Only about 15% used the AI tutor. What did the university trial measure?
An October working paper reports lower final grades after a university introduced a course-integrated AI study assistant. Only about 15% of students offered it used it even once. Those two findings belong together. The University of Maryland study involved 2,379 undergraduates and 30 instructors. Researchers randomized access at the instructor level within groups of identical or similar courses. Students in both conditions could use other AI tools. The experiment therefore estimates the effect of introducing this assistant into those classes, not the effect of personally using it. That distinction doesn't make the result unimportant. A university decides whether to introduce a tool, and the consequences can extend beyond the students who click it. The authors suggest that introducing it may have changed broader AI use, but they don't establish that explanation. There is another wrinkle in the word ‘tutor’: the default mode supplied answers directly. Only two instructors ended the term with the guided tutoring mode selected. These findings concern that configuration and rollout, not every approach to AI-supported teaching. I'd take the result seriously when deciding whether to expand this service. I wouldn't tell an individual student ‘using the tutor lowered your grade’ on this evidence. And before the next rollout, I'd want the university to explain what students should keep doing—practice, reading, asking questions—rather than leave them to discover whether the new help replaces something they needed.
Comments
Before a student types ‘I still don't understand this,’ can they tell who will be able to read it? I'd put that explanation beside the message box, especially when the assistant sits inside a graded course. A student may ask a different question if they think they're talking privately rather than leaving something their instructor can inspect. The paper says instructors and researchers could view interactions and notes that this may have affected what students asked. It doesn't establish that this caused the low uptake. I'd check whether students understand that audience before treating an unused tutor as unwanted help.
I'd keep staffed help on the timetable while testing the assistant, too. A student who doesn't want to leave a readable chat history still needs somewhere to ask a muddled question. Before replacing a drop-in session, check whether students who decline the chatbot can actually get help before the assignment is due. A licence for every student is easy to count as expanded support; it doesn't tell you who can use that support comfortably. That's a condition I'd put on a rollout, not a claim that this university cut its teaching hours.