Tutoring gains did not carry through to the broader test
Can AI tutoring improve learning beyond the problems students practice? A ten-week UK mathematics trial assigned 1,179 secondary pupils to standard hints and videos, human tutoring, or one of two AI tutoring conditions. AI-supported pupils performed better than the standard-remediation group on nearby platform measures and an intermediate quiz, with results similar to human tutoring. That advantage did not clearly extend to the externally scored Star Maths assessment. Crucially, expert educators reviewed every AI draft before delivery. Adding pupils’ historical learning data did not improve outcomes beyond question context alone. The trial supports a supervised tutoring model on proximal outcomes; it does not establish independent AI tutoring or broad achievement gains.
Why it matters for leaders
Require both a curriculum-aligned assessment and an independent transfer measure in a district pilot. Cost the human review workflow explicitly, protect practice time, and ask vendors to justify any additional student-history collection with a demonstrated benefit.
The evidence
The preregistered vendor technical report covers 58 classes in ten schools already using Eedi. Fifty-eight educators supervised responses; 94.7% of 94,385 drafts were approved unchanged.
Who was studied
Ages 12–15; UK state-funded secondary mathematics classrooms.
What the evidence cannot tell us
Ten weeks, existing platform users and changing model versions limit generalization. Human review was integral. Supervisors reported less drafting effort but greater overall workload while managing multiple chats.
Read the original sources
Summary and leadership analysis from the Week ending 2026-10-30 brief. The edition’s selection notes explain source access and evidence limitations.