UC Berkeley Professor Replaces Graded Homework With In-Person Exams as AI Use Grows

For Richard Borcherds, graded online homework no longer gives a dependable picture of what students can do on their own. The University of California, Berkeley mathematician has shifted most assessment to in-person exams without electronic devices, saying AI makes it difficult to distinguish students’ work from machine-generated answers.

Borcherds told the New York Post that online exam averages had risen substantially, while he couldn’t identify which students had used AI. He described graded homework as creating an incentive to submit AI-generated answers: students who do the work themselves may receive lower marks, while instructors spend time assessing work that may not reflect what its author understands.

That’s a practical problem for a course built around mathematical reasoning. A polished answer can show that a problem was solved, but it may reveal little about whether the student can reproduce the reasoning, spot an error or apply the same idea to a new question. If an instructor can’t reliably tell those things apart from a submitted solution, the grade risks measuring access to an answer generator as much as learning.

Grades become harder to interpret

The difficulty has changed alongside the tools. The source article traces the problem to the arrival of GPT-3.5 in November 2022, when popular large language models could produce conspicuous mistakes. Those errors made machine-generated work easier to notice. As the systems improved, that signal weakened: a plausible written answer no longer offers a clear way to tell who produced it.

That doesn’t establish that every unusually strong online result comes from AI use. The article cites research finding that students often score better on online exams, but says the studies have not definitively shown AI was the reason. The concern is instead that, when exams are unmonitored, the resulting scores may stop serving as a reliable measure of what students learned. Better marks alone can’t reveal whether students have gained a skill or received help that the assessment was meant to exclude.

A Brown University professor faced a related gap between take-home and in-person results. His students averaged 96% on take-home midterms, compared with a historical range of 65% to 80%. He then held the final exams in person. Just two students scored within 10% of their midterm marks, and one of those students scored higher. The difference raises questions about comparability between assessment formats; by itself, it doesn’t identify what caused any individual student’s result.

For instructors, that distinction matters. A take-home assignment can still be useful for practice and feedback, but its score may carry less weight as evidence of independent performance when students can consult AI tools. An in-person exam limits that assistance during the assessment and gives a teacher a more direct view of each student’s work. It also shifts more of the grading burden toward scheduled exams, which can test only the tasks and conditions instructors choose to include.

Detection has limits

UC Berkeley has considered AI-detection software, but the university found it unreliable. That leaves instructors to set their own course rules, provided they describe those rules in their syllabi. Clear instructions can tell students what kinds of AI assistance are allowed, but a syllabus can’t, on its own, verify who wrote a particular answer.

The article also points to an example of a coding assistant telling a user to learn to code instead of generating the requested code. That response shows that tools can differ in how they handle requests, but one example doesn’t provide a dependable way to control what students use across a large range of available systems. For instructors, relying on a particular tool’s response would leave the underlying assessment problem in place.

Borcherds’ choice is therefore a change in what his grades are meant to capture: more evidence gathered under conditions where students can’t use electronic devices, and less reliance on homework scores that he considers difficult to interpret. The tradeoff is straightforward. In-person exams can make independent work easier to assess, while reducing the role of online assignments that may help students practise between classes.

The open question is how courses will balance those aims as AI tools continue to improve. For now, Borcherds says he’ll be relieved if AI’s worst effect is simply making homework harder for teachers to grade. His response suggests that assessment design, more than automated detection, may determine whether a grade still tells instructors what a student can do.