Maryland trial finds students offered an AI tutor got lower grades
In a randomised trial at the University of Maryland, grades fell in matched courses although only about 15% of students used the tutor. The working paper has not been peer reviewed.
A University of Maryland trial found that undergraduates offered a course AI tutor got lower final grades, the university said on 16 October 2026.
The working paper has not been peer reviewed. It covered 2,379 undergraduates and 30 instructors in autumn 2025. Instructors were randomly assigned to add the tutor to their courses or not. The tutor was built on GPT-4o and drew on each course's materials.
It’s clear from these results that simply giving people access to these tools doesn't mean they will work as intended.
The clearest result came from 1,353 students in sections of the same course. Their final grades were 4.12 points lower out of 100, which is 0.37 standard deviations. Across all 2,379 students, the authors found no significant evidence of lower grades.
Recorded participation in course activities fell sharply in both samples. Estimated grade losses for first-generation students were more than double those of other students.
Only about 15% of students offered the tutor ever used it, so the authors say the effect may reflect wider AI use prompted by the tutor. Students in both groups could already use ChatGPT and Gemini. About 74% of requests sought information, explanations or solutions. Fewer than 1% asked for feedback on a student's own work.
The Hechinger Report said lead author Jing Liu does not conclude that colleges should avoid AI. A separate peer-reviewed trial in Nature Human Behaviour on 26 October tested 487 management consultants. ChatGPT helped them score higher on data science tasks, but the authors found no evidence they did better on technical questions without it afterwards.
For students and lecturers, the authors say the result applies to this tool as set up in this trial, not to all AI tutors. Only two instructors ended the term with the mode that guides students rather than giving answers. Grades were not an independent test of learning. Maryland Today said 40 courses have the tools and their instructors have received guidance. It does not say whether the university will offer the tool to every course.
Sources
- Maryland Today Study Suggests More Guidance Needed in Use of AI Course Assistants
- EdWorkingPapers, Annenberg Institute at Brown University The Effects of Course-Integrated AI Tutoring on Student Performance and Engagement: A Randomized University Trial
- The Hechinger Report Same researcher, opposite results
- Nature Human Behaviour Generative AI and the temporary upskilling of knowledge workers