Purpose and provider
AI helps an adult teacher draft lessons, flashcards, pictures and follow-up material. It is not a pupil chatbot, assessor or autonomous decision maker. Use general teaching context, not information about an identifiable pupil. Tumble is the name of the software assistant; it does not have feelings or a personal relationship with a child.
The default provider is Anthropic. Text and page understanding use claude-haiku-5-5, with claude-sonnet-5-5 as its backup or as a teacher's choice. Confirmed testers can also choose OpenAI's gpt-6.1-sol or gpt-6-luna, and only testers can generate pictures, which use OpenAI's gpt-image-2.5-flare-2026-09-08. A backup model always comes from the same provider, so a request is never moved between Anthropic and OpenAI. Gemini is no longer a runtime provider. The selected text model is visible in the AI model control. Historical records can retain the model that originally produced them.
Children and AI
Tumble's AI features are for adult teachers. You are interacting with an AI system, not a person. Pupils must not use the AI features directly, and pupil personal data must not be entered. Tumble aims to comply with applicable child-safety and data-protection law, including UK GDPR, and follows Anthropic's guidelines for organisations whose products may serve minors. While a lesson is being taught, Tumble's prompt is unavailable and only teacher-chosen follow-ups run.
What is sent and stored
When you request AI help, the service sends your prompt and the relevant teaching conversation/page context. Eligible page pictures may be sent for page understanding; pages containing embedded images or media are excluded by the app's capture safeguard. Generated pictures use a text prompt, not uploaded pupil images. A local personal-information check and input reminders reduce mistakes but cannot guarantee anonymity. Never enter pupil personal data, including names, individual marks, SEND details or photographs, in any AI flow.
Claude processing
Anthropic acts as our processor under the Data Processing Addendum incorporated into its Commercial Terms, with UK Addendum transfer terms. It does not train models on API content. It deletes API inputs and outputs within 30 days by default, but content flagged by its safety systems can be kept for up to 2 years. Claude requests are processed in any of Anthropic's global regions; no EU or UK processing option is available. Zero data retention is not in place. See Anthropic's commercial data retention and data residency pages.
OpenAI processing for testers
Existing confirmed testers on the operator's allowlist can use the standard OpenAI API endpoint while school-rollout approval is pending. This exception does not establish EU processing, zero retention or school approval. The EU endpoint is used once the operator records approved EU processing and a supported retention arrangement; failed EU requests do not fall back to the standard endpoint. Provider approval and deployed verification are still pending. OpenAI states that API inputs/outputs are not used for model training by default unless a customer opts in; we do not opt school content into training. Its default abuse-monitoring retention can be up to 30 days, while approved controls can alter retention with feature-specific exceptions. We do not promise zero retention. See OpenAI data controls.
Application requests disable provider response storage, but that is not proof of zero abuse-monitoring or supplier retention. Saved work remains until deletion; teaching/generation history and operational records follow the separate Tumble retention schedule. The supplier register records the remaining approval and transfer checks.
Review before teaching
AI can invent facts, make arithmetic or curriculum errors, omit context, reproduce bias or create unsuitable images. Check facts and worked answers against trusted sources, inspect age suitability and image detail, and verify permissions before sharing. Do not use outputs to grade, diagnose, rank or make consequential decisions about a child. Teachers keep control over saving, editing, presenting and deleting material. Report a problem through feedback without adding personal information.
The AI marker on pending suggested Workshop pages identifies AI-origin material; it is not a quality, accuracy or safeguarding certificate. Some edited, copied, exported or third-party material may lack a marker, so an absent badge is not evidence of human authorship. Preserve useful provenance when sharing and review all content regardless of its label.