AI Must Not Do the Learning the Learner Needs to Do

A learner asks an artificial-intelligence system for an explanation, receives a clear answer and completes the task quickly. The submitted work may be better than anything the learner could have produced alone. Yet the central educational question remains unanswered: what intellectual work did the learner perform, and what capability will remain when the prompt, model or connection is unavailable?

Education cannot protect learning by prohibiting every useful scaffold. A teacher, peer, worked example, calculator, library and language tool can all reduce unnecessary difficulty. The task is to distinguish support that opens the learner's participation from substitution that removes the very thinking, practice or judgement the curriculum intends to form.

AI should therefore sit inside a learning-agency cycle: purpose, prior knowledge, a learner-owned goal, productive struggle, bounded assistance, monitoring, revision, independent performance and contextual transfer. The cycle makes efficiency answerable to formation. It asks not only whether a task was completed, but whether a person became more capable of learning and acting responsibly.

The OECD is scheduled to release the first volume of PISA 2025 on 8 September 2026 at 09:30 Central European Time, after this article's publication. More than 760,000 fifteen-year-olds in 91 countries and economies participated, representing about 33 million learners. Kenya, Rwanda and Zambia are among the first-time participants.

The official preview says the new volume will report science, reading, mathematics and computational…