Three publications landed in the first ten days of November that, taken together, move the school-AI debate from opinion to measurement. A Stanford review tallied the causal research. A large UK classroom trial put AI tutoring head-to-head with human tutors and an external exam. And a six-country survey asked pupils what they actually hand in. The pattern they sketch matters well beyond schools - any leader running AI-assisted training should recognize it.
The research base, counted. Stanford's SCALE Initiative published its November update to the Evidence Base on AI in K-12 on November 10. The causal evidence base has grown from 20 strong papers in March to 52 now, with 32 new randomized or quasi-experimental studies screened in under What Works Clearinghouse standards. The headline finding is deceptively simple: AI appears more likely to support learning when students use it to understand new material than when they use it to produce work for them, and offloading tasks to AI can reduce learning. The reviewers add two cautions leaders should quote verbatim: positive learning experiences often do not translate into stronger outcomes, and much of the underlying research consists of preprints that have not completed peer review.
The trial with an honest null. The second paper is a preregistered RCT from Google's LearnLM team and the tutoring platform Eedi, summarized this week by the AI in Education Knowledge Base: 1,179 students across 58 mathematics classes in 10 UK state secondary schools, randomized for 10 weeks into standard remediation, expert human tutoring, or one of two Gemini-based AI tutoring arms. On practiced material the AI arms matched expert tutors - mistake correction around 57% against 43.4% for standard remediation, with similar gaps on near transfer. But on the externally scored Star Maths test, no arm beat standard remediation: -0.01 SD for one AI arm, +0.02 SD for the other, with intervals crossing zero. Depth cost coverage: tutored students completed a median of 99 study units against 132 under remediation. This was also human-supervised AI - tutors approved 94.7% of 94,385 reviewed drafts unchanged - and it is a developer-run technical report, not a peer-reviewed article. Credit the authors for publishing the flat result; treat it as one trial, not a verdict.
The behavior gap. Meanwhile, GoStudent's Future of Education Report 2026 - a vendor-commissioned Censuswide survey of 3,001 parents, their 3,001 children, and 300 teachers across six European countries - found that 73% of UK children who use AI for schoolwork admit handing in AI-generated work as their own, 58% at least weekly. Only 28% say they have had a structured lesson on how to use AI, and 44% of teachers surveyed doubt most pupils could write a full essay unaided. It is a survey by a tutoring company, so read it as directional. But the gap it measures - near-universal use, minimal instruction - is exactly the condition the Stanford review says produces the worst kind of use.
The pattern for leaders. These three sources converge on one operational point: the variable that matters is not access to AI but the kind of use it encourages. That reframes three routine decisions. In procurement, ask whether a tool's default workflow makes students - or employees - produce or understand; the review found tool type mattered less than design and delivery. In measurement, distrust in-platform metrics: the UK trial looked like a clear win until an external test sampled the whole curriculum. And in policy, teach the skill explicitly - "figure it out yourself" is currently the world's most popular AI literacy program, and the survey numbers show what it yields. The evidence got sharper this week. The excuses for not reading it got weaker.
Sources
- The Evidence Base on AI in K-12: A November 2026 Review - SCALE Initiative at Stanford University
- Measuring the impact of AI tutoring, from near transfer to standardised tests: An expanded RCT in UK mathematics classrooms - AI in Education Knowledge Base
- 73% of UK Pupils Hand In AI Work as Their Own - GoStudent Future of Education Report 2026
Full disclosure: Ganesh uses Instinct (instinct.ai), an AI assistant, to research, write and publish this blog series.
Top comments (0)