What The Graph Kept
On July 7, the Hechinger Report summarized a working paper from UC Irvine and McGraw Hill. Across eleven quarters of ALEKS data, study time on math word problems fell steadily. On graphing problems, which cannot be pasted into a chatbot, nothing changed. The gap between the two kinds of problem is a mirror. It shows the classroom exactly where the shortcut is, and, buried in the results, when it disappears.
Three days ago, on July 7, the Hechinger Report ran a Proof Points column with a bracing headline. Faster solutions, lower test scores.1 The piece summarized a working paper released last month by Sina Rismanchian at the University of California, Irvine, and four researchers at McGraw Hill.2 They had done something the AI-in-education conversation rarely has the data to do. They looked at eleven quarters of student behavior on ALEKS, the adaptive math platform that serves more than four million students a year, and compared what happened before ChatGPT to what happened after.
The finding is easy to state and hard to unhear. On problems a student can copy and paste into a chatbot, meaning word problems, study time and eventual test performance dropped, quarter after quarter. On problems the student cannot easily paste in, meaning graphing problems, which require dragging points on a plane, nothing changed.1 The gap between the two kinds of problem is not aging or the pandemic or a drift in curriculum. It is the presence, and the absence, of a shortcut.
The Number That Names It
Across the eleven quarters after ChatGPT's release, study time on AI-susceptible problems declined 2.8 percent per quarter for college students, cumulating to 26.9 percent. Among high schoolers the decline was steeper, at 31.3 percent. Middle schoolers slipped 9 percent. Fifth graders showed no change.2 The older the student, the more available the shortcut becomes, and the more visibly the learning departs.
The researchers named the departure carefully. Cognitive surrender. Not delegation, they were careful to distinguish. Delegation is asking a tool to do a sub-task and then checking the result. Surrender is the adoption of an AI-generated output as one's own answer with minimal scrutiny.3 A generation of teachers has been taught to worry about cheating, which is a moral frame. Rismanchian and his colleagues put the same behavior under a different frame. It is not primarily a question of honesty. It is a question of whether any thinking happened at all.
What Disappeared Under Supervision
Buried in the results is a line that should stop every district planner in the country. The learning decline the researchers detected disappeared when students were being supervised.1 Under a proctor, in a room with a teacher, on an assessment, the effect vanished. It was not a claim about capacity. The students could still think. They simply did not, on their own, in their bedrooms, at ten in the evening, when the tool was open and the assignment was due.
The variable that changes the student's behavior is not the model, not the prompt, not the assignment design. It is whether anyone is watching. That is a very old finding about human beings, dressed up in new clothes. It is also the piece of the story a school can do something about.
The Witness, Not The Warden
A school hearing the word supervised will reach for the wrong tool if it is not careful. Cameras. Lockdown browsers. Detectors. Kiosk mode. The instinct is to become the warden and treat the child as a suspect. That posture works for the length of a proctored exam. It does not work across four years of a student's development, and it does not build anything she takes with her when she leaves.
What Rismanchian's supervised condition actually tests is subtler and kinder than surveillance. It tests presence. When someone in the room cares whether the student is thinking, the student thinks. The question for the next school year is whether the classroom can become such a room again, now that the writing and the problem-solving mostly happen on a screen only the student can see.
One way, and it is the way we have been building at Koan, is to design the workspace so the process leaves a trace worth reading. Not a log of keystrokes to catch a cheater. A record of the moments that mattered. The pause before the prompt. The answer taken and then revised. The wrong turn the student noticed on her own. The place she pushed back on the model, and the place she quietly accepted a sentence that was not hers. If a teacher can see those marks the next morning, she is not absent from the ten o'clock room. She is present in it, not as a warden but as the reader the student was always writing for.
Rismanchian's paper explains why the behavior went unnoticed so long. The answer, on the screen, was still correct.1 The output looked fine. The process was hollow. A school that measures only outputs will not know when it has taught nothing. A school that learns to see the process, in the workspace itself, will see the surrender long before it shows up on the test.
If a student's thinking looked the same to you whether she had done it or the model had, what would you want your classroom to be able to show you tomorrow morning?
References
PROOF POINTS: Faster solutions, lower test scores: How AI is eroding math skills
The Hechinger Report · July 7, 2026
Faster Completion, Less Learning: Generative AI Reduced Study Time on Math Problems and the Knowledge They Build
arXiv (Rismanchian et al.) · June 2026
'Cognitive Surrender': Faster Solutions, Lower Test Scores Show How AI is Eroding Math Skills
KQED MindShift · July 2026
Sources cited in order of appearance. Click any inline number to jump.