What The Students Could Feel

Last week, the Hechinger Report surfaced a study¹ that most of the ed-tech conversation has been waiting for and quietly dreading in equal measure. It is the first large randomized trial of an AI teaching tool built not for the student, but for the teacher.² Researchers at Wharton and the University of Pennsylvania ran it across fourteen middle and high schools in a private chain in Turkey during the spring 2025 semester. One hundred ninety-three teachers. Two thousand eight hundred and sixteen students. More than fourteen thousand student-course observations.³

The tool itself was a custom generative AI assistant built on GPT-4o, designed to help teachers plan lessons, draft materials, and adapt content on the fly. The economics were unambiguous. Teachers who got the tool saved time on prep. They said they liked it. What the study was designed to measure, though, was what happened downstream in the students.³

The headline number was small and quiet. Students of teachers who received the AI assistant reported a drop in intrinsic motivation of about eleven-hundredths of a standard deviation.³ They rated their courses as less important, less enjoyable, and less interesting than students in classrooms whose teachers were unchanged.⁴ Average grades did not move. Among the students whose teachers had been weaker instructors going into the trial, final exam scores fell, and student confidence fell with them.³

The Number That Was Not Being Watched

A dip that small is not the number a school board looks at. It is not the number on a superintendent's dashboard. It is not the number a vendor lists in the pilot report. It is, however, the number a student registers on their fifth-hour Tuesday when the material feels flatter than it did a month ago, and the number a teacher registers as a slow, hard-to-name loss of the room.

The lead author, Wharton's Alp Sungu, said something after the paper landed that reads better the more you sit with it. "Teachers, just like students or coders, might be using AI as a crutch," he told the Hechinger Report. "Instead of doing the actual work, they're using AI to delegate the task, and that lowers the quality of their teaching."¹ The claim is not that AI made teachers lazy. It is that some part of the labor of preparation, the specific and personal act of turning material over in one's own head so a lesson can meet a specific group of children, is not incidental. It is not the exhaust of the teaching. It is the teaching.

The Room Knows What The Rubric Cannot

The most interesting thing about the Turkey trial is that the effect showed up in students, not in teachers. When you ask the teachers, they say they are working better, faster, calmer. When you ask the students, quietly, they say the class feels different, and less. That gap between what the adults report and what the children experience is the gap most classroom-technology adoptions of the past thirty years have refused to look at.

It suggests that the visible surface of a lesson, the slide deck, the worksheet, the AI-drafted exit ticket, is not the layer the students are actually metabolizing. What they are metabolizing is whether the person in front of them has spent time with them in mind. Some of that gets encoded in the pauses between sentences, in the specific example a teacher reaches for, in the way she rephrases when a hand goes up. Those are the things a language model can produce a plausible imitation of but cannot cause a teacher to have felt.

What A Different Kind Of Visibility Would Show

The instinct, after a study like this, is to talk about guardrails. The paper itself frames its findings that way and calls for better teacher training and thoughtful interfaces.⁵ That is right. It is also incomplete. What is missing from the current instrumentation of American classrooms is any honest way to see the substrate underneath the lesson. We measure attendance, standardized scores, and completion rates. We do not measure whether a student paused for six seconds before writing a sentence, whether she came back and revised it, whether she asked for help and then abandoned that help two lines later. We do not measure whether the class today felt like the last one or less than it.

That layer, the small and live evidence of what a student's thinking is actually doing, is where Koan spends its time. Not because a metric solves anything on its own, but because a school that cannot see whether the students are with it is a school that will make decisions in the dark about tools with effects like the ones the Wharton team just measured. A dashboard for the teachers is not the same as a mirror for the room.

Fall 2026

By September, thousands of teachers in American districts will be piloting AI planning assistants. Most of them will report that the tools help. They will not be wrong. The Wharton paper does not argue that the tools should be pulled from teachers' hands. It argues that the tools have effects the adults do not by default get to see, effects that lodge inside the students, and that only careful design and careful looking will surface. The next stage of this conversation belongs to schools willing to admit that the students in the room are still the truest instrument they have.

What would a school notice if it started measuring the same thing the students already do?