Richard Borcherds, a mathematics professor at UC Berkeley and a 1998 Fields Medal winner, has stopped grading homework for correctness. He says widespread AI use has made the practice pointless: today it is almost impossible to tell a submission written by a student from one generated by AI.
His solution is a return to in-person exams with no electronic devices. "Online exams are now a waste of time as several students just use AI to answer the questions," he told the newspaper that interviewed him. "I cannot tell which students are using AI, but I notice the average score increases a lot in online exams." Borcherds also argues that grading for correctness gives students an incentive to hand the work to AI, penalizing those who are honest. Paying to grade AI-generated answers, he adds, is a waste of time and money.
A problem that began with ChatGPT
The trend began when OpenAI released GPT-3.5 in November 2022. Back then models made frequent mistakes, so AI-written homework was easy to spot. The technology has since improved to the point that it is very hard to tell whether it was used. AlexTech has covered how professors are abandoning AI detectors in favour of oral exams and in-class assessment.
This is not an isolated case. Several researchers have found that students score significantly higher on online exams. However, a September university study found that students prohibited from using AI perform worse than those who use it, challenging the stance taken by UC Berkeley Law. Without proving AI was used, the gap was large enough for them to conclude that unmonitored online exams no longer reflect what students have learned. A Brown University professor saw something similar this year: his take-home midterms averaged 96%, against a historical range of 65% to 80%, so he moved the final exams in person.
Why detection is not enough
UC Berkeley also tested AI-detection software and found it unreliable. The university therefore left individual professors free to set their own AI rules, provided they state them in the course syllabus. Borcherds says he has no ideal solution, but if the worst AI can do is make homework harder to grade, he will be "very relieved."
The case shows how traditional assessment is under strain. Many teachers still see homework as useful practice, but argue that grades should rest on work done under controlled conditions. In August, UNAM in Mexico annulled thousands of online entrance exams after AI proctoring flagged suspicious perfect scores.
No comments yet. Be the first!