Skip to content
Share
% 1 min read

How we think about AI safety for kids

Safety for kids' AI is often described with broad strokes — "kid-safe", "COPPA-compliant". We get asked what that actually means in practice. Here's the engineering, plainly.

Three layers

  1. Prompt scope. The model is given a deliberately narrow brief — educational question answering, age-banded. Off-scope prompts are refused.
  2. Content review. Every lesson passes automated classifiers for medical, political, romantic, and violent content. Anything above threshold is quarantined and reviewed manually before it reaches a child.
  3. Output constraints. The lesson format itself is structured — title, explanation, three sub-lessons, each with a specific shape. This narrows what the model can even try to produce.

The refusal protocol

If a kid asks something the system can't safely answer, Xplorer doesn't go silent — he redirects warmly. "That's a great question for a grown-up. Want to save it for later and ask a different one now?" Silence scares kids; a warm pivot doesn't.

The hard rule

If a child seems to be in distress — keywords suggesting harm, bullying, or crisis — the system immediately shows a calm message directing them to a trusted adult, and pauses the lesson. No model, no AI response, no clever pivot. A human interaction, always.

What we don't do

  • We don't use behavioural advertising.
  • We don't retain children's identifying data beyond what's needed for the session.
  • We don't use streaks, badges-with-guilt, or notification-driven retention.

Pick an age band and start filling Xplorer's tank