Abstract
<p>The introduction of large language models has intensified concerns around breaches of academic integrity in higher education. The current narrative on AI-enabled cheating among students has been largely shaped by survey-based measures in academic outlets and anecdotal evidence in media outlets, and personal experiences. We present a multi-pronged method for detecting Concerning AI Usage (CAI) in a preregistered study with 81 undergraduate students at an R1 university in the US. We find that—depending on which signals we rely upon to detect AI usage—41% to 70% of students engaged in CAI. We also investigate the relationships between CAI and demographic features, self-reported intellectual virtue, and learning outcomes, and show the limitations of relying on self-reports for AI usage and intellectual virtue.</p>