CleverCreatorAI
Back to blog
Trust & SafetyMarch 6, 20267 min read

How we think about AI safety when the user is twelve years old

PA
Priya Anand
Head of Trust & Safety

Building an AI tutor for kids means a different set of constraints than building one for adults. Here's how TutorMate's safety architecture works — and why we think it should be the industry default.

When your users include children as young as five, "the model gave a slightly wrong answer" stops being the worst-case scenario. Building TutorMate meant designing for a much stricter set of failure modes from day one — not bolting on guardrails after the fact.

Every TutorMate interaction passes through age-band-specific filtering before it reaches a student, tuned separately for K-2, 3-5, 6-8, and 9-12 cohorts. What counts as an appropriate explanation, tone, and even pacing varies enormously across those bands, and our review processes reflect that — this isn't a single filter with a sensitivity slider.

We also made a deliberate choice to server-enforce hint progressions rather than let a model freely decide how much help to give. A student can't prompt their way to a final answer; the system is structured so that getting unstuck always involves doing some of the thinking yourself. That's a safety property and a learning property at the same time.

Parents and teachers get meaningful visibility, not just a privacy policy to scroll past — content controls, progress reports, and clear explanations of what data is collected and why. We think that level of transparency shouldn't be a competitive differentiator; it should be the baseline every product serving young learners is held to.

SafetyPrivacyEdTech
Get started freeMore articles