Many students will often decide to cheat on assessments if they think they can get away with it. Generative AI now makes escaping detection seem promising, though that sense of security can turn out to be false.
Brown University economics professor Roberto Serrano encountered this phenomenon earlier this year and started attracting attention because of it in late June. He gave his Welfare Economics and Social Choice Theory class its first take-home midterm in nearly two decades. The change aimed to reduce anxiety after a horrific campus shooting in December.
As a result of the new exam format, the class size rose sharply from around 30 students to 86. Once this expanded group’s midterms were in, the outstanding results looked highly suspicious. The average score hit 96 per cent, with 40 perfect papers. Serrano and his graders then tested the questions on ChatGPT, only to discover that the AI outputs matched the students’ unusually convoluted style and specific arguments nearly verbatim.
Serrano told the class he suspected AI-assisted cheating. With his dean’s approval, he switched the final exam to an in-person format. He warned that he would void the midterm if the final results differed dramatically, and they did.
The supervised final averaged just 48.6 per cent, the lowest in the course’s history. Twenty-seven students dropped out or skipped the final. Twenty-two of them had scored perfectly on the take-home midterm. Only a select few performed similarly on both examinations.
This appalling discovery led Serrano to void the midterm. He raised the final’s weight to 80 per cent of the grade and lowered the pass mark to 40 per cent. Nineteen students still ended up failing the course.
“We cannot afford to have a society in which a significant fraction of our best young minds think that cheating is OK,” the professor said. “That leads to a declining society, to a failed society… We cannot choose to become idiots.”
This Brown case adds weight to other recent AI cheating problems across American schools. Princeton University ended its 133-year tradition of unproctored exams under the honour code in May. Faculty voted to require proctors because AI made hidden cheating too simple.
In 2024, a UC Berkeley-led study surveyed more than 95,000 undergraduates at 20 research universities. Roughly two-thirds used generative AI and more than 9 per cent of users admitted they used it to cheat. Daily users cheated at far higher rates than occasional ones.
These examples showcase the same pattern. Easy access to AI turns unsupervised work into a temptation few resist when they expect no consequences. Academic institutions are now tasked with adapting their curriculums and teaching methods to address this issue. Although using artificial intelligence is highly useful for brainstorming and learning new concepts, copying and pasting its responses into an exam is not acceptable.
Read more: Number of students using AI to cheat grows at alarming rate
Follow Rowan Dunne on LinkedIn
rowan@mugglehead.com
