The Students We Don't See

This month, researchers at North Carolina State University and Carnegie Mellon published a study with a quietly devastating title: "Sticky Help, Bounded Effects."¹ It will be presented at the 16th Learning Analytics and Knowledge Conference in Bergen at the end of April. The dataset is large by classroom-research standards: 1.4 million student-system interactions, 339 middle schoolers, 14 classes, all using MATHia, an AI-powered intelligent tutoring system already in wide use across U.S. schools.²

What the researchers found is the kind of finding that should change how we think about classroom AI. Once a teacher helped a student, that student was more likely to receive help again in the next session, and the next. The pattern persisted even after controlling for whether the student was currently struggling. Meanwhile, students who sat idle, who quietly stopped engaging, were largely invisible to the teacher's attention loop. Help "stuck" to certain students. Disengagement did not.

The researchers also found that when help did arrive, it correlated with learning gains within the same session, but did not translate into later skill acquisition. The intervention worked in the moment. It did not compound.

What the Dashboard Did Not Show

The promise of AI in classrooms has always rested on a simple bargain. Teachers cannot watch every student at once, but a well-designed system can. Dashboards can highlight who is struggling, who is racing ahead, who needs a nudge. Give the teacher better information, the argument goes, and they will distribute their attention more wisely.

The NC State study suggests this bargain has a hole in it. Teachers had access to dashboards. They could see student progress. And still, their attention clustered. The students who had once received help became the students who continued to receive help. The students who quietly disengaged became, in effect, invisible.

This is not a story about negligent teachers. The teachers in the study were doing what teachers have always done. They were responding to signals. The problem is that the signal of "currently struggling" is loud, while the signal of "currently disengaged" is quiet. A student who has stopped trying does not raise a hand. A student who has gone numb does not appear on the dashboard as a red flag. They appear as nothing at all.

The Quiet Half of the Classroom

If you have spent time in a classroom, you already know this pattern. Every teacher has the students they "have" and the students they "lose." It is rarely a moral failing. It is the result of a system that surfaces visible struggle far more readily than invisible withdrawal. The student who messes up loudly gets a coach. The student who quietly stops caring gets nothing.

What the study makes newly visible is that AI-powered tutors, despite all their promise, can replicate this pattern rather than disrupt it. A dashboard that shows performance is still a dashboard about output. It tells a teacher who is getting answers wrong. It does not tell them who has stopped trying.

This is the gap between data and insight. A system can collect a million data points and still miss the child in the third row who closed her laptop ten minutes ago and is staring at the floor.

The Visibility That Matters

The future classroom, we believe, will not be built around more performance dashboards. It will be built around attention coverage. Teachers will not just see student data. They will see their own patterns of engagement, their blind spots, the students whose names have not come up today, the students whose process has gone flat.

The researchers themselves point in this direction. They argue for real-time analytics that track attention coverage and highlight under-visited students.² That is a small phrase with enormous implications. It means designing tools that watch the teacher as carefully as they watch the student. Not for surveillance, but for service.

This is part of what we are building at Koan. When Aidan, our AI tutor, works with a student, every revision, every pause, every shift in reasoning is captured in the WorkHub. But the WorkHub does not stop at the student level. It also surfaces the texture of an entire class. A teacher can see at a glance who has gone quiet, who has revised heavily and stalled, whose engagement patterns have shifted compared to last week. The goal is not just to flag struggle. It is to make the invisible students visible, even when their disengagement is silent.

What We Choose to See

The "sticky help" pattern is not new. What is new is that we can finally measure it, and that we have the tools, if we choose to build them, to break it. Every previous generation of education technology promised more data and more efficiency. The lesson of the NC State study is that more data is not the same as a better classroom. The data has to point at the right thing.

Right now, most classroom AI points at the loudest students. The quiet ones, the disengaged ones, the ones who have stopped trying, slip through. A truly visible classroom would do something stranger and more humane. It would help the teacher notice the absences, not just the alarms.

If a student stops trying and no one notices, did the classroom fail them, or did the technology built to help them simply look the wrong way?