Richard Borcherds Stops Grading Homework at UC Berkeley as AI Makes Correctness Checks Pointless

Richard Borcherds, a UC Berkeley mathematics professor and Fields medalist, no longer grades his students' homework for correctness. After seeing how many of his students used AI to consistently receive top scores, he concluded there was "no point" in paying readers to grade homework for correctness, because AI use in his courses is now unavoidable.

What Richard Borcherds changed in his UC Berkeley math courses

Borcherds now grades homework on completion: a student who hands in the assignment receives credit. He said the arrangement is "not a very good system," but that at the moment he has no alternative.

The change was driven by measurement failure rather than by a campus rule. Because AI-assisted submissions reliably produced top scores, paying human readers to check correctness stopped generating useful information about what students had actually learned.

Borcherds says there will soon be nothing human mathematicians can do better than AI

"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," Borcherds said.

The professor said he no longer knows whether to encourage students to go into the math department. Five years earlier, he said, it seemed that being a mathematician was one of the safest careers from the AI boom. That assessment no longer holds.

"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 said.

Borcherds also predicted that institutions will lag the technology. He described a plentiful amount of "inertia" in the academic system and said he would not be surprised if things stayed the same in the math department for a couple of years despite AI developments. He compared the situation to physical paper journals, which continue to publish even though he considers them "utterly pointless" for two decades.

He framed the moment as historically unusual:

"There have been plenty of technical breakthroughs in the past — hand weaving was put out of business by looms, and horses were put out of business by motor cars, and so on. This is the first time anything like this has happened in mathematics, though."

What UC Berkeley policy actually says about AI in coursework

UC Berkeley does not impose a campus-wide rule requiring or banning AI use in coursework; AI policy for a class is left to the discretion of the faculty member teaching it.

The Berkeley Division of the Academic Senate asks that faculty include a statement in their syllabi about course expectations, including the use of AI tools for coursework, according to Anne Marie Richard, assistant vice provost for undergraduate education and chief academic technology officer. Richard said that direct policies on AI use can vary dramatically with course level and focus area.

UC Berkeley has tried several AI detection tools, including Turnitin and Rumi, but has determined that these tools are not adequate enough to be standardized on campus. Instead, the campus maintains a growing database of sample AI use statements in the AI Hub, the campus directory for AI tools, training and policies, for faculty to use in their syllabi.

The trigger for the interview: OpenAI's claimed Navier-Stokes result

Borcherds revealed his grading change in a sit-down interview following OpenAI's announcement that AI agents apparently solved the Navier-Stokes existence and smoothness problem, one of mathematics' most difficult unsolved problems. The Daily Californian's own coverage of that announcement used the word "apparently," and a companion Daily Californian article described the UC Berkeley math community as left at a crossroads in the wake of the claimed proof.

The Hacker News discussion: grading homework was already broken

The Hacker News thread on the story (176 points, 257 comments) split between readers who saw the change as an overdue correction to a transactional grading model and readers who saw it as a real loss for student learning.

Homework points were always a transaction, and AI exposed the cost

The most-upvoted line of argument was that homework grading was a weak measurement long before AI. Commenter parsimo2010 noted an irony: Borcherds did not grade homework himself in the first place, he paid assistants to grade it — common practice for tenured professors with grant money and graduate students.

"You have to figure out a new transactional structure that can't have your tuition made irrelevant by a $20 subscription."

That commenter's proposed fix was to stop grading homework entirely and put 100 percent of the grade weight into proctored assessments with some kind of firewall against AI tools, arguing that any student disadvantaged by that change was probably also disadvantaged by classmates using premium AI subscriptions on take-home work.

The case for exams, vivas, and in-person assessment

Several commenters argued that the reliable fix is to move assessment into supervised settings:

jruohonen recommended in-class exams, either written or oral, with the point that failing to do homework then reflects only on the student.
grey-area proposed 0 marks for homework, all marks on the exam and a viva, with feedback still available but explicitly not counting toward grades.
remywang, who teaches, said in-person quizzes motivate students to do the homework anyway because it prepares them for the quiz.
oldsklgdfth recalled a 2006 high school physics teacher who gave credit simply for handing in homework without assigning it a grade, and noted that at a school like Berkeley, "the professor have office hours for me to ask questions."

Multiple commenters noted this problem predates AI

VyseofArcadia said there was already no point in grading homework at the lower undergraduate level a decade ago because Wolfram Alpha existed, and suggested that graduate-level courses may be handled by Moore-style classes where students present at the chalkboard rather than with slides while the professor critiques.
jasonjmcghee said many professors at Berkeley did not use homework as part of grades when they attended, with midterms and finals worth 100 percent of the grade.
TrackerFF described the Norwegian model: no marks or points for homework at all, a final grade based entirely on a pen-and-paper final exam or a mix of exam and projects, and mandatory homework used only to qualify for the final exam. As a teaching assistant, that commenter saw obvious cheating — identical step-by-step solutions with identical errors — but noted that cheating on homework could not help a grade.
bubblethink argued there is no point in grading homework or essays, that India and China already rely on standardized tests and in-person exams, and that universities are proxies for IQ filtering.

The counterpoint: AI also grades, and it is available to everyone

Commenter peri-cl argued the same technology that breaks take-home grading also democratizes feedback:

"AI's can now grade pure math problems with the accuracy and rigor of a postdoc. And unlike PhD tutors, they're universally accessible to all. It's the start of a golden age for those of us who are curious autodidacts."

Under that view, a student working through a math textbook can submit an attempted theorem and proof to a model and see misconceptions and lapses in rigor exposed immediately.

Skepticism about the professor's framing

Not every commenter accepted the framing. Commenter tzmake claimed that Borcherds has been working with Epoch AI since at least 2024 and is issuing standard talking points about looms and horses while claiming he could not have foreseen the situation, calling him "an AI shill, just like Tao." That is a commenter allegation in the thread and is not corroborated by the Daily Californian reporting.

The learning-loss argument

Other commenters focused on what disappears when the homework signal does:

vhiremath4 argued that even peer-to-peer cheating required someone to do the homework, which acted as a forcing function, and that AI removes it entirely. The commenter said a cohort of students, especially 18-20 year olds, will not do well on exams because they need "a bit of a stick."
analog31 put the shift in one line: "We don't yet know if AI has replaced mathematicians. But if AI can do homework and exam problems, then it means AI has replaced students."
bananaflag said they never graded homework or anything except the final exam, so students had no incentive to use AI, and offered the classroom line: "AI can replace me teaching you, it cannot replace you understanding the material."
thenoblesunfish argued that people will remain intrinsically interested in mathematics and in understanding something well enough to explain it, so the task is to find new ways to assess students rather than to rethink the purpose of education.
p0w3n3d suggested a low-tech check: ask students directly about their homework and penalize those who cannot answer questions about it.
arjie asked what the point of grading homework for college students is at all, and proposed splitting the single number into two certified grades — conscientiousness measured by attendance and mastery measured by exam.
proc0 suggested higher education should return to being about academia rather than industry skills training, since companies can now run their own extended training and hire people out of high school.
mkl95, an online CS student, said higher education as we know it faces extinction in less than a decade and argued that only truly elite universities that educate efficiently deserve to survive.

Who Richard Borcherds is

Richard Borcherds is a UC Berkeley mathematics professor and a Fields medalist. According to HN commenter seanhunter, he is best known for proving the "Monstrous Moonshine" conjecture of Conway and Norton — a correspondence between modular functions and representations of the Monster Group. When the conjecture was made in the 1970s it rested on the observation that a certain number matched the cardinality of the Monster Group, an apparent coincidence that John H. Conway playfully called "Moonshine," which is where the conjecture's name comes from. Borcherds also runs a YouTube channel with lectures on number theory, category theory and other topics.

Key takeaways

Richard Borcherds, a UC Berkeley math professor and Fields medalist, stopped grading homework for correctness because AI-assisted submissions consistently earned top scores; he now gives credit for completion.
Borcherds called completion-based grading "not a very good system" and said he currently has no alternative.
He said that if AI keeps improving at its present rate, there will soon be nothing human mathematicians can do better than AI, and that he does not know what career advice to give students.
UC Berkeley sets no campus-wide AI rule for coursework; the Academic Senate asks faculty to state AI expectations in syllabi, detection tools such as Turnitin and Rumi were judged inadequate for campus-wide standardization, and faculty can draw sample AI statements from the campus AI Hub.
The interview followed OpenAI's announcement that AI agents apparently solved the Navier-Stokes existence and smoothness problem.
Hacker News commenters divided between those who said homework grading was always a weak, transactional measure that AI simply exposed, those who proposed moving all weight to proctored exams and vivas, and those who argued that removing the homework incentive will cost a cohort of students who rely on incremental deadlines to learn.