Researchers at the University of Maryland, College Park, led by Jing Liu, ran what they describe as the first large randomized trial of AI tutoring across many subjects at a US public university. In the fall semester of the 2025-26 school year, 30 instructors teaching sections in STEM, the social sciences, the humanities and education were randomly assigned to give their students access to a GPT-4o study assistant built into the course website. It drew on materials the instructor chose. The paper is dated September 30, 2026 and was released as Annenberg Institute EdWorkingPaper 26-1598 in October (Liu et al., 2026).
Among sections of the same course, access to the tutor lowered final grades by 0.37 standard deviations and cut recorded participation in the course's online activities, such as discussion posts and quizzes, by 0.90 standard deviations. Homework results showed no gain. Estimated losses were larger for first-generation students. In a survey, students who used the tutor were more likely to say they asked their instructor fewer questions about course content (Liu et al., 2026). The authors' reading is that the tutor may have replaced the effort and contact that normally drive learning, rather than adding to them.