The tool reads the sentence, not the learning

Seoul National University caught a group of students last month who had used AI on their assignments, and Korean universities are now installing detection tools. A search for ways around those tools returns a stream of tips, and a dentistry student at Yonsei University described classmates who slip through by changing a few words.[1]

Seen from the student's side, the mechanism is plain: what the detector scores is the surface of the sentence, so the cheapest countermeasure is a surface edit. Swapping a handful of words costs minutes, while writing the assignment costs an evening, and the tool cannot tell the two apart. There is a fair alternative reading — the detectors may work on most students and only the determined few route around them, in which case how many were caught matters more than how easily the workaround spreads. The reporting names the workaround and the catch at Seoul National University, but no rate for either.[1]

The schools with published figures changed the exam

The measured picture elsewhere is not kinder to detection. At the University of Reading, human graders identified AI-written exam answers only 6 per cent of the time. Nearly 90 per cent of Harvard undergraduates say they use AI for coursework and about 25 per cent say they used it in place of assigned reading, so the population being screened is close to everyone.[2]

So the institutions with numbers moved the lever they can still pull, which is the shape of the task. The University of Sydney runs two lanes, proctored assessment beside coursework where AI is used openly; the University of Chicago Law School is piloting device-free first-year classes with offline exams and oral defences. The size of what that format decides shows up at Brown, where an economics professor recorded an average of 96 per cent on take-home midterms and 48.6 per cent on the in-class final. Korean campuses and these universities are working under the same constraint — neither software nor a human grader separates the two reliably — and only one of the two groups has published what its own format produces.[2], [1]

The signal worth waiting for

The Korean experts quoted are already pushing in that direction. Lee Jae-sung, who teaches AI at Chung-Ang University, argued the question should move from whether AI was used to whether the content was properly checked and whether anything in it is wrong. Lee Kyoung-jun of Kyung Hee University put the principle as using AI well and answering for the result. Universities are distributing guidelines that write those rules down.[1]

A guideline is a statement of intent until someone measures what it changed, and the Brown professor showed the cheap way to measure it: the same cohort, one take-home task and one in-room task, both averages published. If a Korean university publishes that pair before the end of 2026, the gap between the two averages will say whether the new rules moved anything or only moved the paperwork. Until then the honest reading for a student, a parent or a teaching assistant is that the detector settles the format of a sentence and nothing beyond it.[1], [2]