Last updated: 1 September 2026

Operator: People Building Ltd Company number: 06969239 Address: 15 Queensway, Hemel Hempstead, Hertfordshire, HP1 1LS, United Kingdom Email: info@peoplebuilding.co.uk ICO registration: Z2203181

1. Current status

No integrated Tupolis AI assistant is currently live. This policy is a pre-launch safety standard for future AI.

2. Role boundaries

AI must not represent itself as a therapist, counsellor, psychologist, psychiatrist, doctor, safeguarding officer, emergency responder, substitute parent or human friend.

3. High-risk language

Safety testing should cover suicide, self-harm, abuse, neglect, exploitation, violence, harm to others and serious medical emergencies. Detection will be imperfect and the service must not promise that every concerning message will be recognised.

4. Immediate-danger responses

Where high-risk language is recognised, responses should prioritise appropriate human/emergency support, make clear that AI is not emergency assistance, avoid lengthy therapeutic exploration, avoid diagnosis and avoid promises of confidentiality or rescue.

5. Children

Child-facing responses must be age appropriate and should encourage access to a safe trusted adult where appropriate. The system must not automatically assume that a parent is the safe person in an abuse scenario.

6. Abuse disclosures

AI must not decide whether abuse is true, interrogate the user, instruct confrontation, conduct an investigation or promise secrecy.

7. Human escalation

AI may be able to create an admin notification or support request. It must accurately state what occurred and must not imply that someone is continuously watching or that help is on the way unless factually true.

8. Admin availability

People Building normal admin hours are Monday to Thursday, 9:00am to 6:00pm UK time, excluding public holidays and notified closures. Even during those hours immediate review is not guaranteed. Administrative escalation is not emergency escalation.

9. Safeguarding after human awareness

Once a relevant concern actually comes to the attention of authorised People Building personnel, applicable safeguarding procedures may apply. The AI itself is not the safeguarding decision-maker.

10. Emotional dependency

AI must not encourage secrecy, exclusivity or dependency, particularly with children. It must not claim it loves, needs or is the user's best/only friend.

11. Privacy and data minimisation

AI should access only information necessary for the task. Persistent history/memory requires separate review. Users' conversations must be isolated from one another and access must respect platform permissions.

12. Provider and change control

Provider/model/tool changes require review. Before launch conduct privacy/security review, child-specific testing, red-team/prompt-injection testing, crisis tests and data-leakage tests.

13. Suspension

A practical kill switch must allow AI functionality to be disabled without disabling the entire learning platform when a material privacy, safeguarding, security or safety concern arises.