UC Berkeley Math Professor Stops Grading Homework Due to AI

Why grading for correctness broke down

Grading take-home problem sets rewarded AI generation instead of actual comprehension.

Richard Borcherds, an award-winning mathematician known for his work on infinite-dimensional algebras and quantum field theory, has seen enough of machine-generated problem sets. The Berkeley educator determined that assessing remote coursework has become an exercise in futility. According to Borcherds, AI tools have improved so rapidly that it is practically impossible to separate genuine student workings from synthetic output.

The core failure lies in the incentive structure. Borcherds pointed out that grading for correctness actively drives students toward automated assistance. When a student knows an algorithmic model can generate an accurate mathematical proof instantly, doing the work by hand exposes them to lower marks.

"Online exams are now a waste of time as several students just use AI to answer the questions," Borcherds stated. He added that attempting to detect the infraction is pointless, while universities waste money paying teaching assistants to evaluate machine-made text.

The death of automated AI detection

Borcherds is far from the only academic facing an honesty crisis in the classroom, but his response highlights a broader administrative retreat. UC Berkeley previously tested commercial AI-detection software and concluded the platforms were fundamentally unreliable, throwing both false positives and missed flags. Rather than imposing a broken campus-wide surveillance system, Berkeley allowed individual faculty members to set their own classroom policies.

Statistical verification models struggle across technical disciplines, much like how AI code models fail security checks in professional software engineering. Without dependable automated filters, professors must choose between policing assignments manually or changing the evaluation format entirely.

Returning to pen, paper, and physical rooms

Borcherds chose the low-tech path out. His math courses now base grades almost entirely on in-person, proctored examinations where students have no access to electronic devices.

The shift strips away the conveniences introduced during years of remote learning. While homework may still be assigned as practice, it no longer carries the high-stakes point values that invite shortcuts. For professors teaching advanced technical disciplines, stripping electronics from the test room remains the only defensible way to verify what a student actually knows.

What it means for college students

Expect the return of analog testing to accelerate across universities in the US and Europe. As automated detection utilities fall out of favor with faculty boards, departments handling quantitative subjects are shifting weighting away from weekly take-home assignments and piling assessment percentages into midterm and final exams held in physical lecture halls.

The illusion that automated proctoring and detection tools can salvage remote assessments is fading. By discarding homework points and turning back to pen and paper, higher education may finally be accepting the realities of generative software.