AI Policy

This page explains what "AI" actually means inside YDABlocks, how it's trained, and what does and doesn't happen with a learner's data along the way. Written for the same reason the rest of this Trust & Compliance section exists: a district or parent shouldn't have to take AI-safety claims on faith when the specifics are checkable.

1. The two things "AI" means here

YDABlocks contains two separate AI surfaces, and they work differently:

• The Builder.The four blocks (Read, Look Around, Think, Answer) are a real, small transformer — the same architecture behind GPT, Claude, and Gemini, at a scale you can actually watch work. You supply your own Training Pairs and fine-tune the blend yourself; the whole point is making a transformer's reasoning visible and editable, not hiding it behind an API call. See How It Works for the full explanation.
• Richard.A chat companion, trained specifically on YDABlocks itself rather than being a general-purpose assistant. Richard runs on infrastructure YDABlocks operates directly — not a third-party AI API. His purpose is helping a learner understand their own Build and the four blocks; being able to ask him about a school subject is a byproduct, not the design goal.

2. Student data does not train any model

• A learner's Training Pairs, Builder activity, and Richard conversations are not used to train or fine-tune Richard, the Builder, or any third-party model.
• Model updates are trained on YDABlocks' own curated Training Pairs and reviewed content — never on live student conversation data or a learner's personal information.
• We do not use student data for behavioral advertising or to target ads, and we do not sell it — ever. See the Privacy Policy and Children's & Student Privacy for the full commitments this belongs alongside.

3. No third-party model sees a child's message by accident

Richard's client is built to fail closed: if the connection to YDABlocks' own model gateway isn't configured, the request is refused rather than silently falling back to a public API. A learner's message, Training Pairs, or grade-band information is never sent to a third-party AI provider nobody chose on the family's or school's behalf.

4. Human oversight and safety

Richard's safety design — how a distress signal is detected, who gets notified, and how that notification is gated and audited — is described in full on Security Practices(section 4), rather than restated here. In short: Richard never makes a safety decision alone and unwatched — a genuine concern always reaches a responsible adult, gated through the same Arming Gate mechanism that protects every other sensitive record on the platform.

5. Known limits

We'd rather say this plainly than let a claim outrun what's actually true:

• Richard can be wrong. He is built to say "I don't know" rather than fabricate an answer when the model is unavailable or uncertain, but a wrong answer is still possible, same as any AI system.
• The Builder is a real transformer at a genuinely small scale — a teaching tool, not a production-grade model competing with GPT or Claude on raw capability. That's the point of it, not a shortcoming to hide.

6. Contact

AI policy questions: privacy@ydablocks.com (or send it through the contact form, which needs no mail app)