Teachomatic what to automate, and what not to

What We Still Don't Know

Most writing in this area is organised around what is known. This page is organised around what is not, because the gaps are larger than the findings and they are the part that determines how confidently anyone should be talking.

Reviewed August 9, 2026.

For a workplace comparison of the underlying concept, Monitask’s discussion of cognitive offloading shows how the same behaviour is framed outside education.

The pattern behind the gaps

Before the list, the reason it looks the way it does.

Research answers cheap questions first. Cheap means: quick to measure, easy to define, visible within a term, fundable. Expensive means: long-term, hard to define, requiring years of follow-up, and awkward for anyone with a product.

For an external perspective on education research and evidence, see Nature.

So the questions with answers are the ones that were easy to ask, and the questions parents ask first are the ones nobody has funded. That is not a conspiracy; it is what a young field looks like. But it means the shape of the evidence base tells you about research economics rather than about importance.

The five open questions

Long-term effects on learning. Almost nothing. Studies run for weeks or a term. Whether gains persist, whether they transfer, whether anything looks different after three years — unstudied, because a three-year study takes three years.

Effects on independence and metacognition. The most commonly voiced worry: does routine assistance erode the ability to work through difficulty alone? Plausible on theoretical grounds, difficult to measure, and near-absent from the literature. Anyone who tells you it definitely does, or definitely does not, is going beyond the evidence — including anyone using it to argue with a class.

Effects on inequality. Genuinely open and it could go either way. Adaptive support might most help students with least help at home. Or better-resourced schools might extract more from the same tools, as has happened with every previous technology. Both stories are coherent, neither is settled, and it matters more than most of what is measured.

Effects on wellbeing and social development. Barely touched. The one adjacent authority — the American Academy of Pediatrics' January 2026 statement on digital ecosystems — deliberately folds AI into media use generally and shifts the framing to displacement and context rather than hours. That is sensible guidance and it is not evidence about learning.

What happens to teaching itself. If prep and feedback are increasingly drafted, what does the job become over a decade, and what happens to the skills that were maintained by doing that work? Nobody is studying this, and teachers are the ones who will find out.

What we do know, briefly

For balance: teacher-side time savings are reasonably supported and modest, purpose-built tutoring systems have produced large gains in controlled trials, and detector accuracy claims do not hold up evenly across student populations.

Three findings. Against five structural unknowns, most of which are more important.

How to act inside uncertainty

Not paralysis, and not pretending. Four principles.

Match confidence to stakes. Being wrong about lesson planning costs an hour. Being wrong about a detector costs a student their record. Apply scepticism in proportion.

Prefer reversible decisions. A tool you can stop using is different from restructuring assessment around one. Where the evidence is thin, keep the exit.

Watch your own classroom. With five open questions at population level, local observation is not a poor substitute for research — it is the only evidence you have about your students. It is weak evidence and it is yours.

Say "we don't know" out loud. To students, to parents, to colleagues. It costs less credibility than it seems and buys a great deal, because the alternative is being caught overclaiming later.

What would change our minds

Worth stating, because a position that cannot be moved is not a position.

Multi-year follow-up showing gains persisting or dissolving. Well-designed studies on independence and metacognition, in either direction. Evidence on distributional effects across school types. Independent evaluation of detection tools against a published standard — which does not currently exist anywhere.

If any of that arrives, these pages change and the change gets logged rather than quietly edited.

The short version