Purdue: 45 per cent of students on a C programming course show signs of using an LLM, and they perform less well in the written exam

The Argus tool analyses the behaviour and style of code across six years of graduate intake; its alerts show a negative correlation with marks in invigilated examinations, but require human review.

Professors at Purdue University present Argus, a system that identifies C programming assignments likely written with the help of a language model. Rather than analyzing the final rendering alone, it combines behavioral indicators (how the work progresses as the exercise progresses) and style indicators to flag abnormal trajectories.

Six years of data

Applied to the spring sessions of a six-year high-volume second-level course, the tool reports 45% of spring 2026 enrollees as having patterns consistent with LLM-assisted development. These reports are significantly associated with poorer results in written, face-to-face and proctored exams.

The authors insist on the risk of false positives: an unfounded accusation weighs heavily, for both the student and the teacher, especially in mass courses. They believe that no automatic detection tool should be used without a structured human review procedure, and call for a rethink of course design and academic rules.