Sticky Help

A few weeks ago, in Bergen, Norway, a paper called "Sticky Help, Bounded Effects" was presented at the 16th Annual Learning Analytics & Knowledge Conference.¹ The title is a small piece of poetry on its own. Researchers from North Carolina State and Carnegie Mellon had spent the 2022 to 2023 school year watching what happened when middle and high school teachers used an intelligent tutoring system called MATHia. They logged 1,437,055 interactions across 339 students in 14 classes at 10 different U.S. schools, then asked a single quiet question.²

When a teacher chose to step away from their desk and help a student, who did the teacher choose?

The answer, after all that data, is gentle and unsurprising: the same students they helped last time.¹ In a classroom where an AI tutor is supposed to free the teacher to roam, to notice, to read the room with new clarity, the teacher's attention still gravitates to the desks it already knows. The students who were helped before are the students helped again. The students nobody touched yesterday are mostly not touched today.

The researchers gave this a name. They called it sticky help.

There is a second finding, just as important. When teacher help did arrive, its measurable effect was bounded. The lift it produced was visible inside the same session, in the next few problems the student attempted, and then it faded.¹ The help was real but it did not travel.

These are not, on their face, criticisms of teachers. The teachers in the study told the researchers what every teacher knows in their bones: they would love to spend one-on-one time with every student in the room, and they cannot.² The math is the math. Twenty-eight students, fifty-three minutes, one body, two hands. The choice of where to walk is not a moral choice. It is the choice the day makes for you.

What the Tutor Sees, and What It Does Not

In theory, an AI tutor should solve this problem. It logs every keystroke. It knows which students are stuck, which are coasting, which are progressing in clean lines. It produces a heat map of need.

In practice, the heat map and the teacher's walking route are two different documents.

The AI tutor sees one band of student behavior, the band that happens inside the software itself. It does not see who has not eaten today. It does not see the student who needs a question phrased differently before they can begin. It does not see the friend group that has just rearranged itself across the room and changed the chemistry of three desks at once. The teacher sees those things. The teacher also sees, by reflex and by history, the students they have already worked with, because the relationship is the thing that calls them over.

The result is a room with two parallel forms of intelligence, neither of which can see what the other sees. The AI flags engagement states. The teacher walks the route the day calls for. The students who fall between those two attentions sit there, quietly, doing or not doing the work.

The Visibility a School Could Actually Use

This is the gap a different kind of tool could fill, and it is the gap we built Koan to address. Not by removing the teacher's instinct, which is the most valuable instrument in the room, but by giving the teacher a quiet, honest record of where their attention went, and where it did not.

When a teacher in a Koan classroom looks back at last Tuesday, they can see whose writing they touched, whose work they read carefully, whose drafts moved between sessions and whose drafts paused. They can see who Aidan, our AI tutor, spent twenty minutes with on a paragraph that never made it into the final piece, and they can see which student wrote three drafts without ever asking for help. The record does not tell the teacher who to walk toward next. It returns to them, in low light, the information their day was too fast to hold.

The researchers in Bergen ended their paper with a careful sentence. They suggested the findings could be used to build tools that help teachers track their interactions and ensure each student gets the attention they need.² That sentence is the next decade of school technology, compressed.

The first decade of AI in classrooms gave teachers more help. The second decade has to give teachers more sight.

The students who fall through the cracks in a class with an AI tutor are not falling through because the tutor is bad. They are falling through because two systems of attention are operating in the same room without ever comparing notes. Visible work is the bridge. The teacher walks the route their instinct knows. The AI logs the practice it can see. A patient record of student work can show, calmly, which students have been carried by both, and which have been carried by neither.

If a teacher could see, at the end of each week, exactly which students they had touched and which they had not, would they walk a different route on Monday?