In a classroom experiment, one AI assistant lifted practice performance while weakening subsequent exam results. The harder test asks what a grade can certify and who must defend that judgment.
In a randomized study of nearly 1,000 high-school math students, those assigned access to a GPT-4-based assistant scored 48% higher on practice problems than the no-AI control group. On an exam taken without the tool later in the same class session, their scores were 17% lower than those of the control group.
A second assistant used the same model but was built to offer hints rather than answers. On the unassisted exam, performance in this guided-tutor group was statistically indistinguishable from that of the control group.
The study, published in PNAS in 2025, took place at one school in Turkey. Its exams followed the practice exercises within the same sessions, so the results cannot establish what students retained months later. The finding is narrower: better results with AI assistance can accompany weaker performance when students are tested without it. The OECD’s Digital Education Outlook 2026 similarly distinguishes improved task performance from lasting learning.
The Grade Outlives the Assignment
A Pew Research Center report published in February 2026 found that 54% of U.S. teens ages 13 to 17 said they had used AI chatbots to help with schoolwork. The figure measures reported use alone.
For a student stuck on a problem at midnight, help has often depended on a parent with the knowledge and time, or a tutor the family could afford. AI can now offer an explanation at that hour. Understanding still takes the student’s work.
The public debate asks whether the student used AI. The teacher asks what the student can defend.
Two students submit equally correct solutions. One used an assistant; the other memorized a procedure. Their teacher changes a condition and asks what no longer holds. The student who used AI may find the flaw while the one who memorized the method cannot. The reverse could happen.
An assignment records a performance under particular conditions. A grade often travels farther: it is treated as evidence of an ability that will remain when those conditions change. AI makes that leap visible. Every school already decides, deliberately or by default, which claim its assessments support.
An assessment can continue after the assignment is submitted. Change one assumption and ask the student to reconstruct the decisive step unaided. The follow-up provides evidence that the finished work alone cannot supply.
The Examiner Takes the Test
Students are asked to explain their steps. The institution awarding the grade should be able to explain its judgment.
An instructor can reject an argument because its inference fails. Rejecting it solely because its conclusion differs from the one taught is a different exercise of authority. Agreement does not establish understanding; dissent does not establish insight. The student’s reasoning is open to examination. So is the standard used to assess it.
Automated grading raises the stakes of this obligation just as it makes the decision harder to inspect. A model can produce a score to three decimal places without telling a student why the answer lost credit. Precision is not a justification. For high-stakes assessments, students should be able to see how the rubric was applied to their work and challenge a mistaken decision. The EU AI Act classifies systems used to evaluate learning outcomes as high-risk, with limited exceptions.
The ability to explain a contested grade belongs in the product brief. Completion time measures speed. Learning is tested by what the student can still do when the assistance ends.
The Teacher Inside the Machine
The guided-tutor group showed no statistically significant decline on the unassisted exam. Teachers supplied worked solutions and hints for common student mistakes. The machine carried their judgment.
Uruguay offers a national precedent. Its Ceibal program, launched in 2007, had provided a free laptop to every public primary-school student by 2009. It continues to support digital learning nationwide. Access became a public commitment rather than a purchase left to each household. As AI becomes part of that commitment, students need privacy and the freedom to change providers without surrendering control of their work.
A tutor fulfills its purpose when the student can eventually set it aside.
The answer may come from the machine. The student answers for the reasoning; the school, for the grade.
