning assessments so AI assistance can't substitute for the thing you're actually measuring. Brown is making the right structural changes for its own courses. The question is whether the rest of higher education follows, or whether credential inflation continues until the signal is worthless.

---

FAQ: ChatGPT Cheating at Brown — Quick Answers

How many students were caught using ChatGPT at Brown? At least 50 students in ECON 1170, identified by economist Roberto Serrano using ChatGPT output matching and exam pattern analysis.

What was the midterm score distribution? 96 out of 100 average. 40 perfect scores out of 89 students. The in-person final averaged 48 out of 100.

What happened to the students who skipped the final? 27 students who took the midterm didn't show for the final. Twenty-two of those 27 had scored a perfect 100 on the midterm.

What are Brown's penalties for AI-assisted cheating? Ranging from assignment failure to expulsion.

Can you detect AI cheating on take-home exams? Serrano caught it through grader reports, ChatGPT output matching, and identical exam patterns. Not automated detection tools. Scalable detection remains unsolved.

---

What Serrano exposed isn't just about 50 students who got caught. It's about the assessment format that let them try. Take-home, closed-book exams were reasonable in the past. They're not reasonable now. The only question is how long it takes the rest of academia to admit that.

If you're evaluating people, stop trusting the credential. Start demanding the work.