You can engineer safety, but you can not simulate a child

A child once asked Una something they called disgusting. When you swim in the ocean, are you swimming in whale poop? The logic was perfectly sound. Whales are huge. Whales live in the sea. What else do you need?

Una answered honestly. Yes. Then she explained why it matters: whale waste feeds the tiny ocean plants that produce much of the air we breathe. The child went from grossed out to genuinely curious in the span of one conversation. That is the kind of thing you cannot plan for. It is also the kind of thing that tells you everything about whether you are building this right.

What safety testing actually looks like

Safety, you can engineer. Genuinely. You write the guardrails. You define the hard nevers: no violent content, no personal data, no adult themes, no attempts to replace a parent or a therapist. These live in Una's constitution, a document that shapes every response she gives.

Then you put them in a test suite. Automated checks that run before anything ships. Red-teaming, adversarial prompts, edge cases we have thought up and cases we borrowed from AI safety research. Each one either passes or it does not. If it does not, nothing moves forward.

This is engineering. Hard and slow and occasionally tedious, but fundamentally a problem you can solve. You write the test. You run the test. You watch it pass. Repeat until you are confident. Safety is not magic. It is rigour, applied consistently.

What no test can predict

Here is the thing no automated suite has ever caught: a real child.

Children do not follow a test plan. They follow their own curiosity, straight past every assumption you made. No prompt anyone would think to script begins with a whale poop hypothesis. No red-team session produces the exact tangent an eight-year-old takes after hearing about the Amazon. No benchmark measures what happens when the joke only makes sense at that age.

Kids ask questions that are sideways and literal and deeply logical all at once. They circle back to something you said three exchanges ago. They make connections that surprise you. That is not a flaw in the interaction. That is the whole point of it.

The automated suite is good at catching the things we can predict. It cannot catch the things we cannot. And the things we cannot predict are precisely where a child lives.

So we test both ways

We run the automated checks because they are the floor. Nothing goes out without them. But we also test with real children, because they are the ceiling check. The thing that tells us whether the floor is anywhere near where it needs to be.

When a child talks to Una and hits a wall, a weird non-answer, a moment where the warmth disappears and something robotic comes back, we notice. When a child lights up because Una took their strange question seriously instead of redirecting it, we notice that too.

Unomundi is a cultural-discovery app where children aged 6 to 12 explore countries and cultures through stories, games, and conversations with Una, a warm AI guide. The experience only works if Una can hold a real conversation. Not a simulated one. A real one, with all the unpredictability that implies.

You cannot get there with prompts alone. You get there by watching what happens when a child actually talks to her, sitting with the moments that do not work, and fixing them before anyone else's child sees them.

What that whale poop question taught us

The child who asked about swimming in whale poop was not being difficult. They were doing exactly what we want every child on Unomundi to do. Taking something they half-knew, following it to its logical conclusion, and asking out loud without embarrassment.

Una met them there. Honest first, curious second. The child came in wanting to be grossed out and left knowing something genuinely strange and beautiful about how ocean ecosystems work. That pivot, from disgust to wonder, is not in any spec. It is what happens when the system is working.

We could not have scripted it. But we can build something that makes it possible. That is the whole job.

What this means if you are a parent

You are probably already asking the right questions. Is it safe? Who checks it? What happens if my child says something unexpected? Those are good questions. You should keep asking them of every app on your child's device.

For Unomundi: yes, there is a test suite, and it runs before anything ships. Yes, the guardrails are real and documented. And yes, we also test with actual children, because the safety checks are necessary but not sufficient. A guardrail that works in a lab and fails in front of a curious eight-year-old is not a guardrail.

Safety, you engineer. Curiosity, you have to witness.

Sign up to test our beta now

FAQs

How does Unomundi keep Una safe for children?

Unomundi uses an automated test suite with guardrails, hard limits, and red-team checks that run before anything ships. We also test with real children, because automated checks catch what we predict and real kids surface what we cannot.

What happens if a child asks Una something unexpected?

Una is built to meet unexpected questions honestly and with curiosity. When a child asked whether swimming in the ocean means swimming in whale waste, Una answered truthfully and turned the moment into a conversation about ocean ecosystems.

Is Unomundi's AI safe for kids aged 6 to 12?

Unomundi is a cultural-discovery app where children aged 6 to 12 explore countries and cultures through stories, games, and conversations with Una, a warm AI guide. Una operates under a published constitution of hard limits, tested automatically before every release.