CurricuLLM LogoCurricuLLM
SchoolsAI tutorFeaturesPricingTraining HubDevelopers
An AI tutor belongs in the middle of the learning, not at every step
17 October 2026

An AI tutor belongs in the middle of the learning, not at every step

Dan Hart

Dan Hart

CEO, Co-Founder, CurricuLLM

A school AI tutor should not be present at every step of learning. Used in the middle, as a hint or another example while a student practises, it can help. Used at the start or the end, it can skip the thinking.

What we mean by an AI tutor is set out on the AI tutor for schools page.


When the help should arrive

A consensus report from education researchers at the University of Minnesota asked 30 researchers to look at how generative AI might affect 65 learning processes. The interesting conclusion was about timing.

When a student first meets an idea, an immediate answer can skip the work of building an understanding. During practice, an AI tutor can offer another example, a hint, or feedback that strengthens what has started to form. When the student has to use that knowledge on a new problem, they need the chance to think without the tool again.

That is a better design question than "should students use AI?" A good tool asks the student to try first, helps when they are stuck, and then gets out of the way. The authors are clear that this is expert consensus, and that classroom evidence still has to catch up. The question to hold onto: are we designing the tool to improve the work a student submits, or to strengthen what they can do alone next time?


Less preparation time, same quality of materials

The Education Endowment Foundation published an independent randomised trial of Aila, Oak National Academy's lesson-planning assistant, with primary teachers in England.

Teachers using Aila spent 24 per cent less time preparing lessons and resources, about 49 minutes a week. Materials assessed independently were no worse. Teachers who used it more saved more time. The largest gains were for less experienced teachers, and for those less confident in their subject or their pedagogy.

The trial write-up is here. It is evidence about planning time and resource quality, not about whether students learned more. That is still a result worth having, and it is the right shape of evidence: a trial, a comparison, and a limit stated in the open.


A proof you cannot explain

OpenAI has released mathematical results from an internal model, with computer-checkable versions of many of the proofs. A formal check can say whether the steps follow. It cannot say which idea made the proof work, how it connects to earlier work, or what a person should try next.

An independent advisory group has argued that labs should help the mathematical community build that understanding: share how results were found, record the attempts that failed, and support work that lets other people explain the proofs and build on them. OpenAI says it consulted the group and is sharing more of the process.

A result that can be checked is valuable. A result that someone can explain, teach and use to find the next thing is more valuable. The same standard should apply to a school AI tutor. A correct final answer that the student cannot reconstruct is a weak kind of success.


An audit needs two keys

Anthropic says six researchers from METR now have access to internal tools roughly comparable to its own risk teams, and can publish their assessment without Anthropic's approval, with predefined categories of sensitive information redacted. It is the deepest of three access tiers the company has described for outside evaluators.

The update is here. "Independently evaluated" means very little until you know two things. Could the evaluator see enough to understand the system, and could they say what they found? A school buying a school AI tutor should ask both questions of the vendor, not just for a badge.

One in ten Australian businesses have halted an initial AI rollout over security concerns, according to an Adobe survey reported by news.com.au. The same research points to missing pathways: no pre-approved route, no standard security review, no settled way to integrate the tool, no budget line. Reviews should produce mitigations. Stopping everything is what happens when the process was built for a different kind of technology. Schools are in the same position, with a higher duty of care. The answer is a path through the review, not a pile of tools staff are already using without one.

Back to all posts
CurricuLLM Logo
CurricuLLM

AI for schools

Product

FeaturesPricingUse CasesSchoolsAI tutorDevelopersFAQ

Resources

Training hubSupportBlogResearchEvidenceEventsPress

Company

About usTrust & SafetyPrivacy policyTerms of useStatusContact