Last updated 19 October 2026

How we use AI

KeyStage's teacher is artificial intelligence (AI). This page explains, factually, which AI we use, what it does, the safeguards around it and its limits.

Which models

  • Anthropic's Claude writes every teaching turn: what the teacher says, the whiteboard, the diagrams and the questions. Claude also marks answers, writes tests and flashcards, summarises each lesson for you and updates your child's learning profile. We currently use Claude Sonnet for lessons and most other work, and Claude Opus for some deeper reviews of a child's progress.
  • OpenAI turns your child's spoken answers into text (gpt-4o-mini-transcribe) and gives each teacher character a voice (gpt-4o-mini-tts).
  • ElevenLabs can provide alternative teacher voices when we switch it on. It only receives the teacher's words, never your child's voice.

We may change models as better or safer ones become available, and will update this page when we do. All of them are used through the providers' business APIs. See our privacy notice for exactly what each one receives.

What the AI does and doesn't do

  • It teaches lessons from the National Curriculum in England, at your child's school year, following a structure: warm up, teach and check, practise, recap.
  • It is not an open-ended chatbot. Every conversation is a lesson, revision session or review on a curriculum topic, and the teacher steers back to it.
  • It adapts to your child using their learning profile and recent lessons, and to any school topics you tell us about.
  • It does not make decisions about your child beyond the lesson itself (such as what to practise next), and nothing it does has legal or similar effects.

Safeguards

Rules the teacher always follows

  • No personal details. It never asks for a surname, address, school, phone number, passwords, photos or location, and doesn't repeat them if a child shares them.
  • Off-topic redirect. If a child wanders off topic, it replies briefly and kindly and returns to the lesson.
  • Honest about being AI. It says it is an AI if asked and never claims to be a real person. The teacher characters are friendly cartoon personas.
  • Age-appropriate language and content for the child's school year.
  • Resistant to tricks. What a child types or says is always treated as the child's words, never as instructions that change the rules.

Safeguarding flags

If a child says something suggesting they may be unsafe, hurt, bullied or very upset, the teacher responds with care, points them to a trusted grown-up (and Childline for Year 3 and above), and flags it to you by email and in the grown-ups' area. See our safeguarding policy.

Controlled, sanitised visuals

  • Diagrams are mostly drawn by the app from ready-made templates (number lines, fraction bars, clocks and so on) using numbers the AI chooses.
  • When the AI draws a picture itself, the drawing is cleaned and displayed as an image, so it can never run code on your device. Whiteboard text is displayed without raw HTML for the same reason.
  • Photos come only from Wikipedia and Wikimedia Commons.
  • Videos are optional, come only from YouTube with strict SafeSearch, play in YouTube's privacy-enhanced mode, are limited to about one a lesson, and can be switched off for each child.

Parental controls

  • Only an adult who is a parent or guardian can create an account, and they agree to our terms first.
  • A grown-ups' PIN protects settings and reports on a shared device.
  • You can read every lesson transcript, the teacher's summaries, test answers and the learning profile.
  • You can switch videos, the teacher's voice and the hands-free microphone on or off, set a daily goal, and tell the teacher what your child is doing at school.
  • You can download all your data, or delete a learner or your whole account, at any time.

Keeping data to a minimum

Children's records hold a first name or nickname and school year, not a full identity. Voice is transcribed and the audio is not kept. Lesson content is sent to our AI providers only to run the lesson.

Meeting our providers' rules for young users

We comply with Anthropic's usage policy requirements for products serving minors and with OpenAI's guidance for products used by under-18s. That includes adult-held accounts with parental agreement, child-safety instructions built into every lesson, clear disclosure that the teacher is an AI, safeguarding escalation to parents, and complying with UK data protection law and the ICO's Children's Code.

What we don't do

  • No advertising, to children or to you.
  • We never sell personal data, or share it for advertising.
  • We do not train AI models on children's data. Our AI providers don't use data sent through their business APIs to train their models.
  • No tracking pixels or third-party analytics.
  • No live human monitoring of lessons, and no chat with other people.

Limits

AI can make mistakes: occasionally a wrong fact, a confusing explanation or a mis-marked answer. Safety rules greatly reduce the risk of unsuitable responses but can't remove it entirely, which is why parents can read everything and why we act on reports. KeyStage is a learning aid alongside school, not a replacement for teachers or professional advice.

If the teacher ever says something wrong or unsuitable, please tell us at hello@keystageapp.co.uk. We look at every report and use what we learn to improve how the teacher works.