Karmactive Team
UC Berkeley math professor Richard Borcherds, the 1998 Fields Medal winner, has stopped grading homework for correctness because AI tools now solve advanced mathematics so reliably that checking problem sets penalises the minority of students still working through problems on their own. Borcherds told the university’s student newspaper that today’s AI models rarely produce the characteristic errors that once allowed graders to identify machine-generated work.
Borcherds has shifted assessment entirely to device-free in-person examinations, treating them as the only reliable way to verify whether a student has learned the material independently. When he asked ChatGPT what AI would do to mathematics, the system said it did not know. He won the Fields Medal—mathematics’ highest honour—in 1998 for his work on the Moonshine conjecture and infinite-dimensional Lie algebras. His decision carries particular weight because mathematics was widely expected to resist AI disruption for years longer than writing-heavy subjects.
For STEM college students, the safety cushion of using weekly homework sets to boost a low exam average is disappearing. As faculty abandon correctness-based grading and universities like UC Berkeley opt out of unreliable AI detectors, course grades are shifting toward high-stakes, device-free in-person exams. Students who relied on AI to clear weekly problem sets now face unassisted midterms and finals as the dominant measure of their academic standing.
The institutional response at Berkeley extends beyond any single faculty decision. UC Berkeley’s Research, Teaching, and Learning Commons formally reviewed a pilot program for Turnitin’s AI-writing detection tool and opted out, citing academic research showing detection tools produce unreliable results and cannot consistently distinguish AI-assisted work from human writing. The RTL Commons now directs faculty to set clear syllabus expectations about AI use and to redesign assessments around in-person, device-restricted formats rather than attempting detection after the fact.
For decades, mathematics departments relied on paid student graders to mark problem sets, assuming minor calculation quirks would expose machine-generated work. Borcherds has observed that current AI models have eliminated those telltale errors entirely, creating a counterproductive arrangement: universities paying human graders to evaluate AI-produced mathematics while simultaneously penalising the small number of students turning in imperfect human work.
Does UC Berkeley use Turnitin to detect AI homework? No. Following an internal pilot program and a review of academic studies, UC Berkeley’s Research, Teaching, and Learning Commons opted out of Turnitin’s AI-writing detection tool. The university advises faculty to set clear syllabus rules and shift toward in-person assessments without electronic devices.
Borcherds summarised the stakes directly: “If the worst effect of AI on society is that it is harder to assess homework then I will be very relieved.” His in-person exam model offers one practical framework, but how universities will adjust course-level grade weights and manage impacts on graduate teaching roles remains unresolved.
For related context, see Karmactive’s earlier coverage and latest explainer.
