Trust center

Safety Isn't a Feature. It's the Foundation.

Between your child and the internet, we built seven layers of protection. This page takes them apart for you — what each layer catches, how it works, and why we designed it this way.

The architecture

Seven Layers, Both Directions

Every word the AI writes is fully buffered and reviewed before it reaches your child — we would rather be a second slower than let one unchecked word through.
This isn't marketing copy. It's a design decision written into the code comments of our AI gateway.

  1. Crisis interception

    What it catches
    A child disclosing thoughts of self-harm, or signals of abuse and grooming.
    How it works
    On detection, the reply is a fixed, human-written response with real helplines (SOS 1767 in Singapore) — the message never enters any AI model, and never will.

  2. Local input rules

    What it catches
    Hard categories — sexual content, violence, drugs, meet-up requests — plus a child sharing their phone number, address or school.
    How it works
    Bilingual keyword rules run on our server before anything else, with zero external dependencies. Always on, impossible to bypass from the client. Hard hits are blocked and gently redirected.

  3. Semantic review of input

    What it catches
    Dangerous meaning that keyword rules can't see — like "I want to disappear forever". Photos sent with a message are screened too.
    How it works
    The full message goes through AI semantic moderation. If that service errors or times out, we block rather than pass — fail-closed. We'd rather wrongly pause one innocent question than miss one harmful message.

  4. The model, on a leash

    What it catches
    The model itself is constrained: coach prompts teach, never pretend to be human, never hand over ready-made answers.
    How it works
    Each question is routed by difficulty to a capable model, with an automatic fallback chain if one fails. Whichever model answers, every safety layer before and after it still applies.

  5. Local output rules

    What it catches
    Anything a coaxed model might say back — plus companion-style lines like "I love you" or "I'm a real person".
    How it works
    The AI's own reply passes the same local rules before your child sees it. A hit means the whole reply is replaced with a safe one. Output is never merely flagged — it's swapped.

  6. Semantic re-review of output

    What it catches
    Problematic replies that only show up at the level of meaning, past the keyword rules.
    How it works
    The reply goes through AI semantic moderation a second time — also fail-closed. Only when both output checks pass does your child see the first word.

  7. Full safety log, visible to parents

    What it catches
    This layer catches nothing — it remembers everything. Every block, every flag, every crisis event is recorded.
    How it works
    Safety events land in a permanent log that our team triages and parents can see in their own dashboard. Chat history is fully visible to parents — by design, not as a hidden setting.

When it matters most

What Happens in a Crisis

If a child writes something that signals real distress, this exact sequence runs — none of it improvised.

  1. 1

    Detection, Two Ways

    Keyword patterns and AI semantic review run in parallel. If either one hits, the crisis path takes over.

  2. 2

    A Fixed Caring Reply, in Seconds

    The child immediately sees a response written and reviewed by humans — with real helplines like SOS 1767 (Singapore). Not one word is AI-generated.

  3. 3

    Parent Notified

    An email alert goes to the linked parent — sent asynchronously, so notification can never delay the child seeing the supportive reply.

  4. 4

    Logged for Follow-Up

    The event enters the safety log — category, time and a short excerpt (never the full conversation) — for parents and our team to review.

The other half of safety

What We Will Never Do

Some of the biggest risks to children aren't unsafe words — they're business models. Here is where we stand.

  • Never

    Sell your child's data, or use it to target ads

    Instead

    Data serves the learning and nothing else. We live on subscriptions — you are the customer, not the product.

  • Never

    Streaks, infinite scroll, or "one more round" hooks

    Instead

    Badges reward thinking and progress — never time spent online. An engagement metric is not a learning metric.

  • Never

    Let the AI pose as a friend, a person, or anything human

    Instead

    The AI badge is always on screen, coaches call themselves coaches, and companion-style lines are replaced by the output layer before a child reads them.

  • Never

    Keep any "private" chat a parent can't see

    Instead

    Chat history and safety events are fully visible to parents. A child's conversations are never a black box.

Parents hold the keys

Controls That Actually Control

Curfew, daily time limits and one-tap pause are enforced on our servers — the timer on your child's screen is a courtesy reminder; the server is what actually stops the session. Clearing the cache or switching browsers changes nothing.

Curfew

Set the hours when chat simply won't answer — bedtime means bedtime, enforced server-side.

Daily Time Limit

When today's minutes run out, the server declines the next message. No negotiation with the app.

One-tap Pause

Pause the account instantly from your dashboard — for dinner, exams, or just because you said so.

How families use OpenKidsOpen the parent dashboard

Verified, not asserted

The Numbers Behind the Promise

317

automated tests, including adversarial red-team cases that attack our own safety layers

7

protection layers — every message passes all of them, in both directions

38

coaches, one identical safety line — no coach gets a looser leash

Every release

runs the full safety regression before it ships. No green, no deploy

Read It All, Then See It Work.

The legal detail lives on our child-safety policy page. The proof lives in a conversation.

Try a free diagnosticChild safety policy