How BuddyAI uses AI safely

A teacher approves AI material before a class sees it

Generated worksheets are drafts. A draft goes to a review queue and must be approved by a teacher before it can be published to pupils, and the person who approves it cannot be the person who created it. A draft whose questions are not sufficiently supported by the textbook cannot even be submitted for review; the software refuses it. The teacher remains the teacher of record, and nothing reaches a class on the model's authority alone.

The tutor gives hints, not answers

When a pupil asks for help with a problem, the tutor is instructed to give the next step rather than the final answer, so the work stays the pupil's own. Answers are drawn from the textbook material retrieved for the question, and each supporting reference is checked against what was actually retrieved. A reference the model produces that does not match a real passage is discarded rather than shown, because a citation to a page that does not exist is worse than no citation.

A pupil may also ask by voice from the classroom home. The recording is turned into text, and the reply is read aloud, by a speech service; the question itself is answered by the same hint-only tutor, with the same textbook citations and the same child-safety screen as a typed one. Nothing is heard by the software until the pupil presses the microphone, and the microphone closes when they press it again.

Material shown to a child is screened first

Text on its way to a pupil passes a child-safety check before it is stored or displayed. If that check cannot reach a clear verdict, the material is blocked rather than allowed through, so a failure falls on the safe side. Blocks are recorded for school staff to review; the recorded entry names the feature and the reason, and deliberately does not keep a copy of the flagged text.

What this is not: it is a filter on the content of messages, not a safeguarding system. It will not recognise a child who is unhappy or at risk, and no school should treat it as though it might. Concerns about a child belong in your existing safeguarding process.

We do not profile children

BuddyAI does not collect or infer wellbeing, mood, attention or behaviour data about a pupil. There is no engagement score, no focus rating and no measure of a child's state of mind, and the ban is on drawing the inference at all rather than on the wording used for it. What a pupil sees about themselves is work set and work completed, counted from records their own teachers created. Child-facing screens never rank a pupil against classmates or show a class average.

There is no advertising, analytics or session-recording software anywhere in BuddyAI, and no pupil data is used for advertising or sold to anyone. An automated test in our codebase fails the build if such a component is ever added.

Who can have an account

Accounts are created by the school, not by children signing themselves up. A pupil cannot register, so access follows from a school's own roster and the consent it holds from families. BuddyAI does not ask a child for a date of birth and does not verify age independently of the school.

A school can additionally set a minimum grade below which pupils do not reach the open-ended AI features at all, so younger classes work through their teacher instead. This is a choice each school makes: unless the school sets a minimum, no such restriction applies, and the setting is shown as unset until they do.

The model, and what it is not given

BuddyAI uses Google Gemini. A request carries the question and the textbook passages needed to answer it, and does not carry a pupil's name, contact details or marks. Where that processing happens is set out on our data residency page, including the respect in which it does not yet stay in India.