Only 15 Percent of Students Used the AI Tutor. Grades Still Fell.

The University of Maryland tested an AI tutor inside real college courses before rolling it out. Only about 15 percent of students ever used it. In sections of the same course, final grades still came in about four points lower.

So how does a tool most students never touched change a class? That question is why this AI tutor study matters more than the grade number. It suggests that switching on an AI tool may change what a course says to every student in it, including the ones who never open it. The authors do not claim to have proved that, and I will show you where they stop. You can still act on the question this week, whether or not your campus ever puts a tutor in your course site.

The study went up in October as a working paper, and Jill Barshay of The Hechinger Report brought it to a wider audience on Monday. Credit goes to Maryland first. Provost Jennifer King Rice, the paper’s senior author, told Maryland Today that rather than roll the tool out to everyone “just because we could, we should slow down and take the time to study it first.” Then it made the troubling results public.

What the Maryland AI tutor study found

In fall 2025, Jing Liu and colleagues assigned instructors at random, within groups teaching the same or similar courses, either to add a virtual study assistant to their course site or to teach without it. The study included 30 instructors and 2,379 undergraduates from 11 of the university’s 12 schools and colleges. The assistant ran on GPT 4o, drew on materials the instructor selected, and sat in the sidebar of the learning management system.

The cleanest comparison is the group teaching sections of the very same course: 22 instructors and 1,353 students. Here is what came back.

• Final grades: 4.12 points lower on a 100 point scale, or 0.37 standard deviations. Letter grades came in 0.11 lower on a four point scale.
• Participation: about 26 fewer recorded actions, such as discussion responses and quizzes, a drop of 0.90 standard deviations. Across all 30 instructors the drop was nearly identical, at 0.89.
• Homework: the share submitted, and the share submitted on time, did not detectably change. The authors suggest students may have kept meeting requirements while doing less around them.
• First generation students: estimated final grade losses were more than twice those of their peers, 5.10 points against 2.23 across the full sample.
• Actual use: 14.63 percent of students with access ever used the tool, and about 40 percent of those users had a single session.

Now the limits, stated plainly. This is a working paper. Across all 30 instructors, the authors report no significant evidence of lower grades, which they say is likely due to weak alignment across different courses and grading. Course grades are not an independent test of learning. This is one university, one tool as configured, and one term. And it is no argument that AI tutors cannot work. The authors point to Harvard physics research in which a purpose built tutor produced larger immediate learning gains.

Why would a tool almost nobody used change anything?

The authors do not claim to know. They do have a suspect, and the word I keep coming back to is signal. Rice put it this way: “It may be the case that giving access to the VSA signaled an acceptance of AI tools more broadly, which prompted students to use other available AI agents.” VSA is the virtual study assistant. Students in both groups already had other campus AI tools, including ChatGPT and Gemini.

The paper adds evidence, all of it descriptive. In an end of term survey with an 18 percent student response rate, 41 percent of respondents in tutor classes said their instructor encouraged generative AI or built it into the course, against 11 percent in the other classes. The lower grades were not confined to the students who used it. In those same course sections, raw average final grades were 85.19 for users and 85.66 for non users, against 90.22 in the classes without the tool.

Look at what the instructors did, too. Of the 10 treated instructors who answered their survey, all mentioned the tool, 40 percent demonstrated it, and none built it into a task or an assignment. Only two ended the term in the mode that guides students toward an answer rather than handing it over.

The students who did use it mostly asked for answers. Of 3,460 requests, 73.8 percent sought information, explanations or solutions. Practice and test preparation came to 11.3 percent. Feedback on a student’s own attempt came to 0.7 percent. Liu told Maryland Today that this kind of use might make study time seem more efficient while bypassing the “productive struggle” students need for learning. And about 57 percent of tutor class respondents agreed they asked the instructor fewer questions about course content, against 44 percent elsewhere.

Here is my read, and it is my read rather than the authors’ finding. A tool in the sidebar, a mention in class and no further word is a message. Students have to guess what it says. The tool may matter less than the sentence nobody wrote to go with it.

Students say your words steer them

A second paper, released last week, points at the same lever. The 2026 Dartmouth AI Survey, from Bruce Sacerdote and Mason Parris-Bacon, includes a representative sample of 1,003 U.S. college students. In that sample, 71.2 percent agreed that when an instructor expressly permits or prohibits a particular use of generative AI, that guides how they use AI on the assignment.

The same survey shows what fills the silence. Eight percent said they paste in AI text with little editing, while 49.7 percent said their peers do. And 38.7 percent said they feel they have to use AI on out of class assignments because so many peers do. The authors suspect some of that gap is students not admitting it, and that much of it may be peer misuse being overestimated.

Self report is not behavior, and saying a rule guides you does not prove that it does. But set the two papers side by side and the practical point is hard to miss. In one, an unexplained tool seems to have moved a class. In the other, students say an explicit instruction moves them. You control the second one, this week, at no cost.

Run the signal check on one course

You do not need a tutor in your course to do this, because students are reading your course either way. Here is the check, in six moves.

1. List what students can see. A tool in the course site, a line in the syllabus, something you said on day one, a campus tool someone else switched on.
2. Write what a reasonable student would conclude. One sentence, in their words and not yours. This one stays private, and it is usually the uncomfortable part.
3. Say it expressly for the next two assignments. What AI may do, and what it may not. Name the task, not the tool, because the tools keep changing.
4. Say what comes first. What does a student try alone before asking anything? A bad attempt is still an attempt, and it gives a tool, or you, something real to respond to.
5. If a tool is in the course, say what it is for. Practice and feedback on work already tried were the rare uses at Maryland. Say those are the uses you want. If your tool has a setting that guides rather than answers, look at it.
6. Send one kind of question to a person. Not “my door is always open.” Name one kind of question you want brought to you, and say where and when.

Then post the result where students will see it, next to the tool if there is one.

What this will not do

Straight talk. No study here tests a card like this, and I am not going to tell you it raises grades. It is a judgment call. It will not stop a determined student from pasting in a prompt, and it will not fix an assignment that rewards a polished file with no thinking behind it. If that is your problem, AI Cheating Is a Design Problem. Draw It, Then Defend It. covers the assessment side, and Alaska’s Draft AI Policy Says a Human Decides. Your Syllabus Should Say Where. covers the policy language.

Grades are a blunt instrument in both directions, too. A University of California, Berkeley paper reported by The Wall Street Journal in May found A grades growing faster in writing and coding classes since ChatGPT arrived, and its author said the results suggest students relied on AI to do better, not that they were learning more. So if your campus offers you a tool, skip the adoption numbers. Ask the question from the essay Solve Everything: if a student uses an AI tutor for an hour, do they still have the skill 180 days later?

One disclosure. Harvest Kernel sells AI tools, including coaches for students, so weigh my read with that in mind. This study is a reason for us to be careful too.

Start with one course this week

The free session walks you through all six moves on your own course, and you leave with a finished AI card, ready to post.

Then the next thought arrives on its own. That is one course and two assignments. You may teach other sections, and every assignment left this term needs its own line. The Syllabus Generator builds a student centered syllabus grounded in your Course Record, the record of your course, your outcomes, your policies and your voice that every tool in the toolkit uses. Full access for 14 days, cancel in one click. Write this first card by hand anyway, because this sentence has to be yours.

If you would rather test your wording on other instructors, the Harvest Kernel Learning Community is free, and today’s question there is simple: if I asked a student in your course what you think about AI, what would they say?

Students read your silence on AI as an answer. Give them a better one.

Sources

• Jing Liu, Tracy Sweet, Marty Haoyuan Chen and colleagues, “The Effects of Course-Integrated AI Tutoring on Student Performance and Engagement: A Randomized University Trial,” EdWorkingPaper 26-1598, Annenberg Institute at Brown University, October 2026. edworkingpapers.com
• “Study Suggests More Guidance Needed in Use of AI Course Assistants,” Maryland Today, October 16, 2026. today.umd.edu
• Jill Barshay, “Same researcher, opposite results,” The Hechinger Report, November 2, 2026. hechingerreport.org
• Bruce Sacerdote and Mason Parris-Bacon, “Impacts of AI in Higher Education: A Panel Survey of Undergraduates,” 2026 Dartmouth AI Survey, October 2026. ai.dartmouth.edu
• “‘A’ Grades Are Suddenly Everywhere Since the Arrival of ChatGPT,” The Wall Street Journal, May 2026. wsj.com
• “Solve Everything.” solveeverything.org