A seven-month trial of AI tutoring tools in UK schools has produced results that are encouraging in a specific, limited way. Students using the software showed higher engagement with practice material, and teachers reported that it freed time for one-to-one support. The tools did not, on their own, produce the kind of learning gains that would justify replacing instructional time.

The findings, discussed by educators and researchers this week, fit a pattern emerging across several countries: AI tutors are useful when they extend a teacher’s reach, and unreliable when they are asked to carry a lesson alone.

What the software does well

Adaptive practice is a good fit for current systems. A model can generate varied questions at an appropriate difficulty level, give immediate feedback, and keep a student working through material they find tedious in worksheet form. These are real benefits, particularly in large classes.

The same systems struggle with diagnosis. When a student holds a persistent misconception, a tutor that responds to the surface form of an answer may reinforce it rather than surface it.

The teacher’s role changes, not disappears

Teachers in the trial described spending less time marking routine work and more time intervening where students were stuck. That is a plausible productivity gain, but it depends on the tool being accurate enough that the teacher can trust its output, and on class sizes that leave room for the intervention.

In settings where those conditions do not hold, the same software adds a layer of supervision work rather than removing one.

The policy implication

Procurement decisions are often framed as a choice between AI and traditional instruction. The trial suggests that framing is wrong. The relevant question is what mix of teacher time, software and class structure produces better outcomes per pound spent, and the answer is likely to vary by subject, age group and school.