A professor's retrospective on flagging 207 students for AI-written code

Jeffrey Turkstra, who teaches CS 240 at Purdue, published a long retrospective on the semester his AI-detection tool went wrong in public. His tool, Argus, flagged about 207 of 599 enrolled students for code it judged could only have come from a large language model, and the university sent those students an email demanding they confess a violation by a deadline or fail the course. The timing, near the drop deadline, and the framing drew a backlash from students, parents, and administrators, and Turkstra eventually reversed the retroactive penalty and let students who had dropped re-enroll. Of 117 who returned the form before the reversal, 108 admitted responsibility and 9 disputed it.

The detail worth keeping is the performance gap. Turkstra reports that students flagged as using an LLM scored 10.75% lower on the first proctored midterm and 14.5% lower on the second than their peers, which he reads as evidence that leaning on generated code hurts learning rather than helping it. He is careful about method: Argus does static analysis to find anomalies he argues only an LLM would produce, not a trained classifier guessing at style. His own lessons are procedural, deploy such tools at the start of a term with the institution aligned, and build in due process so honest students are not swept up.

Why it matters

If you teach or set academic policy, this is a concrete case of detection at scale going sideways, and the midterm gap is a usable argument that shortcutting with an LLM shows up later on supervised exams.