The risk of AI cheating is evident from the LMS data for the first eight weeks: 73 per cent correct predictions

Of 52 Turkish students, 44 per cent were classified as high-risk based on their final exam papers. Five browsing variables are sufficient to predict this classification right from the start of the term.

The study involves 52 first-year students in computer science and educational technology at a public university in Turkey taking an introductory programming course. The question raised is whether the risk of using AI during the final exam can be anticipated from the digital traces left in the educational platform during the first eight weeks.

The labeling is based on the exam logs: copies, loss of focus of the window, right-clicks. Twenty-three out of 52 students, or 44.2%, were classified as high-risk during a face-to-face proctored exam. Five variables selected from 27 candidates were then used to predict group membership, tested with logistic regression, naïve Bayesian, random forests and gradient boosting, in leave-one-out cross-validation with pre-processing and selection redone at each fold. Logistic regression came out on top, with 73.1% accuracy.

The most regularly selected variables are course module consultations, assignment submissions, and the number of distinct days of access to the videos. The authors set an explicit limit on the use of their model: these predictions are aimed at early pedagogical support, not the establishment of misconduct or the opening of disciplinary proceedings.