---
title: "Caltech Stops Grading Math Homework for Accuracy Because AI Solves It"
description: "A 17-year Caltech professor and a Fields medalist at Berkeley reach the same conclusion independently"
author: "Luis Goa"
published: 2026-11-16T11:12:01Z
modified: 2026-11-16T13:55:21Z
url: https://rews.cc/a/caltech-stops-grading-math-homework-for-accuracy-because-ai--4f55ab
language: en
tags: ["education", "ai", "caltech", "homework", "students", "us"]
publisher: "Rews (https://rews.cc)"
---

# Caltech Stops Grading Math Homework for Accuracy Because AI Solves It

*A 17-year Caltech professor and a Fields medalist at Berkeley reach the same conclusion independently*

By Luis Goa · November 16, 2026 · https://rews.cc/a/caltech-stops-grading-math-homework-for-accuracy-because-ai--4f55ab

## In brief

- Caltech’s Yi Ni now grades math homework only on completion, saying AI can complete nearly all college math problem sets
- UC Berkeley Fields medalist Richard Borcherds made the identical change independently, calling correctness-grading pointless
- Epoch AI’s FrontierMath Tier 4 leaderboard shows top AI scores rising from under 2% in 2024 to 100% in November 2026 on a revised problem set
- OpenAI and Google DeepMind models reached IMO gold-medal-equivalent scores (35/42) in July 2025, versus silver in 2024
- An MIT survey of 268 early-career mathematicians found 76% feel less motivated to do research because of AI

Yi Ni has taught math at Caltech for 17 years. This term he stopped grading homework on whether the answers are right. Students still have to turn problem sets in, but the grade now records only completion, because Ni cannot tell which answers came from a student and which came from a chatbot, as he told Business Insider’s Truman Dickerson after [Dickerson sat in on one of Ni’s lectures](https://www.businessinsider.com/ai-math-class-university-homework-caltech-careers-2026-11).

“There is a problem in the education process,” Ni said. “Virtually all the homework in college math, you can use AI.” The class Dickerson watched was entirely chalk and blackboard, eight boards filled with formulas over the course of the hour, no laptop, no projector. The analog format didn’t matter: the homework students take home from it can be solved by software whether or not the lecture used any.

Ni isn’t alone, and he isn’t first. Richard Borcherds, a Fields medalist (mathematics’ closest equivalent to a Nobel, awarded only every four years) who has taught at UC Berkeley since 1993, told the student paper *The Daily Californian* in late October that he made the identical change for the identical reason. “AI can now solve many undergraduate mathematics problems, often better than the students themselves,” Borcherds said, as [Newsweek reported](https://www.newsweek.com/berkeley-professor-grading-homework-ai-12619345). Grading for correctness, he said, “adds a huge incentive for them to use AI for homework instead of doing it themselves, and penalizes students for honesty.” He now grades on completion too, and says exams given in person, without computers, are the only assessment he’s confident still measures a student rather than a model.

Both professors land on the same fallback for the same reason: in-person, closed-device exams are the one format AI can’t reach. Everything that leaves the classroom is now presumptively contaminated.

![A robotic hand holds a pencil against a glass dome enclosing a wooden school desk and chair, with blank paper and another pencil on the desktop.](https://rews.cc/img/c3e764a25bf11463cdc3a1bae234072c2ab48e4c.webp)

## Why math broke before writing did

Ni drew a specific contrast with AI-generated prose, which professors across disciplines have learned to flag by its tics: short declarative sentences strung together, overused em dashes, a habit of posing a question only to answer it. Math has no equivalent tell. A correct derivation looks the same whether a sophomore or a language model produced it, and a problem set is graded on whether the final answer and the steps are valid, not on style. The channel detectors use for prose, flagging statistical regularities in word choice, doesn’t exist for a sequence of algebraic manipulations that either follows the rules or doesn’t.

That wasn’t always true. When Epoch AI released its FrontierMath benchmark in 2024, a set of several hundred problems written by professional mathematicians specifically to be hard for AI, top models scored under 2%. The hardest tier, Tier 4, is meant to approximate research-level work: the kind of problem that takes a working mathematician hours or days. Progress since then has been close to a straight line up. GPT-5 reached roughly 25% on the full benchmark in 2025. By November 2026, Epoch AI’s leaderboard for the corrected Tier 4 set shows OpenAI’s GPT-6.1 Sol at 100% and Anthropic’s Claude Opus 5.5 at 95%, according to [Epoch AI’s published leaderboard](https://epoch.ai/frontiermath/tiers-1-4). Epoch cautions that its 2026 revision fixed errors in 42% of the original problems, and that scores on the corrected set run about 12 percentage points higher than the same models scored on the flawed original, so the two eras aren’t directly comparable. The direction, not any single number, is what Ni and Borcherds are reacting to.

![A massive pyramid made of wooden abacus beads and rods rises in a rocky desert, topped by a computer chip beneath a blue sky.](https://rews.cc/img/1303dc3c82e98666bbd219bdd0bf00e7b770ee32.webp)

The same climb shows up at the high-school end of the difficulty range. In July 2025, both OpenAI and Google DeepMind announced that general-purpose reasoning models had scored 35 out of 42 points on the 2025 International Mathematical Olympiad, the threshold for a gold medal and the highest cutoff in the competition’s history. The IMO itself certified Google DeepMind’s Gemini Deep Think result; OpenAI’s was graded by three former IMO medalists the company engaged itself, not by the competition, as [Google DeepMind’s statement](https://deepmind.google/blog/advanced-version-of-gemini-with-deep-think-officially-achieves-gold-medal-standard-at-the-international-mathematical-olympiad/) and contemporaneous reporting by [TechCrunch](https://techcrunch.com/2025/07/21/openai-and-google-outdo-the-mathletes-but-not-each-other) and [Axios](https://www.axios.com/2025/07/21/openai-deepmind-math-olympiad-ai) recorded. Only about 8% of human IMO competitors earn gold, Google DeepMind noted. A year earlier, DeepMind’s AlphaProof had managed only silver.

Then there’s the open-problem claim that neither Ni nor Borcherds could have ignored even if they wanted to. In September 2026, OpenAI said an internal, unreleased model, running roughly 10,000 coordinating AI agents that exchanged 2.7 million messages over 88 hours, produced a proof addressing part of the Navier-Stokes existence and smoothness problem, one of the Clay Mathematics Institute’s seven Millennium Prize problems, each carrying a $1 million award, as [CNBC reported](https://www.cnbc.com/2026/09/09/openai-navier-stokes-math-problem-solved.html). The Clay Institute changed the problem’s status on its website to “active” rather than solved or unsolved, and said its verification process would be “deliberately unhurried.” Mathematicians surveyed by Scientific American in October found that OpenAI’s approach diverged from an independent proof track already underway by other researchers, and the claim remains formally unverified. Borcherds brought it up unprompted in his own interview as the moment that reordered his sense of how much time mathematicians as a profession had left.

![A giant metal drawing compass traces a smooth circle on the ocean beside a sandy beach with an empty wooden chair under a cloudy blue sky.](https://rews.cc/img/c2e8782127752cce93a704174f5427a1240b3f38.webp)

That unease isn’t confined to two professors. A survey of 268 early-career mathematicians (graduate students and postdocs across roughly 60 institutions), [published on arXiv in late October](https://arxiv.org/abs/2610.40584) by three MIT math researchers, found 76% had grown less motivated to do research because of AI developments, and a similar share said they’d grown less likely to stay in academic math after finishing their degrees, a finding [rews covered earlier this month](https://rews.cc/a/76-of-young-mathematicians-in-mit-survey-say-ai-has-sapped-t-34c554). Among respondents who do use AI in their own research, 54.8% said they disliked having to.

## No policy, just workarounds

Caltech has no institution-wide rule for how math courses should handle AI; the choice is left to each professor. Sathvik Redrouthu, a sophomore math major who took Ni’s class last year, told Business Insider he avoided using AI on homework because doing so would be “shooting yourself in the foot” once an in-person exam arrived, and that a strict no-AI norm “pushes the average student to learn properly.” He also said this year’s syllabuses at Caltech increasingly carry an explicit label, “AI allowed,” rather than leaving the question unaddressed. Borcherds, for his part, told Newsweek he doesn’t know of a good answer for fully online courses, where there’s no exam room to retreat to.

Neither professor has proposed what replaces homework as a signal of whether a student understands the material, beyond falling back on proctored exams. The open question both are circling is the same one the MIT survey’s respondents are living with from the other side of the desk: if a model can already match or beat a graduate student on math problems built to be hard for exactly that model, what is the homework, or the degree, actually credentialing.

## See also

- [FrontierMath (Wikipedia)](https://en.wikipedia.org/wiki/FrontierMath) — en.wikipedia.org · Background on the benchmark, its funding and its 2024 controversy
