Why a UC Berkeley Math Professor Stopped Grading Homework Because of AI

Why a UC Berkeley Math Professor Stopped Grading Homework Because of AI

Grading college math homework used to mean spotting the exact algebraic slip where a student dropped a negative sign. Now it mostly means reading immaculate, machine-generated proofs that the student couldn't explain if you paid them.

That's the reality behind the viral story of a UC Berkeley mathematics professor who stopped grading homework altogether because of generative AI. Instead of spending hours evaluating problem sets produced by large language models, the course structure shifted entirely toward proctored, in-person exams and ungraded practice.

It sounds radical. Honestly, it's completely overdue.

If you've spent any time around university STEM departments recently, you know homework grades have become a statistical mirage. When an assignment carries even 10% or 15% of a final grade, the temptation for an overwhelmed sophomore to paste a differential equations prompt into an AI solver is overwhelming. And once one student does it to protect their GPA, the curve breaks for everyone else.

What Actually Happened in Berkeley's Math Classrooms

The core problem in higher-level mathematics isn't just that AI gets the right answer. Early models struggled with basic arithmetic and multi-step logic, giving professors a false sense of security. Modern reasoning models don't just spit out a final number—they write clean, step-by-step LaTeX proofs that look indistinguishable from a diligent graduate student's work.

A professor at the University of California, Berkeley—one of the top mathematics departments in the world—faced a straightforward dilemma. Teaching assistants were spending dozens of hours every week grading problem sets that clearly came straight out of an AI interface.

Worse, students were turning in flawless homework while failing midterm exams on the exact same material.

The gap between homework scores and exam scores became impossible to ignore. You can't accuse half a lecture hall of academic dishonesty without hard proof, and AI detectors are notoriously unreliable for mathematical notation. Standard proof techniques look similar no matter who writes them. There's no "voice" to analyze in a linear algebra proof.

So the professor pulled the plug on graded homework. Problem sets are still assigned every week, and solutions are still provided so students can check their understanding. None of it counts toward the final letter grade. Your entire grade comes down to what you can prove with a pencil and paper in a monitored room.

The Illusion of Competence

There is a massive psychological trap built into AI-assisted studying, and math exposes it faster than almost any other subject.

When you read an AI-generated proof, every step makes total sense. You nod along. You think, Yeah, I would have thought to substitute that variable there. Psychologists call this the fluency illusion. Recognizing a correct path is radically different from being able to construct that path from a blank page.

Math isn't a spectator sport. You don't build neural pathways by reading solutions—you build them by getting stuck on a single integral for forty-five minutes, trying three dead-end approaches, and finally seeing the structure of the problem.

When students outsource that friction to an AI tool, they rob themselves of the only part of homework that actually produces learning. Grading those submissions doesn't help the student, and it certainly doesn't help the teaching assistant burning out at midnight over a stack of synthetic LaTeX files.

Why AI Detectors Fail Completely in STEM

Administrators often ask why professors don't just run homework through plagiarism or AI detection software. Anyone who has actually graded a calculus or real analysis course knows why that fails.

First, mathematical prose is heavily standardized by design. Every textbook uses phrases like "Let epsilon be greater than zero" or "Without loss of generality." An algorithm looking for predictable token patterns will flag a completely honest math major every single time.

Second, smart students don't just copy-paste raw output anymore. They ask the AI for the solution, rewrite the variable names, change a few connecting words, and submit it. Proving misconduct in front of a university conduct board under those circumstances is a bureaucratic nightmare that consumes months of a faculty member's life.

Dropping homework grades bypasses the entire policing arms race. You don't need surveillance software when there is zero grade incentive to cheat on practice problems.

The Hidden Costs of Exam-Only Grading

Of course, shifting 100% of a course grade onto midterms and finals isn't a clean fix. Every veteran educator knows the downsides of high-stakes testing, and they're worth taking seriously.

Test anxiety is real. A brilliant student who understands the material cold can freeze up during a 50-minute midterm, misread a single boundary condition, and tank their grade for the semester. Graded homework historically acted as a safety net—a way for steady, hardworking students to buffer a bad exam day.

There's also a pacing issue. Without a weekly deadline attached to points, procrastination skyrockets. Many eighteen-year-olds simply won't touch an ungraded problem set until three days before the midterm, at which point cramming advanced math is basically impossible.

Professors using this model have to compensate by making weekly quizzes in discussion sections—short, 10-minute handwritten checks that keep students accountable without requiring hours of TA grading. It's a compromise, and it demands much tighter coordination across the teaching team.

How Students and Educators Should Adapt Right Now

If you're a student taking quantitative courses right now, treat ungraded homework like a high-stakes test anyway. Sit down with a blank sheet of paper, close every browser tab, and try to solve the problem for at least twenty minutes before looking at a hint. If you use an AI tool at all, restrict it to explaining why your specific attempt failed rather than asking it to generate the next step.

If you're an instructor or department coordinator watching homework averages hit 98% while exam averages sit at 54%, stop pretending the old grading rubric works. Reallocate your teaching assistants' hours away from grading take-home problem sets and put that time into running live problem-solving sessions, oral check-ins, or short in-class quizzes. Stop paying humans to grade work written by machines.

AJ

Antonio Jones

Antonio Jones is an award-winning writer whose work has appeared in leading publications. Specializes in data-driven journalism and investigative reporting.