I built an AI for my child, then taught my child to doubt it

Chen-Kuo Chou, the developer of Evie · Adelaide, South Australia

My son was seven when he asked me how fast light travels.

I taught data science at university for six years. I knew the answer. What stopped me was noticing my own hand — already moving toward a device, ready to hand my son's question to a machine that would answer it faster, more completely, and in a friendlier voice than I was about to.

I didn't stop because the machine would be wrong. It would have been right. I stopped because of two questions I couldn't answer as quickly as it could.

The first was where his question would go.

Where the words go

Every AI assistant a child can reach is a network call. That isn't a scandal; it's just the architecture. The models are large and the phones are small, so the words travel to a data centre and the answer comes back. Every product in this category works that way, and every one of them manages the risk the same way: a privacy policy, a filter running on their servers, and a promise.

I'd spent six years teaching people what you can do with a dataset once you have one. So a promise was never going to be the part that reassured me. I wanted the architecture to be different.

The constraint I set was absolute, and it was the only one: nothing leaves the device. Not "encrypted in transit". Not "deleted after thirty days". Nothing leaves, because there is nowhere for it to go — no backend, no account system, no analytics, no attribution pixel. If I'm wrong about something, there's no server to breach.

That constraint is what made it hard. Speech recognition, a language model and a synthesised voice all have to fit on a phone the family already owns and run fast enough that a small child doesn't wander off mid-sentence. Evie runs Gemma 4 E2B converted to Core ML on the Apple Neural Engine, with Kokoro for the voice and Apple's on-device recogniser for speech. Replies stream sentence by sentence, so the voice starts talking while the rest of the answer is still being generated. Put the phone in airplane mode and it still answers. That demo is the whole pitch, and it's why the privacy label — Apple's own, not mine — reads Data Not Collected.

What it cost, in the parts nobody sees: the model had to be split into chunks to run on the Neural Engine at all. It gets exactly 2,048 tokens of context, and if a prompt plus its reply overflows that, the decoder doesn't throw an error — it hangs silently, and a child sits watching a "Thinking" animation that will never finish. It decodes greedily, with no temperature and no repetition penalty, so a long uncertain answer can fall into a loop and repeat itself until something stops it. On one iOS release the voice engine crashed during warm-up because of a compiler regression, and the fix was to split it into seven stages. None of that is visible in the app. All of it is the price of the server not existing.

That was the first question. The second one changed the product.

The part I didn't plan

The second question was what happens to a seven-year-old who can always get the answer.

I set out to build a safe AI for my son. What I slowly understood is that safety, framed that way, is the wrong problem — and the framing itself teaches the wrong lesson. A child who grows up with an AI that is always right, always confident and always agreeable is being trained into the one habit that will actually cost him: believing the machine. Filtering harder doesn't fix that. It hides it.

Answers stopped being scarce the moment these models shipped. Questions didn't.

Schools are still largely built to produce good answerers, and I'd spent six years standing in front of rooms full of them.

So Evie ships a lessons program, and it teaches children to distrust AI. Fact versus opinion. "Is it really true?" What AI actually is, mechanically, under the friendly voice. Black, white and grey. Quantity versus quality. Each lesson is authored and reviewed step by step, with one open question at the end that the AI itself responds to, and a companion guide so a parent can have the conversation afterwards.

I'm aware this puts me in a strange commercial position: I sell an AI whose own curriculum tells your child not to take my product's word for it. I think that's the only honest version of this product, and it's the part I'm proudest of.

The rating I chose on purpose

Evie is rated 13+ and is deliberately not in the App Store's Kids Category.

I could have applied. The Kids Category is a better badge, a better shelf and an easier sell to exactly the parent I'm trying to reach. But it requires a content standard I could only claim honestly with a filter checking every single thing the model says — and a fully offline, memory-constrained app can't run that filter today. The model is open-ended. That's what it is.

So the rating describes the model accurately, and the safety work goes where it can be real: a parent mode behind a maths gate a seven-year-old can't walk through, allowed hours by day, a classifier running on the device that flags concerning things a child says — to the parent, never to a cloud and never to me — and a report path on any reply. Parents are told all of this plainly at first launch, before they hand the phone over.

I've been told this was a bad business decision. It may well have been. It's still the right one, and I'd rather explain it than let a parent discover the gap on their own.

What I'm actually building

The more AI arrives, the more the thinking has to be the part we protect — especially for children. They will need to know what questions to ask. That isn't people against the machines; it's an order of operations. Build the human first, so there's something for the AI to amplify.

The goal was never an AI that thinks for a child. It's an AI a child can practise thinking against — at home, on a device the family already owns, without a word of it leaving the room.

My son uses it. That's the entire quality bar.