We will prove it raises grades and saves teachers time — or say plainly that it didn't.
Claims about AI in schools are cheap. Our approach is to publish how we will measure impact before any results exist, use comparison groups, and report the outcome whatever it is. No results are shown here yet: the first pilot has not reported.
Pupil attainment
Baseline and end-of-term assessment on the same curriculum topics, marked against the exam board scheme.
Topic-level mastery from the knowledge graph, so gains are tied to what was actually practised.
Comparison with matched classes not using Safe AI, adjusted for prior attainment.
Reported as effect size and months of progress — the measure the Education Endowment Foundation uses.
Independent thinking
How often pupils attempt a step before asking for help, tracked over the term.
Share of tutor turns that were hints or questions rather than explanations.
Whether pupils need fewer hints on the same topic later — the sign that learning stuck.
Teacher time
Minutes spent on marking, feedback and report drafting, logged before and during use.
Every AI-drafted mark or report is signed off by a teacher, and the edits they make are counted.
Teacher workload survey at the start and end of each term.
Safe and compliant throughout
Safeguarding alerts, false alarms and response times reported to the safeguarding lead.
Pupil data stays in the UK and is never used to train models.
A data protection impact assessment is supplied before any pilot begins.
How a pilot runs
Step 1
Agree
School and Safe AI agree the topics, classes, comparison group and success measures in writing.
Step 2
Baseline
Pupils sit a short baseline; teachers log a typical week's marking and feedback time.
Step 3
Teach
One term of normal teaching with the tutor available; teachers keep full control.
Step 4
Report
Same assessment and time log again; a report goes to leaders and governors, good or bad.
Honest limits: one school's pilot is promising, not proof. Robust evidence needs several schools and, ideally, an independent evaluator. Until then we describe results as early findings. See our live safety results.