AI Doesn't Make You Dumber. It Makes You Dependent.

AI Doesn't Make You Dumber. It Makes You Dependent.

A Brown economics professor's students averaged 96% on a take-home exam, with AI in hand. Take the AI away, and the same class collapsed to 48%.

That 48-point gap went viral this week as yet another cheating scandal. That's the wrong read. What Roberto Serrano stumbled onto by accident is the one thing no AI productivity study ever measures: not what a tool gets you while you're using it, but what's left once you unplug it. Here, the answer comes down to a single number: half.

And this isn't just about students. It's coming for the developer who stopped reviewing the code they ship, the lawyer who stopped drafting their own clauses, the consultant who stopped building their own slides.

The number is real, we checked it against the source. The story everyone's telling about it, though, gets three things wrong, and the real version is scarier than the shorthand.

A take-home exam, offered for good reasons

Roberto Serrano has taught economics at Brown for thirty-four years. His course, ECON 1170, covers welfare economics and social choice theory. It's an advanced, heavily mathematical class, the kind that scares off anyone just browsing the catalog.

On December 13, 2025, a campus shooting killed two students. Serrano had just agreed to mentor one of the victims. The following spring, several of his students told him how anxious they felt about sitting a proctored exam. So, for the first time in nearly twenty years, he made the midterm a take-home.

The course usually pulled in around thirty students. That semester, 86 signed up. Serrano himself credits the spike to the promise of a take-home exam. An act of compassion turned into a selling point, and that's already the whole story in miniature.

Grades that make no sense

The midterm landed on March 5. The class average: 96 out of 100. Forty of the 86 students turned in a perfect paper.

For context, this exam has historically averaged between 65 and 80. And Serrano had deliberately made it harder that year, precisely because students had unlimited time. A tougher exam producing the best average anyone's ever seen should thrill any professor. Instead, Serrano ran the questions through ChatGPT.

What came out looked a lot like what students turned in. In his words, the answers were "roughly correct, but very roundabout, in a very convoluted style." Study groups submitted identical homework, down to the same errors in the same spots. Serrano estimates at least fifty of his students cheated. That's his own count, and it's worth keeping in mind for what comes next.

The exam room, and the no-shows

Serrano wrote to his class. He laid out his suspicions and announced a proctored final.

Eighteen students dropped the course immediately. Nine more skipped exam day. Twenty-seven gone out of 86. And here's the detail that outweighs every hot take: of those 27 who vanished, 22 had scored a perfect 100 on the midterm.

Fifty-nine students sat the proctored final in May. The average cratered to 48.6. Three papers scored zero. Nineteen students failed. It's the worst result this course has ever recorded.

The real number isn't the one everyone's quoting

This is where it's worth slowing down, because the most damning statistic isn't 96 versus 48.

Of the 59 students who sat the final, only two scored within ten points of their own midterm grade. Two. Exactly one student did better in May than in March.

That difference matters. A harder final knocking the whole class down a notch is normal, it happens everywhere. But it doesn't scramble the ranking: the strong students in March are still, roughly, the strong students in May, just a few rungs lower.

Here, the correlation vanished entirely. Whatever looked like skill in March had evaporated by May. No amount of exam difficulty explains that.

Now, the nuance everyone's skipping

This needs saying plainly, because most of the coverage buried it: this isn't a scientific experiment. Serrano never re-ran the same exam without a machine, whatever you might have read elsewhere.

These are two different tests at two different points in the semester: a midterm in March, a final in May. They're also two different groups, 86 students on one side, 59 on the other. Comparing two averages calculated on different populations, with different content, is a field observation, not an experimental protocol. And the class sat that final in May right after being collectively accused of cheating, which doesn't exactly help anyone think straight.

Add in Brown's response: the university says it's taking the matter seriously, but that the professor didn't hand the academic code committee enough to build cases against anyone. No disciplinary action followed. Nobody was found guilty of cheating. These are documented suspicions, not a conviction.

Except the nuance doesn't undercut the number. It strengthens it, and that's the twist in this story.

The 27 students who left before the final were the most suspect of the bunch, since 22 of them had landed a perfect 100. Had they stayed to sit the exam, the average would probably have dropped even lower. So 48.6 isn't an exaggeration: it's a conservative estimate, reached only after the prime suspects had already walked.

The whole affair reads more like a crime scene than a lab: no protocol, but plenty of evidence. And all of it points the same way.

The blind spot in every study

Back to what this actually measures, now that we know what it's worth.

Every single study on AI productivity compares a worker using the tool to one who isn't. None of them compares that same worker, six months later, with the tool taken away. We've already seen how badly AI productivity numbers hold up under scrutiny: it's the same blind spot, and it's structural. Nobody has an incentive to measure the dependency they're creating.

Serrano measured it by accident. And his result doesn't say his students are bad. It says that between March and May, what looked like skill was actually just access. Take away the access, and there's nothing left to take.

This is where the story stops being about a university. A developer who ships code they no longer review gets their own 96%, right up until the day they have to debug something in production that nobody actually understands. Dependency stays invisible as long as the tool keeps working. That's exactly what makes it expensive.

Serrano has a blunt line for all this, and it's not kind to diplomas: if all you do is press a button and let the machine do the work, do you really think you need a Brown degree for that?