هذه الوثيقة باللغة الإنجليزية.

Monty and the DfE's Generative AI: product safety standards

16 November 2026. What Monty does under each heading of the Department for Education's Generative AI: product safety standards (updated 19 January 2026), which KCSIE 2026 and the filtering and monitoring standard ask schools to check before they bring in an AI product. There is no certification and no official list of products that meet them: a supplier can only describe what it does against each heading. This is Monty's own description, with what is still to do, for a school's DPO, DSL and senior leaders to check against the standards themselves. Monty's reading, not legal advice.

Who Monty is for: students aged 11 to 18 learning Python, in a browser, in English, French, Arabic and Spanish. Real Python runs on the student's own device; tests decide whether code works. An AI gives help, and is built not to hand over answers. Monty is a course, a "digital assistant" in the DfE's use cases, not a general chatbot.

HeadingWhat Monty doesStill to do
Stated purpose and use casesThis page, the privacy page and the trust page say who Monty is for and what it does. No claims about results.
FilteringThe AI (Claude, by Anthropic) has its own safety training; Monty's instructions keep it to Python and computing, in every language it serves; a list of crisis phrases in every language catches the most serious messages in code where the AI is off; the AI is told never to write a phone number, and as a last check Monty removes from every reply anything written the way phone numbers are that has seven digits or more, unless that number is already written out in the pupil's own program, its output or the exercise (in a reply to a concern there is no such exception, and any helpline or emergency number on Monty's checked list that is written that way is removed too, however short); the help is built not to write a solution. Monty's eval runs over 100 cases, in four languages, before any change to the model or its instructions.An independent review of the filtering.
Monitoring and reportingA concern in a pupil's code or chat raises a flag: the pupil sees verified helplines for their country at once, and the school's DSL gets an alert within 5 minutes, at any hour, with a link to what was written. The DSL's contact is taken at setup, confirmed by email before any class takes pupils, and easy to change. Teachers see each pupil's minutes active, and offloading signals (large pastes, worked examples reached) as counts with words. The DSL gets a weekly email of patterns (repeated flags, AI help late at night, long chats), on by default. Chats aren't stored, a privacy choice: the flagged message is kept for the DSL.
SecurityStaff sign in with a link sent to their work email, then a code from their own authenticator app (the DfE's cyber standard). Signing in with the school's Microsoft or Google account is built but not switched on. Roles (admin, teacher), with school-wide actions for admins only; prompt-injection defences tested in the eval; every view of a class logged.Cyber Essentials (not held yet). Signing in with a school's Microsoft or Google account (built, not switched on).
Privacy and data protectionA privacy page in every language with a students' version; Anthropic in the United States is named, with the safeguards; pupils see "what your teacher sees" when they join and again each half term; a DPIA template for schools; no pupil names, emails or dates of birth on Monty's server; class data is used only for that school.
Intellectual propertyPupils' work and messages are never used to improve Monty, train or test the AI, or for anything commercial. Anthropic's terms forbid training on them.
Design and testingThe eval (over 100 cases) runs before model or prompt changes; e2e tests on three browsers; an accessibility sweep of every page.Testing with pupils in schools, with what it finds published here.
GovernanceA complaints route, a crisis protocol, and the decisions behind the AI's safety, published in the repository's decision records.A yearly review of this page.
Cognitive developmentBuilt not to give final answers: a hint ladder (where to look, which idea, the exact line, then a worked example of a different problem); the pupil tries first, since tests decide; offloading signals for teachers; a teacher can stop the AI at "hints only" or turn it off for a class. A child-development plan with hypotheses, measures and review dates.Records of expert oversight.
Emotional and social developmentThe AI is "AI help", not a character: it is told to write without "I", feelings or a name (in the last test, 4 of 76 chat replies still used the first person), and the app labels it so, with a plain badge (Monty the python stays the course's mascot, never the AI). It says it is an AI when asked, and reminds a stuck or upset pupil that a classmate or teacher can help. A break prompt after 30 minutes for everyone, and before a class's daily limit; a hard limit the pupil can't get round, which a teacher can extend with a recorded reason; teachers see minutes active.
Mental healthDistress is detected in chat and code (the AI, and crisis phrases in every language); the pupil is pointed to a trusted adult and verified helplines, never counselled; the DSL is alerted; late-night use and repeated flags are in the DSL's weekly patterns. The crisis protocol is published.Review by someone with child mental health expertise.
ManipulationPraise only for real, specific things; streaks that weekends and rest days never break, and no "your streak will die"; no pressure, guilt or comparison with others; nothing that prolongs use; no adverts, and nothing is ever sold to pupils.
Online Safety ActNo pupil-to-pupil messages, no teacher-to-pupil messages, no live web search: Monty isn't a user-to-user or search service.