Fields Medalist and UC Berkeley math professor: Continuing to grade by correctness no longer makes sense
According to a report on October 23 by The Daily Californian, the student newspaper at the University of California, Berkeley, the school’s mathematics professor Richard Borcherds, a 1998 Fields Medal winner, has changed the way he grades homework. He said that because students using AI to complete assignments has become widespread, many are continuing to receive high grades as a result, and continuing to grade by correctness has no point, while paying graders to mark AI-generated answers is also a waste of time and money.
Borcherds now gives grades only based on whether homework is submitted (completion-based), and correctness no longer counts. He acknowledged that this is not a very good system, but said there is no better way at present. Final grades mainly depend on in-person exams that prohibit electronic devices; he said online exams have become a waste of time because some students simply use AI to answer, and average scores have risen noticeably.
He further said that it has become difficult to distinguish between human- and AI-generated homework. In the past, AI would expose itself through characteristic mistakes, but now AI has improved to the point that it rarely makes obvious errors. Grading by correctness would only encourage more students to use AI, and would instead punish those who honestly do the problems themselves. He hopes that eliminating correctness-based grading will reduce students’ motivation to go directly to AI for answers, increase their opportunities to do homework themselves, and emphasized that students still need to truly understand the material in order to perform well on exams. In a follow-up email interview with Newsweek, Borcherds reiterated that AI can now solve many undergraduate mathematics problems and often does so better than students themselves.
Regarding career prospects, Borcherds said he is no longer sure whether he should advise young people to pursue mathematics research. Five years ago, mathematics seemed to be one of the safest professions in the face of AI disruption, but he believes, If AI continues to improve at present rate, then, in a few years, there’ll be nothing human mathematicians can do that AI can’t do better. He even asked ChatGPT how he should advise students, and ChatGPT also said it did not know what the future would look like. The interview took place after OpenAI announced that an AI system seemingly solved the Navier-Stokes existence and smoothness problem.
At the university level, Berkeley’s Academic Senate requires faculty to state in their syllabi what the course expectations are for AI tools, but specific policies vary greatly depending on the course level and field; the school had tested AI detection tools such as Turnitin and Rumi and found them insufficiently reliable, so it did not implement a campus-wide unified policy, and the policy is left to individual faculty members to decide.
Source:
The Daily Californian (Berkeley student newspaper), 2026-10-23, primary interview source: https://www.dailycal.org/news/campus/academics/ai-use-leads-uc-berkeley-math-professor-to-see-no-point-in-grading-homework/article_bdf1eae2-c08b-491f-b9f1-81a7cea177de.html
Verification points: stop grading for correctness, grade based on submission, no better solution, ChatGPT Q&A, judgment on the future of mathematics, university AI policy.
New York Post / The California Post, 2026-10-25/26: https://nypost.com/2026/10/26/us-news/uc-berkeley-math-professor-drops-online-tests-as-ai-use-spikes/ and the preceding piece
Verification points: in-person exams with electronic devices prohibited, inflated online exam scores, distorted incentives from grading by correctness.
Newsweek, 2026-10-26 (including Borcherds’ email reply): https://www.newsweek.com/berkeley-professor-grading-homework-ai-12619345
Verification points: AI can now solve undergraduate math problems better than students, hopes to reduce reliance on AI, exams remain effective.
ABAB AI Insight
Historical and institutional path: In university homework grading, the traditional function has been both “practice feedback” and “proof of ability.” Once AI largely invalidates the latter, the center of evaluation is forced to shift sharply from “process homework” to “closed, in-person exams,” following the same logic as paper-and-pencil exams used to counter open-book work or ghostwriting, except that the technical barrier is higher and detection is harder. Berkeley did not mandate unified detection tools, but left discretion to faculty, reflecting the inertia of academic freedom at U.S. research universities more than the urgency of standardized assessment.
Capital and technological path: Generative AI companies such as OpenAI are rapidly improving mathematical problem-solving ability, directly affecting the higher-education assessment market. Teachers giving up paid grading amounts to acknowledging that the marginal cost of detection and manual grading has become higher than the benefit; exam proctoring and in-person testing may in turn create new demand for on-site assessment, oral exams, or proctoring services.
Horizontal comparison: Similar practices have appeared sporadically in other STEM courses (some advanced math courses already place more emphasis on submissions and exams). But a Fields Medalist publicly questioning the value of advising students shifts the issue from “cheating” to whether the discipline itself still has a human comparative advantage, which is more jarring than ordinary teacher complaints and harder to absorb with measures such as “strengthen integrity education.”
Judgment: In the short term, this is a partial failure of the assessment mechanism and a defensive adjustment; in the medium term, if AI’s mathematical ability continues to improve at the current slope, the function of university mathematics education as “proving human understanding” will shrink further, with greater reliance on in-person reasoning, oral defenses, and original research training that AI cannot directly take over. Borcherds’ own admission that it is “not a very good system” shows that there is no clean solution at present, only trade-offs.
Cognitive law:
When detection costs exceed the costs of incentive distortion, the system will abandon fine-grained screening and move toward coarse-grained process confirmation (turning in homework earns credit) plus high-threshold in-person verification.
Once tool capability generally exceeds the average level of those being evaluated, the original mapping of “correct answer = understanding” breaks down, and grading degrades from an ability signal into a compliance signal.
The moment when top practitioners publicly question career advice often comes before the ordinary market reaches consensus, because they are the first to perceive the speed at which comparative advantage is being eroded.