Six Years of Commit History
WEST LAFAYETTE, Ind. — Researchers at Purdue University have published one of the first large-scale longitudinal studies of LLM use in programming courses. Argus, described in the paper “Argus: Academic Integrity in the Era of Generative AI” (arXiv, September 28, 2026), analyzes students’ full Git commit histories across 13 assignments in a large-enrollment CS2 course, combining behavioral and stylistic indicators — advanced C idioms not taught in the course, irregular comments, compressed working spans, and “bursts” of 30+ new lines written at 15+ lines per minute.
45% by Spring 2026
The study spans six Spring offerings (2020–2026), 3,893 student-semesters and 42,695 homework submissions. Mean scores sat near zero in the pre-LLM era (1.1 in 2020, 0.5 in 2022), rose modestly (2.0 in 2023, 1.8 in 2024), then jumped: 4.2 in 2025 and 11.9 in 2026. In Spring 2026, 45% of enrolled students exhibited patterns consistent with LLM-assisted code development.
The Performance Link
Crucially, the researchers correlated the flags with written, in-person, closed-note, proctored exams. The result: a significant negative correlation between flagged LLM use and exam performance — one that has strengthened in the most recent offerings. The pedagogical implication is stark: students who rely heavily on LLM assistance do not develop the independent programming ability the course is designed to produce.
Triage, Not Verdicts
The authors are careful: Argus is a triage tool, not an automated accusation machine. False positives carry serious consequences in large courses, so any automated detection must be paired with a structured human review process. In Spring 2026, course staff reviewed the top 300 flagged students — three reviewers working independently, then reaching consensus, with what the paper calls “extreme benefit of the doubt” for students.
Why It Matters
The 45% figure turns “are students using AI?” from rumor into six-year evidence. The negative correlation gives institutions a concrete stake in AI policy: it is not just about academic integrity, but about learning itself. Expect ripple effects on course design (more in-person assessment), academic policy, and the debate between detection tools and redesign. And a warning for edtech: detection alone will not solve this — the real battleground is assessment design.