Child-development plan
First version 10 November 2026; updated 16 November 2026. The DfE's standards ask edtech that uses AI with children to publish how it supports cognitive and emotional development, how it will know, and when it will check. This is Monty's: design hypotheses, what will be measured, and review dates. Measures use counts Monty already has for classes (never chats), and only for schools that agree to share them for this purpose; until then, they use the public site's anonymous counts.
Hypotheses
| # | Design choice | We expect | Could go wrong |
|---|---|---|---|
| 1 | Tests decide, and AI help is built not to give the answer (the hint ladder: where to look, which idea, the exact line, then a worked example of a different problem) | Students do the thinking: most fixes come after a hint at level 1 or 2, not after the worked example | Students learn to climb the ladder fast, so hints become answers in steps |
| 2 | "AI help", not a character (no "I", no feelings, a plain badge) | Students treat the AI as a tool and turn to classmates and teachers as people; no attachment to the AI | Replies feel colder, and struggling students give up sooner |
| 3 | Break prompts and daily limits, extended only by a teacher with a reason | Shorter, regular sessions over the week, rather than long late ones | Limits cut off students mid-task, or push work to devices outside school |
| 4 | Offloading signals for teachers (large pastes, worked examples reached) | Teachers talk to students who lean on pasting or the worked example, early | Signals read as accusations; students avoid hints to stay off the list |
| 5 | No streak pressure (weekends and rest days never break a streak; no "your streak will die") | Steady use without anxiety about missing a day | Less return without the pressure |
| 6 | Help in the student's language, at their level (word budgets by age band; four languages) | Younger and multilingual students understand the help first time | Short replies leave older students without enough |
What we measure
| Hypothesis | Measure | Where from |
|---|---|---|
| 1 | Share of activities passed after a hint at level 1 or 2, compared with level 4; worked examples reached per pupil a week | Progress records (hint level at pass); class counts |
| 2 | Share of AI replies with first-person phrasing (target under 5%); reports of AI replies by students; the eval's tone criterion | The eval; reports |
| 3 | Minutes per session and per day; share of AI use late at night; how often teachers extend a limit, and why | Class counts; teachers' recorded reasons |
| 4 | Large pastes per pupil a week, and whether they fall after a teacher sees them | Class counts |
| 5 | Days active per week, against weeks with a broken streak | Progress records |
| 6 | Reading-level and language checks in the eval; checks per pass by band and language | The eval; progress records |
Results are published here, as counts across all schools, never one school's or one pupil's.
Reviews
- After the first two terms in schools (summer term 2027): the first measures, and changes to the design if a hypothesis fails.
- Every year in November, with the DfE's standards page and the crisis protocol.
- When the AI model changes: hypotheses 1, 2 and 6 again, through the eval, before release.
- No adviser on child development or children's mental health has reviewed this plan yet.