Fields Medal winner Richard Borcherds now gives homework credit for completion as artificial intelligence reshapes how he evaluates student work.
BERKELEY, Calif. — A prominent University of California, Berkeley mathematics professor has stopped grading homework based on whether students get the answers right, saying widespread use of artificial intelligence has made traditional take-home assignments a poor way to measure individual understanding.
Richard Borcherds, who won the Fields Medal in 1998 and is a professor in Berkeley’s mathematics department, said he now gives students credit for turning in homework rather than grading it for correctness. Borcherds said students can use increasingly capable AI systems to produce strong solutions outside the classroom, making it difficult to know whether a correct submission reflects a student’s own work. He has acknowledged that completion-based grading is an imperfect response but said he has not found a better alternative.
The change affects Borcherds’ own teaching rather than representing a campuswide grading policy. Berkeley records show he is teaching MATH 250A, Groups, Rings, and Fields, during the fall 2026 semester. The graduate-level course covers topics including group theory, rings, modules and fields. University guidance allows instructors to set course-specific expectations for generative AI and encourages them to state clearly how the technology may or may not be used on assignments.
Borcherds told the Daily Californian that he saw little value in paying readers to check homework for correctness when students could use AI to generate highly accurate responses. He said he had seen students receiving consistently high scores with help from the technology and concluded that AI use had become difficult to avoid. His concern is not simply that AI can supply answers, but that grades based on those answers may no longer show who actually understands the mathematics.
In separate comments reported by the New York Post, Borcherds said grading for correctness can create an incentive for students to use AI while putting students who complete the work independently at a disadvantage. He also said there is no easy way to distinguish between work produced by a student and work produced with AI assistance as the systems become more capable. Rather than trying to determine the source of every answer, he has shifted the homework portion of his assessment toward whether the assignment was completed.
The approach reflects a broader challenge facing colleges as generative AI tools become more capable at writing, coding and solving technical problems. Berkeley’s teaching guidance says instructors should explain their expectations for AI use and may allow, limit or prohibit it depending on the course and assignment. The university has also advised faculty to focus on learning goals and to consider whether assignments should measure skills students are expected to perform independently of AI.
Berkeley has been cautious about relying on automated AI-detection systems as a solution. Campus guidance says instructors should not assume such tools can conclusively establish whether a student used generative AI. That limitation leaves professors with choices such as changing assignments, conducting oral or in-person assessments, emphasizing the reasoning process or establishing course-specific rules about when AI assistance is acceptable.
Borcherds has said his concerns extend beyond homework grading to the future of mathematics itself. In an interview with the Daily Californian, he said continued rapid improvement in AI could eventually leave few mathematical tasks in which humans outperform machines. He also expressed uncertainty about what advice to give students considering mathematics careers, a striking position for a scholar who has received one of the field’s highest honors.
For now, Berkeley continues to leave much of the practical response to individual instructors. Borcherds’ completion-based homework system stands as one example of how faculty are adapting assessments when take-home answers can be generated with outside technological help. His comments also underscore a larger unresolved question for universities: how to measure what students know when producing a correct answer no longer necessarily proves they produced it themselves.
Author note: Last updated October 26, 2026.