Mamta Motwani

Position

AI is in the making. It is not in the product.

Every learning product on this site was built with AI-assisted development. Not one of them calls a model when a learner opens it. That distinction is the whole position, and everything below is how it is held.

The useful question about AI in curriculum work is not whether to use it. It is which decisions it is allowed to make. Answer that badly and you get material that is fluent, plausible, aimed at nothing in particular, and impossible to defend to a teacher who asks why.

What follows is not a policy written to be published. It is the set of rules that actually governed the building of 24 shipped products, 1,197 pages and 16 assessments a programme — extracted from the case studies, each of which says the same thing at greater length about one project.

Six rules

  1. 01

    AI belongs in the making, not in the product.

    A learning product that calls a model at runtime is a product whose behaviour changes without anyone deciding it should, and whose users are sending their work somewhere. Neither is acceptable when the user is a child.

    In practice

    Nothing in any of the 24 shipped products calls a model. Every one runs with the network unplugged, which means no child’s work is transmitted anywhere — not as a policy, but because there is no code that could.

  2. 02

    The structure is fixed before the model is asked for anything.

    A model asked to produce a unit will produce something unit-shaped. Asked to expand a structure that has already been decided, it is useful. The difference is who decided what the thing is for.

    In practice

    In the Grade 2 heritage unit, AI expanded a fixed lesson structure into narration and turned scripted lessons into first-pass slides. It was never asked what the unit should be about, where it should end, or which expectation a lesson meets.

  3. 03

    The most valuable use of AI is adversarial, not generative.

    Generated material is easy to get and hard to trust. Using AI to attack your own work — at a volume no reviewer can reach — finds the defects that reading never will.

    In practice

    The Math at Home practice questions were run at volume specifically to find what they could produce and, more usefully, what they could not. That is how a page promising whole numbers was caught showing decimals in 88% of rounds.

  4. 04

    Never let the thing that made the material be the thing that judges it.

    A model asked whether its own output is good will say yes, fluently. Verification has to run against the finished artefact, mechanically, on rules a person wrote down.

    In practice

    Sixty-one automated checks gate every release, running against the pages a child is actually served rather than the material behind them. AI helped build the checks. It is not permitted to be the check.

  5. 05

    Some decisions are the job, and they do not get delegated.

    What a level 2 response actually looks like from an eight-year-old, where a seven-year-old’s attention runs out, what counts as above a Grade 3 ceiling — none of these fall out of any evidence base. They are professional judgements, and they have to be defensible to a colleague who disagrees.

    In practice

    Every curriculum expectation was mapped to its lessons by hand against the Ministry wording. Two were declared out of scope rather than stretched — a model asked to summarise coverage reports full coverage, because that is what the material appears to support.

  6. 06

    The right pedagogical call is often the wrong commercial instinct.

    Stars, streaks and a percentage are what sells a learning product. They are also what turns a diagnostic into a rehearsal. Holding that line is the judgement, and no tool will hold it for you.

    In practice

    The Check instrument in Math at Home has no score, allows one attempt, and reports to the adult rather than ranking the child. Every part of that is defensible on formative-assessment grounds and commercially the harder choice.

Project by project

What AI actually did, and what it was not allowed to decide. Every case study on this site carries the full version of its own row.

What is not claimed here

  • No claim that any of this makes learning better. What is measured is that specific, identified defects were caught by machine rather than by reading — a smaller claim than quality, and the one the evidence supports.
  • No figure for time saved. None has been measured, so none is offered.
  • Nothing here tells you what your own organisation permits. Tool approval and data handling are your policies to state, and they change.

Where these rules came from.

Not from a policy document. Each one was arrived at while building something, and each is argued at length in the case study of the project that produced it.

Read the case studiesThe wider practice