UC Berkeley Professor Replaces Homework Grading with In-Person Exams Amid AI Concerns
A Fields Medal-winning mathematician cites the difficulty of distinguishing AI-generated work from student submissions.
Richard Borcherds, a distinguished mathematician at UC Berkeley and a recipient of the Fields Medal, has altered his grading methods due to the increasing prevalence of artificial intelligence use among students. Borcherds has stopped grading homework for correctness, attributing this decision to the challenge of identifying AI-generated submissions. Instead, he now primarily bases student grades on in-person exams where electronic devices are prohibited.
Borcherds stated that online exams have become unproductive because many students utilize AI to answer questions, making it impossible to discern genuine student work. He noted a significant increase in average scores on online exams, which he believes may not accurately reflect students' understanding.
Research supports Borcherds' observations. A 2024 study published in Discover Education analyzed exam scores from an introductory statistics course and found a notable rise in scores when exams transitioned from in-person to online formats following the advent of AI tools like ChatGPT. Similarly, a study in the Teaching of Psychology journal compared two sections of an undergraduate personality psychology course and found that students taking the final exam online without monitoring scored significantly higher than those taking it in person. The study concluded that unmonitored online exams might inflate scores and not accurately represent students' abilities.
Borcherds explained that AI technology has advanced to a point where AI-generated work rarely contains the characteristic errors it once did, making it difficult to distinguish from human-created content. Consequently, he now credits students mainly for completing their homework assignments, rather than for the accuracy of their answers. He believes that grading for correctness incentivizes students to use AI and penalizes honest students, while also making the task of grading AI-generated content a waste of resources.
UC Berkeley requires instructors to outline their policies on AI use in course syllabi. While the university has explored AI detection software, the tools have not proven sufficiently reliable for campus-wide adoption, leaving individual professors to establish their own guidelines. Borcherds expressed a pragmatic outlook, suggesting that if the most significant societal impact of AI is making homework assessment more challenging, he would consider it a relief.