Dumlao, Wang, Xie, Hu, Bar, Chaney, Gold & Teplitskiy (2026) — University of Michigan. arXiv preprint (cs.CY). 📄 Full text (arXiv)
This large-scale observational study tests the "GenAI substitution hypothesis" — the concern that students offload cognitive effort to generative-ai and earn inflated grades without learning. Using syllabus and administrative data from a large U.S. university (2015–2025; 156,135 students; 87,936 course offerings), the authors measure each course's GenAI susceptibility with a human-validated LLM pipeline that extracts assessment types from syllabi, then apply a difference-in-differences design comparing outcomes before and after ChatGPT's release while modeling COVID-19 effects as persistent or transient. They find no significant differential effect of GenAI availability on grades overall or among previously lower-performing students, and no significant effect on self-reported understanding; effects on subject interest are significant only under a transient-pandemic assumption. The findings temper alarm about grade inflation and satisfaction erosion, complementing ai-availability-student-motivation and the mixed picture in generative-ai-reduced-study-time-math. The null result is notable against theoretical worries about cognitive-offloading and situates student-experience concerns in higher-ed on firmer empirical ground.
Related Pages
- ai-availability-student-motivation — related study on how AI availability shifts student motivation
- generative-ai-reduced-study-time-math — companion finding on GenAI and study effort
- cognitive-offloading — the mechanism the substitution hypothesis presumes
- student-experience — self-reported understanding and interest outcomes
- higher-ed — large-university institutional context