AI & Intelligence Policy

We use the term ‘AI’ to refer to two distinct things: (1) the statistical and rule-based models that turn student behavioural signals into teacher intelligence, and (2) the large language model (LLM) used to draft natural-language summaries in our monthly Learning Intelligence Report. Both are narrowly scoped and described below.

We model five signal families captured during assessment: temporal (response timing), response pattern (which distractor was chosen), confidence (overconfidence and hesitation markers), navigation (re-reads, skips, abandonment), and longitudinal (retention over time). Models operate on quiz-session data only. They do not draw on social media, location, biometric, or off-platform data.

Recommendations surfaced to teachers (“reteach this”, “check that student”, “skip this in revision”) are produced by deterministic rules applied to modelled signals, with severity ranking based on frequency, consistency, and misconception specificity. The teacher always sees the reason a recommendation was made.

• It does not make automated decisions about individual students that produce legal or similarly significant effects.
• It does not grade, rank, or label students for reporting outside the teacher’s own dashboard.
• It does not share student data with any third-party AI provider.
• It does not generate content about identifiable students using an LLM.
• It does not train third-party foundation models on student data.

All intelligence outputs are advisory. The teacher or organisation retains full decision-making authority. We provide the reason for every recommendation so the human in the loop can accept, adapt, or reject it.

Our monthly Learning Intelligence Report uses Anthropic’s Claude model to draft natural-language commentary from pre-aggregated, non-identifying statistics. No individual student record is sent to the LLM. The draft is reviewed by a member of the StartSprint team before publication.

We monitor signal-to-outcome calibration across subject and cohort to detect systematic bias in recommendation quality. Where a pattern is found we pause the affected rule, investigate, and publish a change note. Teachers can flag any recommendation they consider unfair via the in-product feedback button.

We publish a summary of model changes and any material shift in recommendation logic in our product changelog. The current list of third-party model providers and the role each plays is maintained at privacy@startsprint.app on request.

We update this policy when our use of AI materially changes. Updates are published at least 14 days before taking effect and are notified to active teacher and organisation accounts.

Questions about how we use AI: privacy@startsprint.app