Students Asked the AI Tutor for Answers. Almost No One Asked for Feedback.

On October 7, 2026, Kathryn Palmer reported in Inside Higher Ed on a study the University of Maryland ran on its own AI tutor. In a randomized trial during the fall 2025 semester, involving 2,379 undergraduates and 30 instructors, some course sections were given a Virtual Study Assistant built on ChatGPT. The results were murky. Only about 15 percent of students with access chose to use it, and access marginally reduced final grades and course participation. The university’s provost acknowledged that the study raised questions it hasn’t yet answered.

Palmer and the experts she interviewed focus on the design problems. Students already had ChatGPT and Gemini. Instructors needed far more training. Field trials bundle so many factors together that clean conclusions are hard to reach. All of that is fair. But one finding deserves more attention than it received.

Look at what students asked for

Among the students who used the tool, 73.8 percent wanted information, explanations or solutions. About 11 percent used it to practice or prepare for tests. Less than 1 percent asked for feedback on their own work.

That number tells us less about the tool than about what students believe school is for. Years of grading have taught them that the goal is the right answer, delivered on time. So when a machine offers answers instantly, they take answers. Very few think to ask the more useful question: here is what I think, so where am I wrong?

Asking for feedback requires a student to put their own thinking on the table first. That takes a sense of themselves as learners, and most were never taught to build one. Melissa Loble of Instructure once told me no one had ever taught her how to learn, a conversation I wrote about in Readiness Begins With Knowing How You Learn. The Maryland data shows what that missing lesson looks like once AI arrives on campus.

A watched room changes the questions

The researchers point to another reason use stayed low. Instructors could see student chat logs, and the authors suggested that visibility may have shaped whether students used the tool and what they were willing to ask.

Of course it did. Few students will admit confusion in a space their professor is reading. When people know they are being watched, they manage how they come across, and the questions that matter most are the first to go quiet. I made a similar point about recorded workplaces in When Every Word Becomes Training Data.

Then there is the setting itself. In default mode, the assistant simply returned the correct answer. Tutor mode guided students toward it instead. Few instructors switched tutor mode on. The institution, in other words, defaulted to answers too.

One tool can’t fit every learner

The study’s lead author, Jing Liu, put the adoption problem plainly: “You cannot force people to only use one tool.”

The low adoption carries its own message. Students had ChatGPT and Gemini within reach, and a campus tool with a professor reading along had to compete with tools they could use on their own terms. A campus that wants a say in how students learn with AI has to offer something those tools can’t: help that knows the student, the course and the kind of thinking the course is meant to build.

Campuses have good reasons to want a single, sanctioned AI tool. It protects student privacy, controls cost and lets researchers study what actually happens. That consistency is worth keeping. But a common tool only earns its place when it adapts to the person using it. A first-year student who needs confidence, a senior polishing a thesis and a working parent studying at midnight need very different things from the same assistant. USC’s Stephen Aguilar suggested that the opportunity for colleges lies in smaller models that do a few things very well, which points in the same direction.

What colleges can do now

Make feedback the default. Turn on tutor mode, and design assignments that ask students to share their own reasoning with the AI before it responds.

Separate the practice space from the grading space. Let students work through confusion where mistakes aren’t on display, and give instructors the overall patterns rather than every transcript.

Teach students how they learn before teaching them a tool. A short reflection at the start of a course helps: Where do I usually get stuck? What kind of help do I actually use? Those answers give any AI something personal to work with.

Train instructors as seriously as you configure the software. Liu’s call for much deeper instructor training is the most practical takeaway in the study.

The lesson in the smallest number

Maryland deserves credit for studying its own tool in public and being honest about what it could not prove. The most useful lesson may sit in its smallest number. AI will answer whatever students ask. The work of education now is helping students become people who know which questions are worth asking, and who feel free to ask them out loud.

Want to explore these ideas further? Learn more about my work on leadership, identity and conviction at www.theglennllopis.com.