TL;DR

New UK research finds the school debate about AI has been aimed at the wrong target. Cheating barely registers; what worries teachers and students alike is whether the output is correct. More than half of teachers reckon the checking involved wipes out any time the technology gives back.

What the survey found

Up Learn questioned 248 teachers and 2,591 students, with fieldwork running from 1 May to 17 August 2026. Among teachers, accuracy and reliability topped the list of concerns at 73% — and 56% said verification means it may save them nothing at all. A similar share, 54%, was unconvinced the output would meet the standard of their own lessons.

Other worries ranked lower but cluster usefully: 48% flagged mismatches with exam board specifications, 40% cited over-reliance, 33% raised privacy and safeguarding, and 19% pointed at absent guidance or training. Just 8% had nothing troubling them at all.

Students converge on the same anxiety from the other side of the desk, accuracy leading for them as well on 74%. Cheating was named by roughly one in four, well behind losing independent thinking ability, at 47%, and becoming too dependent on the tool, at 65%.

How they use it explains why. Eight in ten turn to AI to have a difficult concept explained and 61% to produce summaries, against 40% for help writing essays. That is a cohort trying to be taught, not trying to cheat.

The number that travels

Two thirds of teachers use AI weekly or more, and still report the checking burden. That combination — high adoption alongside unrecovered time — is the finding that generalises well past education.

Most published AI productivity claims measure task completion. They rarely price in verification, because verification lands on the person holding responsibility for the output. Up Learn’s founder Guy Riese argues the answer is turning that scepticism into a skill — “knowing when AI has got it right, and when to look again”.

Looking forward

This is a vendor survey, and Up Learn sells an alternative to general-purpose AI, so the framing serves it. The methodology is disclosed, which is more than most.

The finding also matches what the Turing Institute told the national security community this week: the risk is people accepting flawed output without scrutiny, and the cost of scrutinising properly rarely appears in anyone’s business case.