Steven Watson | EdgeLab | Research and Scholarship | 19 September 2026

An AI assistant can help someone finish a mathematics exercise while leaving an educational question unanswered: what can that person do next? A correct answer is useful. So is the ability to explain it, notice an error and tackle a different problem. A learning tool needs to make room for those achievements.

Designing Against Deskilling, a new preprint by Sebastian Maier and colleagues, offers a useful occasion to reconsider how we judge assistance. The question for this review is how a successful interaction might also become an opportunity to develop understanding.

What the study reports

The preregistered experiment involved 704 UK adults practising fractions. Five randomised groups compared no AI, an assistant alone, feedback about participants’ use of assistance, symbolic points favouring less extensive help, and both interventions. Full solutions required an explicit request. Feedback described participants’ use and prompted reflection before further practice. A six-question unaided test followed.

Feedback reduced requests for complete answers and improved test performance in the preregistered analyses. AI access alone did not significantly change test performance; points showed no clear benefit on either primary outcome. The feedback effects passed one-sided tests, which assess a prespecified direction. However, their two-sided 95% confidence intervals touched or crossed the no-effect value. For test performance, the odds ratio was 1.51, with an interval of 0.98–2.33. That is not a 51% increase in marks. These were immediate outcomes, not evidence of lasting protection from skill loss. Results and statistical tables.

A promising finding that needs careful interpretation

Randomisation and advance specification of the tests strengthen the case for taking this work seriously. They help separate an intervention’s contribution from a researcher’s freedom to select an attractive result afterwards. They cannot remove uncertainty from an estimate or make one setting represent every classroom.

My reading is therefore cautiously positive. A result close to a conventional statistical threshold deserves replication and attention to its practical size. The fair response is neither to dismiss it nor to announce a settled solution. A useful next question is whether an independently conducted study would find a similar benefit, using a clearly specified comparison and an assessment teachers regard as educationally meaningful.

We should also distinguish learning something new from retaining an existing skill. To show that a design prevents deskilling, I would want evidence about what people could do beforehand, how their opportunities to practise changed, and what they could still do later. That requires a timescale appropriate to the claim. A term of learning, a period without the tool and unfamiliar problems would each reveal something different.

An AE reading: the answer now and the capacity to act later

EdgeLab’s autopoietic ecology approach distinguishes the conditions that make an activity possible now from what that activity does to future capacities. Applied to education, this invites us to follow both the completed task and the learner’s changing ability to participate. This is an interpretive framework, not a theory established by the experiment.

Consider a hypothetical pupil who asks an assistant to explain why dividing by a fraction can make a number larger. The assistant supplies an example. The pupil then draws a picture, challenges the explanation and tries another case with a classmate. The AI response has become material for activity that the pupil helps to organise.

Now imagine the same response being pasted into a worksheet and accepted as evidence of understanding. The text is identical, but its educational role differs. To assess what happened, we would need to observe the pupil’s explanations and subsequent decisions, rather than infer understanding from the presence of correct prose.

This connects with EdgeLab’s work on meaning mediation: an expression acquires significance through interpretation and use. An answer can become a question, a justification, a mark or an instruction. Each transition deserves scrutiny. Who interprets it? What evidence makes that interpretation reasonable? What can the learner now explain or challenge?

The responsibility extends beyond the learner

An ecological interpretation should also question the surrounding arrangements. Suppose a school rewards rapid completion while telling pupils to take time over their reasoning. Asking an individual to reflect will not resolve that contradiction by itself. Lesson time, assessment expectations and the availability of human help also deserve attention. This is a hypothetical institutional example, not a finding from the preprint.

Nor should educational design treat every request for help as a failure. A worked example may give a stuck learner a starting point. Accessible support may make participation possible. The aim should be to clarify which capability a task is intended to develop and whether the available assistance helps that particular learner practise it. Making a task harder is not, on its own, an educational achievement.

There is a plausible simpler explanation to test alongside the AE account: a prompt may merely redirect attention towards the upcoming test. That could still be useful. To assess the additional value of an ecological interpretation, researchers would need to investigate the relationships and institutional conditions it identifies, rather than rename an individual learning effect.

What would be worth trying next?

For a classroom investigation, I would combine delayed assessments with short conversations in which learners explain their choices. I would compare using feedback about assistance with equally prominent general study advice. I would also examine whether learners can question the assistant’s account of their work: a chat record cannot directly reveal everything someone has understood or done on paper.

Teachers’ time and pupils’ experience should enter the evaluation too. Does the design create useful discussion, or another demand to monitor? Does it help learners seek appropriate assistance, including human assistance? These are questions to investigate, not benefits to assume.

The practical takeaway is to ask two questions whenever AI supports learning: did it help with this task, and what evidence would show that the learner gained a capability worth carrying forward? Designing for both is a more demanding ambition than producing a convincing answer.

Publication reviewed: Maier, S., Schwabe, K., Schneider, M., and Feuerriegel, S. (2026). “Designing Against Deskilling: Metacognitive Feedback Reduces Cognitive Offloading to LLM Assistants.” arXiv:2609.20143v1, submitted 17 September 2026. This review draws on the full preprint; peer review is not established by its arXiv posting. Source and version history.