Allowing Generative AI in Class: Evidence from a Semester-Long Controlled Teaching Study

Abstract

We report a controlled, semester-long teaching experiment in an upper-division antitrust economics course. Two back-to-back sections were held constant in content, assessment, and grading; they differed only in policy and guidance on generative AI: one section was permitted to use AI with disclosure and structured training, the other was prohibited from using AI and received parallel non-AI study guidance (n = 29 vs. n = 28). We find no detectable effect of AI permission on proctored exam scores or final grades. By contrast, AI access is associated with higher engagement on in-class activities, longer and more concentrated AI sessions in other courses (15–30 minutes), greater metacognitive behaviors (preferring one’s own answers, catching errors, modifying outputs), more positive perceptions—especially regarding efficiency, confidence, and engagement—and stronger intentions to continue using and studying AI, as well as choosing AI-intensive careers. Standardized course evaluations are also consistently higher in the AI section. Taken together, structured AI access with guardrails appears, in our setting, to reshape how students learn and feel about learning, without raising exam scores.