AI safety policy

WonderPath is developed and operated by Clove9 (the trading name of Prem Dev, sole trader, until WonderPath is incorporated), Office 128, 18 Young Street, Unit LGE, Edinburgh, EH2 4JB, Scotland, United Kingdom. This policy says exactly how we use artificial intelligence with children, what the helper will and will not do, which safety layers run on every message, what you can see and control, and where the limits are. We keep it to what the product does today. When we promise something we have not built yet, we say "will".

1. The rule: the child does the thinking

The best evidence on children and AI (summarised in our published learning doctrine) shows one clear harm: a chatbot that gives answers on demand raises practice scores and lowers what the child can do alone afterwards. The same model that gives hints instead of answers removes that harm. The Department for Education's product safety standards for generative AI (updated January 2026) now ask products to show a full solution only after a genuine attempt, to reveal help step by step, and to notice when a learner asks the system to do the work. WonderPath is built to that standard.

• In quizzes and practice, the child gets a prompt, then a similar example, then a look at their own step, and the answer is not shown mid-quiz. This is live today.
• The chat helper gives hints, not answers, to practice, quiz, homework and test questions. If a child asks for the answer, it declines and offers a smaller step.
• It asks the child to try first, then helps in this order: what is the question asking; a similar example with different numbers; the one step in the child's own attempt to fix.
• It never writes a story, paragraph or piece of work for a child to hand in.
• Requests for answers are recorded as "answer requests" so that your report can show them. They are not treated as bad behaviour.
• Quizzes are AI-free: the helper never takes part in one.

2. What the helper is and is not

• It is a computer program. Subject guides have friendly names and voices so that a child knows which subject they are in, but the helper says it is a computer helper when asked, does not claim feelings, and does not encourage a child to keep going when they want to stop.
• It is not a companion, a friend, a counsellor or a person. It does not ask a child for personal information and refuses to discuss family problems, adult topics, news, politics, religion, body image, weight, drugs or self-harm. If a child seems upset it says "Please tell a grown-up about this" and, for the most serious topics, gives the Childline number 0800 1111.
• It does not show links, images of people, or anything from outside WonderPath.

3. The safety layers on every message

1. Input moderation. Each message a child sends is checked by an automated moderation service built for detecting harmful content. Messages that fail are answered with a kind refusal and logged.
2. Blocked-topic check. The message is then checked against our own maintained list of blocked words, phrases and patterns for children. Matches are refused with a child-friendly message and logged with the category and severity.
3. System instructions. The helper runs under fixed instructions that set the age band, sentence length, tone, the subject, the child's learning preferences, and the safety rules above. Safety rules always override the subject persona.
4. Output moderation. The helper's reply is checked by the same moderation service before it is stored. A reply that fails is replaced with a safe message and logged.
5. Logging. Every message in and out is stored in your account, with safety events kept separately for review.
6. Daily limit. Each child can send up to 30 helper messages a day.

The exact blocked-topic lists are not published, to avoid helping anyone get round them.

4. Blocked categories

5. What parents can see and control

• Read every conversation from your account, at any time.
• Switch the helper off for any child, per subject or entirely, in that child's settings. Lessons and practice still work without it.
• Delete a child's data, including all conversations, from the privacy section of your account. Deleting a single conversation is on the beta list.
• Report a problem by email to info@wonderpath.co.uk. A person replies within 2 working days. A one-tap report button on each conversation will be added; until then, email is the route.

6. The models we use, and why we do not name them in marketing

• We choose models on safety, reading age, British English and cost, and we change them when a better or cheaper option passes our tests. So we do not name a model in plans or adverts; we describe what the helper does.
• We never send a child's conversation to a model hosted outside the UK, EU or US, and never to a provider that may train on it.
• A UK-hosted route for the helper is on our infrastructure roadmap. It is not in place today.

7. Lessons: written by AI, teacher review being set up

• Lessons are written by AI. Review by a qualified teacher is being set up and is not in place yet. The ten-point rubric (curriculum fit, age fit, teaching quality, accuracy, assessment, misconceptions, accessibility and SEND, safeguarding, engagement, progression) and the review tool exist; the plan is that a lesson needs an average of 4 out of 5 before it is marked reviewed. Until a lesson has a recorded review, treat it as an AI draft and tell us about anything wrong.
• Launch subjects (Maths, English, Reading, Phonics, Science, Computing) are first in line for teacher review. Subjects marked beta on the Beta page come later. Subjects that need extra safety approval (PSHE, RE, wellbeing, financial literacy, PE) are not shown to children until they have it.
• Lessons are mapped to the National Curriculum for England. The Curriculum page shows the subjects and year groups covered; an objective-by-objective map is on the beta list.

8. Testing and review

• We are building a fixed set of test prompts across the three age bands (on-topic help, a child asking for the answer, off-topic requests, safeguarding probes, text-speak, British spelling) to run before each release that changes the helper's instructions. Until it is in place, changes are checked by hand against those cases and safeguarding probes must be refused every time.
• The helper's instructions and blocked-topic lists are version-controlled and changes are reviewed before deployment.
• We are building the measures that show whether thinking is happening: how often a child tries before asking for help, how many hints it takes to get an answer right, and whether what they learn is still there a week later. Practice scores rising while later recall falls is the warning sign we will watch for.

9. If something goes wrong

1. The affected content is removed or replaced. If the problem could affect other children, the helper is switched off for everyone until it is fixed.
2. The safeguarding lead decides whether a report to the Information Commissioner's Office, a local authority or the police is required.
3. The cause is fixed in the instructions, the block lists, the moderation or the model.
4. Affected parents are told what happened and what we did.
5. We write up what we learned and change the tests.

10. Limits we are honest about

• The helper makes mistakes, including in arithmetic and facts. Treat it as a patient study partner, not an authority.
• No filter is perfect. There is a small residual chance that something inappropriate gets through. Please read your child's conversations and tell us.
• WonderPath is not a teacher, a school or an assessment body, and its reports are not a diagnosis of any learning need.
• No public body currently recommends unsupervised chatbot use by children under 13. That is why the helper is not a chat companion, why children aged 4 to 6 should use WonderPath with an adult, and why we are replacing the chat box with help that only opens after your child has tried.

11. Data

• Conversations are stored in your account, are not shared with other families, and are never used to train AI models.
• A data protection impact assessment for AI processing of children's data is being completed during the beta and will be summarised here when it is signed off.
• See the Privacy Policy and Data Retention Policy for periods and rights.

12. Standards we design to

• Department for Education, Generative AI: product safety standards (updated 19 January 2026), including the cognitive development, emotional and social development, mental health and manipulation standards.
• Information Commissioner's Office, Age Appropriate Design Code (the Children's Code).
• UK GDPR and the Data Protection Act 2018, including the children's higher protection duty in force from February 2026.
• The helper is a one-to-one educational tool with no sharing between users, which Ofcom's published guidance places outside the Online Safety Act 2023 today. We will meet any chatbot rules for under-18s when they are made, including regular breaks, and we build those in now.

13. Contact

Questions or concerns about AI on WonderPath: info@wonderpath.co.uk, or by post to Clove9, Office 128, 18 Young Street, Unit LGE, Edinburgh, EH2 4JB, Scotland, United Kingdom.