Fields Medalist at Berkeley Abandons Online Grading as AI Inflates Scores

Fields Medalist at Berkeley Abandons Online Grading as AI Inflates Scores

A Fields Medal-winning mathematician at the University of California, Berkeley has walked away from grading his students’ online work. Richard Borcherds says he can no longer tell whether an answer came from a student or from a generative AI tool, and the rising averages on digital assessments have made the exercise pointless in his eyes. He has returned to conventional paper exams, with no electronic devices in the room.

  • Key Point: Borcherds, a Fields Medalist who works on infinite-dimensional algebras and quantum field theory, has stopped evaluating online submissions and moved his assessments back to paper.
  • Key Point: He says he cannot identify which students rely on AI, but average scores on online exams climb sharply, which he reads as a signal of machine assistance.
  • Key Point: Berkeley tested detection products including Turnitin and Rumi and found none reliable enough to deploy across the campus.
  • Key Point: For now, Borcherds awards credit simply for submitting homework, a stopgap he describes as far from ideal.

Why digital assessment lost its value

The problem, as Borcherds frames it, is structural rather than a matter of individual dishonesty. His grading model rewards correct final answers, and AI systems are good at producing exactly that. The format therefore pushes students toward the tools while putting those who work without them at a disadvantage.

“Online exams are now a waste of time because several students just use AI to answer the questions. I can’t tell which students use it, but I notice that the average score goes up a lot on online exams.”

He is equally critical of the incentive his own method creates. Paying to mark answers that a model generated, he argues, wastes both time and money, and honest students end up penalized for doing the work themselves.

“Grading for correction adds a huge incentive for them to use AI on the homework instead of doing it themselves, and it penalizes honest students. Paying to grade AI-generated answers is a waste of time and money.”

With no better option available, the professor now records homework as complete on delivery and does not assess its content. He concedes the arrangement is weak, but says he has nothing viable to replace it with at present.

A campus-wide gap in AI policy

The shift comes after Berkeley examined several tools built to flag machine-written text, among them Turnitin and Rumi, and judged none of them dependable enough for standard use across the university. That conclusion leaves instructors without a technical fix and pushes the question back to individual departments and faculty members.

The Berkeley Division of the Academic Senate now asks teachers to state in each course syllabus what use of AI tools is permitted. The specific rules remain up to each instructor and can differ by course level and subject area. Anne Marie Richard, associate vice chancellor for undergraduate education and director of academic technology, points faculty to a growing collection of sample AI statements hosted in the campus AI Hub, a directory of tools, training and policies that instructors can draw on when drafting their own guidelines.

Doubts about the future of human mathematics

Borcherds’ concerns extend well beyond the classroom. He spoke after OpenAI announced that AI agents had resolved one of the hardest open problems in mathematics, the existence and smoothness of Navier-Stokes, and he sees the pace of progress as a warning.

“If AI keeps improving at the current rate, in a few years there will be nothing human mathematicians can do that AI can’t do better.”

Asked what advice to give young people considering a mathematics degree, a path that seemed among the safest from automation only five years ago, he offers none. He recounted even putting the question to ChatGPT, which he says could not tell him what lies ahead either.

He does expect the academic system to absorb the shock slowly. Borrowing a comparison with print journals, which he says have kept publishing for two decades despite being largely useless, he argues that institutional inertia could keep mathematics departments largely unchanged for several years. Technical change has displaced skilled work before, he notes, with looms replacing hand weaving and cars replacing horses.

“But this is the first time something like this has happened in mathematics.”

The immediate consequence is that assessment design, not detection software, becomes the pressure point for universities. If Berkeley’s experience holds elsewhere, institutions may be forced to choose between supervised, device-free exams and grading schemes that no longer measure what a student can do unaided. How quickly mathematics departments adapt will depend less on the tools available than on how long they are willing to rely on inertia.