Signal Learning-Analytics Model Predicts AI-Cheating Risk Early, Not for Discipline

Summary

A study released on arXiv September 29, 2026 tested whether AI-assisted cheating risk in a final exam could be predicted from students' LMS digital traces during the first eight weeks of a semester. The sample was 52 first-year Turkish university students in an Introduction to Programming course. Based on suspicious behaviors like copy events, focus loss, and right-clicks recorded during a proctored final exam, 23 of 52 students (44.2%) were labeled high-risk. Researchers selected five features from 27 candidates and built models using Logistic Regression, Naive Bayes, Random Forest, and Gradient Boosting, evaluated via leave-one-out cross-validation. Logistic Regression performed best, reaching 73.1% accuracy. Course-module views, assignment submissions, and days accessing course videos were the most consistent predictive features. The authors stress the predictions are meant to support early academic guidance, not to establish misconduct or trigger discipline.