The Awe That Wore Off

On Monday, July 28, the Hechinger Report's Proof Points column published a piece with a headline that will not have been read by most district leaders and should be. Teachers save time with AI. Their students may pay the price.¹ Behind the headline is a study posted to SSRN on June 25 by Alp Sungu of the Wharton School, Benjamin Lira of the University of Pennsylvania, and Angela Duckworth.² It is one of the first large randomized trials of a generative AI teaching assistant in real classrooms. One hundred ninety three teachers. Two thousand eight hundred sixteen students. Fourteen middle and high schools in Turkey. Ten weeks of the spring 2025 semester. Fourteen thousand one hundred ninety eight student-course observations.³

Teachers were randomly assigned to one of three arms. Continue teaching as usual. Get access to a GPT-4o teaching tool loaded with the Turkish Ministry of Education's curriculum. Get the same access plus weekly reminders and usage feedback.³ The tool worked. The teachers used it. What it did to their students is the finding worth pausing on.

Students whose teachers received the AI assistant reported lower intrinsic motivation, by about 0.11 standard deviations. They rated their courses as less important, less enjoyable, and less interesting than students in the control group. Average academic performance across the whole sample did not move. But among students of teachers who had been lower performing before the experiment began, both performance and student confidence dropped meaningfully.²³

The line from the paper worth rereading is this one. "One possibility is that the initial awe of discovering the power of AI to instantaneously respond to any request later gives way to an awareness of its unintended negative effects."² The authors are careful. They call what they found a principal-agent problem. The teacher and the student, in the presence of a labor-saving tool, are not the same person and do not want the same things.²

Where The Time Went

The paper's usage data explains where the trade lives. Sixty six percent of the AI conversations teachers logged were about lecture materials, homework, exams, syllabus design, and student reports. Sixteen percent were about instructional support proper: differentiation, correcting a misconception, giving a student feedback, helping with stress.⁴

Two thirds of the labor the tool saved was upstream of the room. Six sevenths of what it did not save was inside it. The teacher went home to a lighter Sunday night. The student walked in on Monday to a class that felt, on the survey, less like her own.

There is nothing surprising in that ratio if you have watched a busy teacher use any productivity tool. She uses it first for the thing that costs her the most sleep. Materials generation costs sleep. Reading a student in the middle of a paragraph does not cost sleep in the same way. It costs presence, which the tool does not know how to save.

The Second Order

The second-order finding is the one that will take a year to be noticed. Teachers, on average, did not change their beliefs about whether AI helped or hurt student learning during the trial. But the ones who had already been using AI heavily before it began became, over ten weeks, more pessimistic. The ones who had barely touched it became more optimistic.²

That is a shape worth naming. Optimism about AI in schools is not evenly distributed across experience. It concentrates in the early hours. It thins with time on task. The people writing your district's policy this summer are, disproportionately, in the first group. The people who will be teaching with the tools in October will move, quietly, into the second.

What The Trial Did Not Measure

The trial measured motivation, confidence, and grades. It did not measure the thing everyone in the debate is actually arguing about, which is the shape of a student's thinking. It could not have. That shape does not fit on a Likert scale, and it does not appear in a standardized final exam. It shows up, if at all, in the record of the work itself. In the pause between two sentences. In the revision. In the moment a student left one tab and came back to another with something new to say.

The finding this study offers is that a productivity gain for the teacher, on its own, produces a motivation loss for the student. What no study of this shape will ever tell a district is what a particular class of students in a particular building actually did during the hour. Only the trace of the work can do that.

That trace is the layer we work on at Koan. Faster lesson plans on the way in and faster feedback on the way out are useful. They are not, on their own, an answer to what the trial found. The answer sits on the other side of the room, in the middle of the hour, in the record of what the student is quietly doing while the teacher's Sunday night gets shorter.

What The Fall Will Actually Ask

By fall, most American districts will have adopted at least one of the teacher-facing AI tools the summer produced. Lesson plans will be faster. Feedback drafts will be quicker. Sunday nights will be lighter. The Turkey trial does not argue against any of that. It argues, quietly, that the second half of the picture has to be measured too, or the gain gets counted and the loss does not.¹

The awe wore off, in the paper, at about ten weeks. The teachers who had used AI most learned to distrust it first. That is a small, human timeline. It is also the one most districts have not yet built any instrumentation for.

If your teachers are saving three hours a week by fall, what will you be measuring on the other side of the room to know what those hours cost?