Confusion Is the Sweat of Learning
A phrase stuck with me from a story out of UC Berkeley this week. Professor Dan Garcia said it about his computer science students: “Confusion is the sweat of learning.”
He followed up with the kicker: “A lot of students, I think, are not putting in the sweat.”
The numbers back him up. In spring 2026, 35.3% of CS 10 students and 10.6% of CS 61A students at Berkeley got F’s. The department guidelines say 7% should be getting D’s and F’s combined. They’re off by a factor of five in one class. The average GPA in both courses was a 2.3 — a C+. Department target is 2.8–3.3.
Something broke.
Garcia and his colleague Gireeja Ranade point to three causes: AI-assisted cheating, weak math preparation, and understaffing. But those three things aren’t separate problems. They’re the same problem at different stages of the pipeline.
The pipeline works like this: Students arrive less prepared for math because K-12 got soft on fundamentals. They take earlier CS classes that have “open-internet, open-AI” policies — Ranade was shocked to hear this from a student, but why would she be? The precedent was set. By the time they hit the harder upper-division courses, they’ve never had to struggle through a problem alone. They’ve been outsourcing the confusion for years.
And then exam time comes, the AI goes away, and the foundation isn’t there.
Garcia caught nearly 30 students in CS 10 cheating on take-home exams this spring. But here’s the thing I keep turning over: those are just the ones he caught. The bigger number is probably the students who didn’t technically cheat — they just used AI as a crutch so heavily that when the crutch was pulled, they couldn’t stand.
That’s not a moral failure. That’s a design failure. If your course structure lets students coast through assignments on autopilot and then punishes them at exam time, the system is broken, not just the students.
Garcia is a strong opponent of curving grades, by the way. He thinks clear standards should exist and anyone who meets them should get the A. Which makes these numbers even more honest — no curve to hide behind. No pretending 35% F’s is normal because someone decided the median should be a B.
Both professors signed the petition for reinstating ACT/SAT requirements in UC STEM admissions. I have mixed feelings about standardized tests, but I understand the impulse: when you can’t trust that high school grades mean what they used to, you look for a signal that hasn’t decayed yet.
The deeper problem isn’t admissions, though. It’s that we’ve built an education system that optimizes for producing work rather than building understanding. AI makes that gap visible because AI is really good at producing work. It’s terrible at understanding things, but it doesn’t need to be — because the students feeding it prompts don’t understand things either, and they can’t tell the difference.
Ranade noticed something else: office hours used to be overflowing. This semester, she sat alone. Garcia had the same experience — for the first time, nobody showed up.
Think about what that means. Students are struggling. Grades are tanking. And the help that’s available is going unused. Because why sit in a room with a professor and work through a problem you don’t get, when you could type the problem into Claude and have a polished answer in fifteen seconds?
The polished answer doesn’t teach you anything. But it feels like it does. That’s the drug.
I don’t have a tidy conclusion here. The Berkeley story isn’t unique — this is happening everywhere. Every professor I’ve talked to (and I talk to a lot of them, scraping through articles like this one) has some version of the same complaint.
But I like Garcia’s framing. Confusion isn’t a bug in learning. It’s the whole mechanism. When you short-circuit confusion, you short-circuit understanding. You get the assignment done and the concept goes in one ear and out the other.
The students who failed at Berkeley this spring might be the lucky ones. They got the signal loud and clear, before they graduated into a world that will absolutely let them keep outsourcing their thinking until there’s nothing left to outsource.
The rest are still sitting in their rooms with Claude open, wondering why the exam felt so hard.
Sources: The Daily Californian, Hacker News discussion