A Human-Centred Rule for Learning with AI

The opening question was whether better work produced with generative AI necessarily means better learning. The evidence gives a conditional answer. This final chapter brings together the relationship among AI access, cognitive effort, use mode, feedback, assessment, and human agency, then turns it into a practical rule for learners, educators, designers, and institutions. The rule is simple to state but demanding to apply: use AI to increase the productivity of human thinking, not to remove the thinking a person is meant to develop.

That rule does not divide tools into good and bad categories. It asks a more useful question: what capability is this activity meant to build, and where does the work that builds it remain? A system can be helpful when it explains, questions, compares, or gives a learner a way to inspect a difficulty. The same system becomes educationally dangerous when it supplies the diagnosis, reasoning, synthesis, or final judgment that the learner needs to practise.

One model: where the work stays

The negative and positive findings across the evidence are not two competing verdicts about artificial intelligence. They are observations of different arrangements of work. In the mathematics trial reported by the OECD Digital Education Outlook 2026, unrestricted GPT Base access raised practice performance by 48% but was followed by performance 17% below the no-GenAI group on closed-book exams. The GPT Tutor, configured to avoid direct answers and support learning, did not create the same practice boost, but its users performed about as well as students who studied without AI.

In the randomized college experiment by Contractor and Reyes, AI access produced a different pattern. Students learning unfamiliar material with access scored 0.27 standard deviations higher on the immediate test and remained 5.1 percentage points ahead about a week later. The experiment’s usage evidence helps explain why: students who used AI for explanations and worked with the material showed more persistent unaided gains than those who used it to generate the work. Neither finding cancels the other. The mathematics study shows what can happen when a tool carries too much of the practice; the college study shows that access can support learning when the learner’s activity remains engaged with the material.

The feedback experiment adds a second dimension. Direct AI critique improved the quality of a revised argument in the moment. Reflective and hybrid designs, which required self-evaluation and comparison, produced stronger delayed AI-free transfer. The difference was not that one model was intelligent and the others were not. It was that the designs placed evaluative judgment in different hands and at different times.

These results support one causal model. AI changes the available path from a task to an outcome. If it supplies a finished answer, it can increase visible performance while reducing retrieval, diagnosis, explanation, and evaluation. If it provides a bounded prompt, a counterexample, a question, or a critique that the learner must assess, it can increase the productivity of those same activities. The outcome is then shaped by four connected variables:

  1. The target capability. Is the learner meant to remember a concept, construct an argument, practise a procedure, search efficiently, or use AI itself as part of the skill?
  2. The location of cognitive work. Does the learner retrieve, explain, compare, and decide, or does the system do those things first?
  3. The evidence of independence. Can the learner retain and transfer the capability after the tool is removed?
  4. The surrounding design. Do feedback, assessment, incentives, teacher involvement, and data practices reward understanding or only a polished result?

These variables interact rather than operate as separate checkboxes. A student may use AI to generate an explanation, but if they then test it against a source, restate it, and apply it to a new case, some of the learning work remains theirs. A student may use a “tutor” that asks questions, but if the questions are so leading that the learner only selects the next supplied step, the apparent dialogue may still substitute for reasoning. A process can be labelled “AI-supported” while still being either active or passive.

The central distinction is therefore not between digital and human work. It is between support that makes human capability stronger and support that makes human capability unnecessary for the task. That is why the performance–learning distinction remains essential even when AI use is permitted. The objective is not to preserve every old step for its own sake. It is to preserve the steps that cause the capability the learner will later need without assistance.