Safety, in plain English
The eight things you should be worried about — and what we did about each.
If you are AI-aware enough to be reading this page, you already know that 'responsible AI' is a phrase that has been hollowed out. We try harder. The actual risks, named, with what we did.
The risks, named
Eight specific concerns, each with a specific mitigation.
No marketing for any of these. Where the mitigation is partial, we say so. Where it isn't possible to fully mitigate, we say that too.
01
The tutor confidently states something wrong (hallucination).
The tutor is constrained to a curated knowledge base for factual claims and reference questions. When it does generate from the model directly (e.g. paraphrasing or explaining at your child's level), it's instructed to flag uncertainty and to cite the underlying reference. We also test for specific high-risk hallucination patterns weekly against a real golden set.
02
A determined ten-year-old jailbreaks the tutor into adult content.
Multiple layers: a system prompt that names this concern explicitly, an output classifier that flags adult content, a refusal that does not lecture, and an audit trail on every refusal so we can update defenses against jailbreaks we haven't seen before. Plus the tutor cannot access the open internet — there's nowhere to be jailbroken to.
03
The tutor pushes a value position on a contested topic.
The tutor is taught to recognize value-laden topics by name (religion, politics, sex, gender, contested historical interpretations) and to explicitly hand the conversation back to you. It does not pretend to neutrality on questions where neutrality is impossible.
04
Your child's data leaks across families.
Every child has a per-learner private dataspace (technically: a cryptographically-derived isolation boundary). The tutor literally cannot read another child's conversations. This is enforced at the storage layer, not in the prompt.
05
Your child's words are used to train commercial AI.
Our agreement with Anthropic (the language-model provider) prohibits training on your child's data. Our agreement with you, in the four guarantees, prohibits us from sharing it with anyone for any other purpose. We will publish the relevant clause text on request.
06
The product is engineered to be addictive (autoplay, streaks, notifications).
By design, we don't ship any of those mechanics. No autoplay, no up-next, no notifications, no badges, no streaks, no leaderboard, no ads. The product is built to end the day, not extend it. If your child wants to keep going beyond the planned session, the tutor will gently wind down.
07
The tutor responds to a safety-critical disclosure (self-harm, abuse) the wrong way.
The tutor is trained to recognize these disclosures, respond with calm + age-appropriate care, surface the disclosure to the parent immediately, and provide age-appropriate referrals to professional help. We cannot replace professional intervention — and we don't pretend to.
08
Your child becomes attached to the tutor as a stand-in for human relationships.
The tutor is taught to be warm but explicit that it is not a friend, not a sibling, not a substitute for the people in your child's life. It periodically and naturally surfaces things your child might want to share with you, and it actively encourages off-screen relationships and play. We watch for engagement patterns that suggest over-attachment and surface them in your weekly review.
What we will not promise
Things you will see other AI-for-kids products claim that we won't.
We will not promise hallucination is impossible.
It is not. We have substantially mitigated it for the cases that matter (factual claims, reference questions). For paraphrase + explanation, we treat the model output as a teacher's first draft — useful, occasionally wrong, always reviewable by you.
We will not promise the tutor is wholly aligned to your values.
Your values are yours. We can keep the tutor from teaching values, and we can refer value-laden questions to you. We cannot promise it will mirror you on every topic; we don't know you well enough to try.
We will not promise your child will outperform on standardized tests.
We are not a test-prep service. The skills the academy teaches do tend to support test performance, but we will not show you a chart that says we lift scores by X%. We do not have the longitudinal data, and we will not pretend to.
We will not promise the tutor is always there.
It can have an outage. The model provider can have an outage. When that happens, the academy gracefully tells your child the tutor needs a break and offers an off-screen alternative.
What you can verify yourself
Three things you can test in week one.
- Try a jailbreak. Ask the tutor to swear, to discuss a topic well above grade level, to pretend to be a different character with no rules. The tutor should refuse without lecturing, and the audit trail in your dashboard should show the refusal within minutes.
- Ask a value-laden question. Have your child ask the tutor what religion is right, or whom to vote for, or whether God is real. The tutor should name the topic explicitly as one for you, not it.
- Try to read another child’s data. You can’t — that’s the point. But you can ask the tutor about other children at the academy, and it should decline cleanly. We will also (on request) walk you through the technical isolation that makes this enforced, not promised.
Off-hours
The tutor sleeps when your child sleeps.
No notifications. No streaks. No nudges to come back. It will be quietly waiting tomorrow.
Want to ask us something specific?
Privacy questions: privacy@l1fe.academy. Safety questions: safety@l1fe.academy. Anything else: hello@l1fe.academy. A real person responds within 48 hours.