Education

Some Professors Stop Grading Homework for Accuracy as Students Turn to AI

Instructors are moving real grading to exams, in-class work and oral checks, but no study yet shows which changes help students learn.

12 min read
Man in a grey blazer and jeans holding a worn book and a brown leather satchel on a gravel path.
Image by Pexels from Pixabay

A Berkeley mathematician and a Caltech professor say they no longer grade homework for correctness because AI can finish it. Survey data show teen chatbot use for schoolwork is common but far from universal, and evidence that AI caused a cheating surge is thin and dated. Families face rules that differ by class.

Two college math professors have stopped treating the accuracy of homework as something worth grading. Their reason reaches well beyond mathematics: when AI can produce correct answers on demand, a correct answer no longer shows that a student learned anything. According to news reports from late October and mid-November 2026, Richard Borcherds at UC Berkeley and Yi Ni at Caltech have both moved away from grading homework for correctness.

Two professors are a small sample, but their choice sits inside a broader shift. Surveys show that a large share of teenagers use chatbots for schoolwork, that many schools lack clear AI rules, and that teachers are experimenting with in-class writing, oral exams and completion credit. What nobody has shown yet is which of those responses works. This article separates what is documented from what is anecdote, and it describes what the evidence suggests for parents and students.

What the Berkeley and Caltech professors actually did

Borcherds, a mathematician who won the Fields Medal in 1998, now grades homework on completion rather than correctness, according to a news report syndicated by Yahoo. He told reporters he had seen students use AI to earn top homework scores. His stated reasoning is that grading for correctness gives students a large incentive to use AI and penalizes classmates who do the work honestly.

The weight of the course has moved elsewhere. Borcherds says final grades now rest mainly on in-person exams. He considers online exams a waste of time because some students use AI on them, according to the same report.

At Caltech, Ni, who has taught there for 17 years, described a similar change in a first-person account published on Yahoo Tech. He no longer grades homework on accuracy. "Virtually all the homework in college math, you can use AI," he said. The same account reports that Caltech has no unified guidance for math professors and that some classes this year carry an "AI allowed" caveat.

A Caltech student who avoided AI offered a practical argument for doing so. Using it on homework, the student said, is like "shooting yourself in the foot" once an in-person exam arrives. That logic is the engine of the whole arrangement: if the exam carries the grade, homework becomes practice whose value depends on whether the student does it.

What these accounts do not show

The headline version of the story, that "everyone" is using AI, is a professor's belief and not a measurement. Ni said he cannot tell whether a given problem set was done with AI. Both professors teach college math, a field where AI is especially capable, so their experience does not describe what happens in a typical middle or high school.

The sourcing also needs care. Much of the Berkeley coverage online traces back to a small number of original interviews, so repeated copies do not confirm one another. UC Berkeley requires faculty to address AI use in their courses but leaves each instructor in charge of the policy, which means two classes in the same department can follow different rules.

Not every instructor is tightening control. Kyle Saunders of Colorado State decided to encourage AI use on some assignments, arguing that a prohibition does not actually produce a class that isn't using AI, according to The College Fix. The claim is his own, relayed secondhand, but it shows the range of responses.

How many students use AI for schoolwork

The best national figures for teenagers come from the Pew Research Center, which surveyed 1,458 US teens and their parents from Sept. 25 to Oct. 9, 2025, and published results on Feb. 24, 2026. Pew found that 54% of teens had used chatbots to get help with schoolwork. About one in ten said they do all or most of their schoolwork with a chatbot's help.

Teen chatbot use, Pew Research Center survey of 1,458 US teens, Sept. 25 to Oct. 9, 2025
Measure Share of teens
Used chatbots to search for information 57%
Used chatbots for schoolwork help 54%
Casual conversation 16%
Emotional support or advice 12%
Do all or most schoolwork with chatbot help about 10%
Chatbots extremely or very helpful for schoolwork about 25%

Heavy reliance is therefore a minority pattern, and the professors' impression of near-universal use in college math is not contradicted by it. The groups differ in age, year and question wording. Pew's separate demographic analysis found that about six in ten Black or Hispanic teens used chatbots for schoolwork, compared with roughly half of White teens. Teens in households earning under $75,000 were more likely to use chatbots for all or most of their schoolwork.

What students think is acceptable

Teens draw lines between tasks. An earlier Pew survey, reported by K-12 Dive, found that 54% considered ChatGPT acceptable for researching new topics, but only 18% for writing essays. For math problems, opinion split almost evenly, with 29% calling it acceptable and 28% not.

A March 17, 2026 RAND report on 2025 surveys found student use of AI for homework rising. Only getting direct answers to homework was widely seen as cheating, at 45%. Most students did not consider other common uses to be cheating, and RAND found that many schools lack clear AI rules. That gap helps explain why some students and some instructors may define the same act differently.

The same RAND report found that 67% of students said AI use harms critical thinking, up from 54% earlier in 2025. These are opinions and not measured effects on learning. An earlier RAND analysis of nationally representative panels found student and teacher AI use rising by 15 percentage points or more over the previous one to two years, with training for both groups lagging and students worried about being falsely accused of cheating.

Did AI cause a surge in cheating?

The assumption that ChatGPT set off a cheating wave is harder to support than it sounds. Stanford researchers working with Challenge Success surveyed students anonymously at 40 US high schools. According to Stanford's Graduate School of Education, 60 to 70 percent of students reported at least one cheating behavior in the previous month. That share stayed about the same, or dipped slightly, after ChatGPT arrived.

Students in the Stanford surveys gave familiar reasons: struggling to grasp the material, too little time, and pressure to perform. Stanford's framing is that cheating is generally a symptom of a deeper problem, and that students cheat less when they feel a sense of belonging and find purpose in their classes. Researcher Victor Lee cautioned, as CNN reported, that schools' approach to AI "could move things in either direction."

What teachers are changing

Moving the evidence into the room

Reuters interviewed about a dozen US middle and high school teachers and experts, in a piece republished by the Express Tribune. Teachers described assigning less take-home work, doing more writing and discussion in class, and redesigning assignments to measure thinking and not just answers. One San Francisco teacher dropped take-home essays in favor of oral exams. The sample is small and not random, and Reuters noted that school systems remain divided over AI, leaving many teachers to work it out largely alone.

College instructors report a similar retreat to older methods. GovTech described Syracuse-area professors returning to pre-technology assessments and in-class essays. NYU's Clay Shirky argues for replacing take-home essays with in-class blue-book essays, oral exams and required office hours, as Yahoo reported. Historian David Bell wrote on Substack that he replaced a short paper with three in-class blue-book exercises and found the course improved, because students had to review the material.

The diagnosis behind the changes

Many of these changes rest on a problem that is not quite cheating. A survey of K-12 educators, described in an essay on Phys.org that was originally published by The Conversation, concluded that the central issue is that AI-generated work obscures evidence of individual learning. The survey is not nationally representative. Its suggested responses include clearer AI-use policies, assessments that focus on process or happen in class, and tasks aligned with learning outcomes.

UK evidence points the same way. Jisc, a British education technology body, wrote in 2025 that in the study it cited, the 17 assessment types tested generally were not robust against generative AI. Either AI-written answers passed, or training markers to spot AI raised false positives. The Jisc analysis is British and covers higher education, so it does not map directly onto American classrooms.

The tradeoffs

Each fix carries costs that its advocates acknowledge. Shirky notes that an emphasis on in-class performance favors some students over others. Bell points out that individual oral exams take an enormous amount of time in a large class. A physics faculty preprint on arXiv, which has not been peer reviewed, calls in-class supervised testing a common default but says the historical concerns about testing remain.

Completion grading raises a separate question. If homework earns credit simply for being done, students have less grade-based reason to put careful effort into it. The coverage does not show anyone having resolved that tension, and the Berkeley reports do not claim Borcherds has. The accounts of what works are teachers' experience and not tested evidence.

Why AI detectors have not settled the question

Software that claims to identify AI writing looks like a shortcut, but reported error rates vary enormously. A 2023 study in the journal Patterns by Stanford researchers Liang and colleagues tested seven detectors on 91 TOEFL essays written by non-native English speakers. On average the detectors wrongly flagged 61% of this genuine human writing as AI-generated. The figures come through a secondary summary on HEPI, a UK higher-education think tank, and not from the original paper.

Other results look far better. HEPI also reports a study of 160 genuine student assignments that found no false positives, and it concludes there is no universal accuracy figure. Turnitin says it launched with a false-positive rate of about 1%. Jisc cites work by Hyatt and colleagues on STEM student writing that found a 1.3% false-positive rate for detectors and 5% for human raters, so people do not clearly outperform the software.

Even a low rate matters at scale. Jisc calculated that a 1% false-positive rate would produce about 4,800 false positives a year at an institution with 20,000 students. The estimates cannot be compared directly, since each used a different design, population and kind of text. The range from 0% to above 60% suggests detectors are a weak basis for an accusation on their own. The risk appears to fall hardest on non-native English writers.

What parents and students can do

No trial has tested family strategies for homework in the AI era. The points below come from teachers' and students' own accounts and from survey findings, and they are context and not individualized advice.

  1. Find out each class's rule. Berkeley and Caltech leave AI policies to individual instructors, and Caltech has no unified guidance for math faculty, so one course may allow AI while the next bans it. RAND found that many schools lack clear rules and that students are unsure what counts as cheating.
  2. Treat homework as practice when exams carry the grade. The Caltech student's reasoning, and the structure of Borcherds' course, both assume that work done with AI will show up as a weakness on in-person exams.
  3. Talk about AI use at home. A CBS News report on the Pew survey carried the headline that teens use AI frequently and many parents are not aware of it.
  4. Learn the process for disputed AI flags. Given detector error rates and students' stated fear of false accusations, knowing how a school handles a dispute is a reasonable precaution.

Stanford's finding that students cheat less when they feel belonging and find purpose in class suggests a further angle. If cheating is partly a symptom of struggle, time pressure and stress, then conversations about workload and understanding may matter as much as policing tools. That is Stanford's interpretation of its survey, not a tested intervention.

What to watch next

The professors' decisions mark a change in what homework is for. When correctness can be produced by a chatbot, graded homework stops measuring skill and becomes either a temptation or a rehearsal, and the grade moves to settings where a teacher can see the thinking. That is a reasonable adaptation, but so far it is an adaptation built on instinct and experience.

Three questions will decide how durable it proves: whether students who skip or shortcut ungraded practice actually do worse on supervised exams, whether in-class and oral formats can be run fairly and affordably in crowded classrooms, and whether schools write rules clear enough that students know what is allowed. Until research catches up with practice, the most reliable guide for families is the specific rule in each specific class.

Frequently asked questions

Why did a Berkeley professor stop grading homework for accuracy?

Richard Borcherds, a 1998 Fields Medal winner, said he saw students use AI to earn top homework scores. He believes grading for correctness gives students a large incentive to use AI and penalizes honest classmates. His course now grades homework on completion, and final grades rely mainly on in-person exams.

How many teens use AI chatbots for schoolwork?

In a Pew Research Center survey of 1,458 US teens conducted Sept. 25 to Oct. 9, 2025, 54% said they had used chatbots to get help with schoolwork. About one in ten said they do all or most of their schoolwork with chatbot help. Use has probably risen since the survey was taken.

Did ChatGPT cause more cheating in high school?

Stanford surveys at 40 US high schools found that 60 to 70 percent of students reported at least one cheating behavior in the previous month, and that share stayed about the same or dipped slightly after ChatGPT arrived. The data are self-reported and run only through fall 2023, so they do not describe the present.

Are AI detectors reliable enough to accuse a student?

Reported error rates vary widely, from 0% in one study of 160 genuine assignments to an average of 61% for non-native English essays in a 2023 Stanford study. Even a 1% false-positive rate could produce about 4,800 false flags a year at a 20,000-student institution, according to Jisc. The evidence suggests detector scores alone are a weak basis for accusations.

What are teachers doing instead of take-home assignments?

Teachers and professors describe in-class writing, blue-book exams, oral exams, required office hours and redesigned assignments that focus on process. Each has costs, including instructor time and possible advantages for some students. No study yet shows which approach produces better learning.

Sources (19)

  1. UC Berkeley professor refuses to grade his students' homework after noticing worrying trend: 'No point' , Yahoo News
  2. I went to a math class at Caltech. The professor no longer grades homework because everyone's using AI. , Yahoo Tech
  3. UC Berkeley math professor no longer grades homework over students' AI use , The College Fix
  4. How Teens Use and View AI , Pew Research Center (2026-02-24)
  5. Demographic differences in how teens use and view AI , Pew Research Center (2026-02-24)
  6. Teens are using AI frequently in their daily lives, and many parents aren't aware, survey finds , CBS News (2026-02-24)
  7. Double the teens using ChatGPT for schoolwork , K-12 Dive
  8. Student Use of AI for Homework Rises as Concerns Grow About Critical Thinking Skills , RAND (2026-03-17)
  9. AI Use in Schools Is Quickly Increasing but Guidance Lags Behind: Findings from the RAND Survey Panels , RAND
  10. What do AI chatbots really mean for students and cheating? , Stanford Graduate School of Education
  11. ChatGPT did not increase cheating in high schools, Stanford researchers find , CNN (2023-12-13)
  12. AI detectors and the fairness gap: read the evidence before you trust the score , Higher Education Policy Institute (HEPI) (2026-08-27)
  13. AI Detection and assessment - an update for 2025 , Jisc National Centre for AI (2025-06-24)
  14. How teachers are remaking classrooms for the AI age , Reuters, via The Express Tribune
  15. AI Nudges Syracuse Professors Back Toward Blue Books, In-Class Work , GovTech
  16. Colleges should go 'medieval' on students to beat AI cheating, NYU official says , Yahoo News
  17. Beating AI in the Classroom , David A. Bell, Substack
  18. Teachers are worried about students cheating with AI, but my survey suggests the deeper issue is learning , Phys.org (from The Conversation) (2026-07-01)
  19. A Workshop Series for Effective Use of AI in Uncertain Times: Building a Physics Faculty Learning Community , arXiv (preprint, not peer reviewed)

Share this article

Back to list