UMD Virtual Study Assistant Tied to Lower Grades
The University of Maryland evaluated its GPT-4o-powered, course-grounded Virtual Study Assistant in a randomized Fall 2025 trial involving 2,379 undergraduates and 30 instructors. In matched-course sections, access coincided with grades 0.37 standard deviations lower and LMS participation 0.90 standard deviations lower, with larger estimated losses for first-generation students. [Working paper](https://edworkingpapers.com/ai26-1598)
The study warns that answer retrieval is not the same as tutoring: even a course-specific, RAG-powered assistant can be pedagogically misaligned. It is not proof that every AI tutor harms learning, but it makes deployment design and learning-outcome evaluation essential.
- –Only about 15% of offered students used VSA; nearly 74% of requests sought information, explanations, or solutions, while fewer than 1% requested feedback.
- –Most instructors retained direct-answer mode, suggesting the product behaved more like an answer engine than a Socratic tutor.
- –Engagement fell even though most students never used the tool, indicating that introducing AI may change broader classroom norms and study behavior.
- –Larger estimated losses for first-generation students raise an equity concern for AI tools marketed as universal academic support.
- –Developers should evaluate independent learning, retention, transfer, and course participation—not just response quality, usage, or retrieval accuracy.
DISCOVERED
2h ago
2026-11-05
PUBLISHED
3h ago
2026-11-05
RELEVANCE
AUTHOR
heynavtoor