How to use AI for personalized children's stories without the creepy parts

“AI safety for kids” gets thrown around as a vibe — something between a marketing checkbox and a vague unease. It’s more useful to break it into the actual things that go wrong, because each one has a different fix:

Those are four different engineering and product decisions, made by four different teams, often inside the same app. A company can get the first one right and the fourth one badly wrong. So “is it safe” isn’t really answerable as one question — you have to check each failure mode on its own.

“We take your privacy seriously” is not an answer — it’s the sentence companies write when they don’t want to tell you the answer. The actual question is: does the AI vendor use your data to train future models, and for how long is it retained?

Most consumer AI APIs default to “yes, we retain and may train on it” unless the company paying for the API has specifically negotiated or configured otherwise — this usually costs more, which is exactly why a lot of apps don’t bother. For Moonmoth, the tier we pay for has zero data retention: your kid’s name, the theme you picked, the story details — none of it is retained by the model provider or used as training data. That’s a contractual setting, not a promise, and it’s the kind of thing worth asking any AI kids’ app directly, in an email, and reading the answer carefully. If support can’t answer in one sentence, that’s information too.

If an app lets you upload a reference photo to generate a likeness, ask what happens to that file after generation. The honest, correct answer is: read once, used to generate the image, discarded — never written to disk. That’s how Moonmoth handles it; photos are in-transit only. If an app’s answer involves “stored securely” or “encrypted at rest,” that means it’s being persisted, which means it’s a smaller but real breach surface that simply didn’t need to exist for the feature to work.

A lot of kids’ apps collect more than they need because it’s easier to build that way — birthdate fields, grade level, school name. None of it is necessary to tell a good bedtime story. Moonmoth stores character archetypes (baby, toddler, big kid, pre-teen, teenager, adult) instead of ages or birthdays, and the parent — not the kid — is the account holder. There’s no reason a storybook app needs to know your child’s exact age down to the month, and if one asks for it, ask why.

This is the one I’d push back on hardest, because it’s not a settings-page fix — it’s a product-design decision made upstream of any privacy toggle. An open-ended chatbot that a child talks to directly, especially one tuned to be warm and always-available, creates a parasocial dynamic that has nothing to do with data retention and everything to do with what the product is optimized for. Common Sense Media has published risk assessments specifically flagging AI companion apps as higher-risk for kids and teens for exactly this reason — the design intent is engagement, not care.

Moonmoth doesn’t have this problem because it doesn’t have this feature: there’s no chat, no companion, nothing the child talks back to. The AI writes a story once, front to back — text, images, narration — and the interaction is the parent picking the theme and the kid listening to a finished thing, the same shape as reading a book. If you’re evaluating an app that does let a kid converse with an AI, that’s a fundamentally different risk category than a story generator, and I’d want a much higher bar of evidence before trusting it unsupervised.

You don’t have to take any company’s word for what “compliant” means. The FTC’s COPPA guidance lays out, in plain language, what a service is legally required to disclose about data collected from or about children under 13 — what counts as personal information, what “verifiable parental consent” has to look like, and what a company must let you do (review, delete) with data collected about your kid. It’s a genuinely useful half hour of reading, and it gives you the vocabulary to ask a sharper question than “is this safe.”

I’ll concede the obvious counterpoint: if what you want is a fully human-vetted, identical-every-time bedtime story, a printed book beats an AI-generated one on every axis except personalization. AI’s job here is very narrowly the personalization — your kid’s name, their favorite animal, the vibe they’re in tonight — not literary quality, and not judgment about what’s age-appropriate beyond the guardrails a company built in. Any app claiming its AI understands your specific five-year-old better than you do is overselling. It doesn’t. It generates a plausible story from what you tell it, inside constraints someone else decided on. That’s a genuinely useful, narrow tool — not a substitute for parental judgment about what your kid should see.

That’s the full list I’d want answered before letting an AI product near my own kid, and it’s the list I built Moonmoth’s story generator against. If you run through it on any other app and get vague answers, that’s worth weighing more heavily than the illustrations.