Richard Borcherds Homework Grading Changes as AI Blurs Evidence of Learning

Richard Borcherds has reportedly stopped grading homework for correctness because AI makes independent student work difficult to identify. The Richard Borcherds homework grading change leaves practice assignments in place while shifting evidence of understanding toward supervised exams. That distinction matters: the reported retreat concerns what homework scores can establish, rather than whether students should practice mathematics.

According to Tom’s Hardware’s October 27 coverage, the UC Berkeley professor now primarily assesses students through in-person exams without electronic devices. Other coverage describes homework credit based on completion. Together, those reports describe a separation between practicing a skill and demonstrating that the student personally possesses it.

The competing approaches already exist within Berkeley. Its fall 2026 statistical computing syllabus permits AI assistance while pairing assignments with in-person checks of understanding. Borcherds’ reported decision therefore raises a specific question for educators and AI users: when software can produce the answer, what should count as evidence that someone understands it?

What the Richard Borcherds homework grading change actually means

The reported change removes correctness as the basis for homework credit; it does not establish that Borcherds has abolished homework.

Tom’s Hardware attributes the decision to Borcherds’ concern that polished submissions no longer reliably distinguish independent work from AI assistance. Its account of his homework assessment change also describes a preference for supervised exams. He reportedly sees higher online scores without being able to identify which students used AI.

The completion-credit detail appears in other coverage, including a summary linked to the New York Post. That account describes completion-based homework credit, narrowing what readers should take from headlines about ending graded homework. Submission can still matter even when correct answers no longer earn differentiated marks.

Borcherds’ standing helps explain the attention. Berkeley’s archive documents his 1998 Fields Medal, awarded for mathematical work including his proof of the Monstrous Moonshine conjecture. That distinction establishes his research credentials. It does not independently validate his assessment policy or measure cheating in his classes.

Several boundaries remain important. The reporting reviewed here does not establish a campus-wide withdrawal from graded homework. Nor does it provide a verified implementation date, a complete grading formula, or a measured cheating rate for his students.

The immediate development is therefore an instructor’s reported response to a problem he says he can no longer resolve through inspecting answers. Treating that response as a Berkeley-wide policy would overstate the evidence. Treating it as proof that every strong submission involves cheating would overstate it further.

The distinction also changes the policy debate. Ending practice assignments would remove an opportunity to learn. Ending correctness-based credit changes the incentive attached to that opportunity, while leaving open how students receive useful feedback.

Consider the difference from a student’s perspective. Under correctness grading, an unresolved problem threatens the assignment score. Under completion credit, the same unresolved problem can reveal something to address before an exam, assuming the course still provides adequate feedback.

That is the potential benefit, rather than an established outcome from Borcherds’ classes. Students might feel less pressure to submit flawless answers they cannot explain. They might also devote less effort to practice if completion feels sufficient.

The Richard Borcherds homework grading story consequently turns on an assessment problem with competing consequences. A course can make homework less attractive to outsource while also weakening the immediate reward for doing it carefully. Whether the balance improves learning requires evidence beyond the announcement itself.

Correct answers are losing their value as evidence

A correct submission establishes something about the submitted work, but increasingly leaves the student’s independent ability unresolved.

Homework has several possible jobs. It gives students practice, helps instructors identify confusion, creates deadlines, and contributes evidence toward a course grade. Those purposes overlap, but they do not always require the same assessment method.

AI makes the separation harder to ignore. A student can possess an accurate solution without having selected its method, understood its assumptions, or recognized where it might fail. The final answer alone cannot resolve those differences.

This does not mean every use of assistance defeats learning. A hint can help someone resume productive work. An explanation can expose a misunderstanding. A completed solution can also replace the reasoning that an assignment was intended to develop.

From the finished submission, these paths can look similar. That creates pressure on instructors who must assign individual credit from work completed outside their observation. It also pressures students who follow restrictions while believing others receive credit through prohibited assistance.

Borcherds’ reported response addresses that incentive directly. If correctness earns fewer homework points, obtaining a flawless answer from software becomes less useful for that particular score. However, completion credit still rewards handing something in, so the incentive does not disappear entirely.

The deeper change concerns the relationship between practice and certification. Practice can involve help, failed attempts, discussion, and revision. Certification requires a defensible basis for saying what the individual can do under specified conditions.

A hypothetical mathematics assignment illustrates the difference. A student submits a valid proof but cannot explain why its central assumption is necessary. The proof remains valid, while the submission provides weak evidence that the student understands its logic.

An examiner who changes that assumption can ask a different question: does the student recognize what breaks? That exchange examines understanding more directly than another inspection of the original document. It also takes time and requires a suitable question.

The same distinction matters outside universities. A generated program might run successfully while its user cannot diagnose a failure. A polished analysis might sound persuasive while its author cannot defend the evidence behind its recommendation.

These are practical implications, rather than findings about Borcherds’ students. They explain why knowledge workers should care about the assessment dispute. Faster production increases the importance of deciding which human judgments still need verification.

For educators, the difficult choice concerns what each assignment should establish. If its main purpose is practice, assistance might fit the learning goal. If it certifies independent competence, access to assistance changes what the resulting score means.

The problem therefore extends beyond catching misconduct after submission. An assessment can become less informative even when nobody has violated a rule, especially if the rules permit substantial assistance. Clear permission does not automatically make a score a valid measure of individual understanding.

Supervised exams restore observation, with limits

Supervised exams strengthen evidence about who performed the work, while narrowing the conditions under which students demonstrate their skills.

An in-person exam without electronic devices addresses a specific uncertainty. The instructor can observe a student attempting problems without access to the tools excluded by the exam rules. That makes authorship easier to establish than with an unsupervised submission.

It does not automatically establish everything a course values. A timed examination measures performance within its particular constraints. Extended investigation, revision, collaboration, and effective use of reference material can require different opportunities.

Research provides a useful historical reference for the concern about unsupervised assessment. A University of Reading experiment, publicized in June 2024, inserted AI-generated answers into undergraduate psychology assessments. Its blind assessment test found that 94% went undetected and that AI submissions received higher grades on average.

The result supports a bounded conclusion: experienced markers can struggle to recognize generated work within an existing assessment system. It does not measure cheating at Berkeley, establish a mathematics-specific detection rate, or prove that supervised exams produce better learning.

That boundary matters because detection and education are different outcomes. An assessment might make unauthorized assistance difficult without giving students effective preparation. Another might support learning while producing a score that cannot confidently certify independent performance.

Borcherds’ reported preference resolves the observation problem more directly than the learning problem. It changes where instructors collect evidence. Whether students become better mathematicians depends on what happens before and after that assessment.

There is also a difference between introducing regular supervised checks and concentrating most credit in a final examination. The former can reveal confusion while students still have time to respond. The latter can make misunderstandings visible only when the consequences are substantial.

Neither arrangement follows automatically from the phrase “in-person exams.” Readers need the actual assessment schedule, weighting, feedback process, and accommodation arrangements to evaluate a course design. The reporting does not supply all those details for Borcherds.

A supervised test can also ask the wrong thing. If it rewards memorizing a familiar procedure, it might establish unaided recall without testing whether the student understands when that procedure applies. Physical presence strengthens attribution, rather than guaranteeing a meaningful question.

The more informative contrast therefore concerns observable understanding and unverified output. Paper is one way to create that observation. An oral explanation, a supervised debugging exercise, or a follow-up problem can pursue a similar goal.

Those alternatives bring their own burdens, including consistency and staffing. They also show why declaring supervised exams the only solution goes beyond the available evidence. The defensible claim is narrower: controlled assessment directly addresses uncertainty about outside assistance.

For the Richard Borcherds homework grading debate, that is a consequential distinction. The reported policy offers a response to unreliable evidence. It should be evaluated as an assessment choice with costs, rather than a universal verdict on how every subject should be taught.

AI detection leaves another judgment problem

Detection software cannot remove the need for a fair process that distinguishes suspicion from evidence.

An AI detector attempts to classify whether software generated or assisted with text. A false positive occurs when it flags human work incorrectly. A false negative occurs when generated work passes without being flagged.

Both errors matter in education. Missing unauthorized assistance can undermine assessment fairness. Incorrectly accusing a student can damage trust and impose consequences on someone who followed the rules.

Berkeley’s published AI detection guidance says the campus opted out of Turnitin’s AI-writing detector after a pilot and consideration of outside research. The guidance also discusses classification problems involving multilingual writers and work that combines human writing with AI assistance.

That institutional position supplies context for the search for other assessment methods. It does not prove that every detector always fails. It establishes that Berkeley has identified sufficient concerns to avoid adopting that particular detection feature for general campus use.

The distinction between identifying assistance and establishing misconduct adds another complication. A course might permit grammar correction while prohibiting generated reasoning. A classification score does not by itself establish whether the student crossed that course-specific boundary.

Nor does a score necessarily explain the student’s process. A conversation about an answer can reveal confusion, but confusion alone does not prove cheating. Students can forget, become anxious, or misunderstand their own independently produced work.

These limitations make the strongest response procedural. Instructors need clear expectations and a defensible way to evaluate concerns. Otherwise, uncertainty about authorship can become uncertainty about whether an accusation was justified.

The reported Richard Borcherds homework grading change avoids some of that conflict by reducing reliance on the disputed artifact. Instead of trying to establish how a take-home answer originated, an instructor collects separate evidence under controlled conditions.

That approach has a clear attraction. It can reduce the importance of proving misconduct for every suspicious assignment. It also leaves an unresolved question: what happens to students who need detailed feedback on the work that no longer receives correctness-based marks?

Removing a score does not require removing feedback, but feedback still demands attention. A student needs to know whether a proof is incomplete, whether an assumption is unsupported, or whether a method applies only under narrower conditions.

An ungraded solution sheet can help, yet comparison requires enough understanding to identify meaningful differences. A struggling student might recognize that their answer differs without understanding why. That is where teaching support remains important.

The skeptical case against completion grading therefore deserves attention. The policy can make assessment more honest about what it knows while leaving students with less guidance about what they misunderstand. Its value depends partly on how the course fills that gap.

There is no published outcome evidence here establishing that Borcherds’ approach improves learning, reduces misconduct, or distributes burdens fairly. Those remain questions for evaluation. His reported concerns explain the decision, but they do not settle its effects.

Berkeley already offers a contrasting approach

AI-permitted practice and supervised checks can coexist, creating an alternative to both unrestricted take-home grading and a blanket ban.

A concrete comparison appears in Berkeley’s fall 2026 Statistics 243 syllabus, taught by Chris Paciorek. Its statistical computing assessments allow AI assistance while using in-person quizzes to check students’ understanding of problem sets.

The published grading structure assigns 20% to problem-set completion, 35% to related quizzes, 35% to a mini-exam, and 10% to participation. The syllabus also describes an oral retake procedure and requires attribution for ideas obtained through AI assistance.

That example does not establish a superior outcome. It does show an existing course design that separates supported preparation from individual demonstration. The comparison is especially useful because it comes from the same university, rather than an abstract proposal.

The policy also changes the meaning of permission. Allowing a tool during practice does not promise that its output will satisfy the later assessment. Students still need enough understanding to explain and apply what they submitted.

This creates a more precise distinction than “AI allowed” versus “AI banned.” Permission depends on the activity. A student might use assistance while working through a problem and then demonstrate understanding without that assistance during a subsequent check.

For computing, that arrangement has a practical rationale. A program can produce the expected result while containing choices its user cannot explain. A follow-up question can test whether the student understands the relationship between the code and its behavior.

That is an illustrative application of the assessment structure, rather than a reported incident from the course. Its importance lies in the connection between the assignment and the verification. The later assessment asks whether learning survived the use of assistance.

The comparison also exposes a weakness in treating all AI use as one behavior. Asking for a syntax reminder differs from requesting an entire solution. Both involve software assistance, but they can leave students with very different responsibilities and understanding.

For AI product developers, the implication is a design question. A tool that prioritizes complete answers might support immediate task completion. An educational setting also needs ways for the learner to remain responsible for reasoning, checking, and explanation.

Claims about learning benefits consequently require evidence beyond the quality of generated output. A convincing answer demonstrates the system’s performance. It does not establish that the user can reproduce the reasoning, transfer it, or identify a plausible error.

Students face a related distinction when organizing study materials. A searchable collection of correct answers can save retrieval time while leaving comprehension untested. Resources on personal knowledge management address organization, but retrieval alone cannot certify mastery.

The relevant question for either assessment model is what happens after assistance. Can the student defend the result, adapt it to a changed problem, and recognize its limits? Borcherds’ reported approach and the statistics syllabus both direct attention toward that question.

Their coexistence also limits sweeping conclusions about Berkeley. The university contains different course designs responding to different teaching needs. The story concerns an instructor’s decision within that variation, rather than a single institutional answer to AI.

What to watch after the grading change

The next useful evidence concerns assessment rules, student understanding, and teaching support, rather than another round of flawless AI answers.

Over the coming months, the first signal is whether additional instructors publish comparable grading changes. The informative documents will specify what homework earns, how understanding is checked, and which uses of assistance remain permitted.

If more courses explicitly separate completion credit from supervised demonstration, the argument for a broader assessment shift strengthens. If policies remain isolated or move toward other approaches, claims of a uniform retreat from homework grading weaken.

The wording matters as much as the announcement. “Homework is ungraded” can describe optional practice, required submission, or work that receives feedback without contributing to the final score. Those arrangements create different incentives and should not be counted interchangeably.

A stronger follow-up would also establish timing. The October reporting tells readers when the decision became news, but not necessarily when each classroom practice changed. Without that distinction, an older adjustment can look like an immediate response to a recent model release.

The second signal is evidence about independently demonstrated understanding. Relevant comparisons would examine what students can explain or solve under clearly described conditions. Higher take-home scores alone would not resolve the concern that prompted the policy.

If students retain or improve their supervised performance while using AI during practice, the case for separating assistance from certification strengthens. If performance declines, educators would need to examine whether students lost practice, useful feedback, or both.

Those findings would still require careful interpretation. Different questions, cohorts, instruction, and assessment conditions can affect scores. A gap between homework and exam performance is a reason to investigate, rather than a standalone measurement of cheating.

The third signal is what happens to teaching support after correctness grading changes. Courses should be examined for the feedback they retain, the opportunities students have to resolve mistakes, and the accessibility of supervised assessment.

If students receive useful explanations and repeated opportunities to demonstrate progress, completion credit looks more credible as part of a wider design. If grading disappears while confusion goes unanswered, the skeptical case becomes stronger.

These signals also matter to employers and professional trainers considering similar changes. A candidate’s completed exercise and their explanation of it answer related but different questions. Verification should match the competence an organization actually needs to assess.

The Richard Borcherds homework grading story is ultimately about the evidence attached to a grade. According to the reporting, he no longer trusts correctness on unsupervised assignments to carry the same meaning. That judgment deserves examination without assuming either that homework has become useless or that exams resolve every difficulty.

For students and educators following the story, start with the actual course rules and the evidence they are meant to produce. Ask whether practice remains useful, whether feedback reaches struggling learners, and whether assessments measure the skills the course promises. For AI users outside education, apply the same scrutiny to your own work: a finished answer can conceal an unresolved understanding gap. Keep the sources and explanations that support your decisions, then test whether you can defend those decisions when the prompt changes. As further details emerge about Borcherds’ approach, judge it by what students can independently explain and apply. What evidence would convince you that learning improved after the grading changed?