UC Berkeley math professor Richard Borcherds no longer checks whether students solve their homework correctly. An assignment now counts if it is simply handed in. The reason is AI: students with chatbots consistently get top scores, and banning AI is no longer possible. The Daily Californian, Berkeley's student newspaper, reported this on October 23.
What changed in grading
At American universities, homework is usually graded not by the professor but by paid readers - upper-year undergraduates or grad students. Borcherds decided there is no point in paying them to check answers for correctness anymore: AI use, as he put it, is now unavoidable.
Homework now earns points for being handed in, not for being right. The professor admits it is "not a very good system", but for now he has no alternative.
This fall Borcherds teaches MATH 250A "Groups, Rings, and Fields", a graduate algebra course with 50 students. This is not school arithmetic: Sylow theorems, Galois theory, finite fields.
Who Richard Borcherds is
Borcherds is not just any instructor tired of cheating. In 1998 he won the Fields Medal, the top prize in mathematics, for proving the monstrous moonshine conjecture, which links the largest sporadic simple group, the "Monster", to modular functions.
He gave the interview after a high-profile announcement from OpenAI. On September 8 the company said its model had proved that a solution to the Navier-Stokes equations, which describe fluid flow, can break down in finite time. It is one of the seven Millennium Prize Problems from the Clay Mathematics Institute, and mathematicians are still checking the proof.
His forecast is blunt: "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." According to him, mathematics has never seen anything like this: looms put hand weavers out of business, cars replaced horses, and now it is his field's turn.
He doesn't know what to tell students
Five years ago, being a mathematician looked like one of the careers safest from AI. Now Borcherds is not sure whether to encourage students to go into math at all. "I have no idea what to advise. I even asked ChatGPT this exact question, and even ChatGPT said it didn't have any idea what was going to happen," he told the paper.
He does not expect quick change, though. The academic system has a lot of inertia, and Borcherds would not be surprised if the department stayed the same for another couple of years. As an example, he pointed to printed academic journals: they keep coming out even though they have been "utterly pointless" for two decades.
What the university is doing
Berkeley has no single AI rule. The Academic Senate asks instructors to state in the syllabus whether AI tools are allowed, and each instructor sets the rules: they vary a lot by course level and subject. Sample wording for syllabi is collected in the campus AI Hub.
Detectors did not work out: the university tried Turnitin and Rumi but found them not reliable enough to adopt campus-wide.
Berkeley saw this coming in the spring
The homework story is not the first one. In June the same Daily Californian reported a sharp rise in failing grades in computer science courses in spring 2026. In CS 10, 35.3% of students got an F, in CS 61A 10.6%, in EECS 127 16.8%. A year and two years earlier, the F rate in CS 10 and CS 61A stayed below 10%.
Dan Garcia, who taught both courses, called a surge in dishonest use of language models the main cause. His colleagues added weak math preparation and a shortage of graders.
What people are saying
On Hacker News the debate is not about whether Borcherds is right but about what should replace homework grades. One camp says grades should come only from in-class exams where AI is out of reach: homework stays practice, not a test. They also point out how much grading costs: grad students spend 10-20 hours a week on it.
Others push back: a homework grade keeps students who need outside structure on track. And a single final exam turns the semester into a lottery: one bad night's sleep and you fail the course.
What follows from this
The professor stopped checking not the homework itself but whether the answers are right. The assignments remain, and students still have to hand them in. What changed is that a correct answer no longer proves the student understood the material.
The catch is that students themselves will pay for it. Homework has always been practice before the exam. If a chatbot solves the problems during practice, it will not help in the exam room - the spring F's in CS 10 showed that clearly.
If you are studying or helping your kids learn, the takeaway is simple: AI works as a tutor, not as an answer key. Ask the model to explain the steps, find the mistake in your own attempt, or give you a similar problem to check yourself. A ready answer copied into a notebook teaches nothing - and judging by Berkeley, grades will soon be given where copying is not an option.
Sources
- AI use leads UC Berkeley math professor to see 'no point' in grading homework - The Daily Californian: interview with Borcherds and Berkeley's AI policies, October 23, 2026
- 2026 Fall MATH 250A 001 LEC 001 - Borcherds' course listing in the Berkeley class schedule
- Navier-Stokes solution - OpenAI's announcement of the proof, September 8, 2026