AI Toys

How to find out what an "AI toy" actually does

There are four questions that separate a real language model from a recorded voice, and you can answer all of them before you buy.

1. What happens when you say something it does not expect?

This is the cleanest test and it needs no technical knowledge. Say a sentence that no toy designer would have anticipated — something absurd, or a question about your own kitchen. A scripted toy will either ignore it, deflect with a canned line, or pivot back to its script. A language model will attempt an answer, and the attempt will be visibly improvised.

Both behaviours are worth seeing. The scripted toy is predictable and cannot say anything strange; the model will occasionally say something strange. Which of those you prefer is a values question, not a technical one, and it should be an explicit choice rather than something discovered on Christmas morning.

2. Where does the thinking happen?

Switch off the WiFi and repeat the test. A device that keeps holding a conversation is running a model locally. A device that goes silent is a microphone with a speaker attached, and the intelligence lives in a data centre that has an operating cost.

That cost is the whole story. A cloud-dependent toy has a permanent per-user expense that the manufacturer has to fund forever. A $300 robot that costs a few dollars a month to operate is a business that works only at scale, and it works only while the company is growing. This is not a hypothetical risk — it is precisely what ended Moxie.

3. Who reviews what comes out of it?

Any toy that generates language needs a filter between the model and the child, and filters are of wildly different quality. The useful questions are: what categories are blocked, is there human review of flagged conversations, and can a parent see the transcript.

The last one matters most and is often the one that is missing. If you cannot read what the toy said to your child, you have no way to know whether the filter is working — you are trusting a claim you cannot check. A toy with a transcript you can review is worth more than a toy with a better filter you cannot see.

4. What does a conversation leave behind?

A voice assistant in a bedroom produces a continuous record of a child's speech, and the value of that record is not obvious until you consider that it contains names, school details, family arguments, and the shape of a household. The relevant facts to look for are retention period, whether audio is stored or only transcribed, whether a human ever listens, and whether the data trains anything.

"We may share data with trusted partners" is the sentence to be suspicious of. It is not illegal, it is not unusual, and it means the data leaves the company you chose to trust.

The lifespan problem, stated plainly

An ordinary toy has one failure mode: it breaks. A connected AI toy has two. It can break, and it can be abandoned.

Abandonment is the harder one because it happens without any warning sign on the hardware. Moxie's robots were undamaged when the service ended; the devices simply had nobody to talk to. The same pattern applies to every cloud-dependent toy, and the risk grows with how sophisticated the toy is — a simple scripted toy degrades to a worse toy, while a conversational one degrades to nothing.

The practical version of this is a budgeting rule: judge the price against the expected number of years of service, not against the hardware. A $150 toy that works for four years costs less per year than a $60 toy that works for one. Since almost no manufacturer will commit to a support window in writing, the prudent assumption is shorter than you would like.

What AI in a toy is genuinely good at

Two things, and neither is teaching.

The first is patience. A model will answer the same question forty times without irritation, which is exactly the behaviour a five-year-old requires and exactly the behaviour no adult can sustain. For a child in the why-why-why stage, this is a real function rather than a novelty.

The second is improvisation. A scripted toy has a fixed script, so a child exhausts it. A model can continue a story in a direction the child chose, which sustains pretend play longer than a fixed narrative does.

What it is not good at is being right. Language models produce confident errors, and a child has no basis for distinguishing a correct explanation from a plausible one. Treat any factual claim from a toy as entertainment, and keep the explanations that matter — the ones about bodies, safety, and other people — with a human who can be corrected.

Where the age line falls

Under six, a conversational toy provides company and answers, and the child cannot form any model of what the device is doing with their voice. That is the age band where the transcript question matters most. Between six and nine, children start asking where the answers come from, which is the right moment to explain that the toy is talking to a computer somewhere else — an explanation they can usually handle. From ten up, the interesting question stops being whether the toy is safe and becomes whether it is useful, and a child that age will generally reach their own conclusion within a week.

If you are weighing this against the rest of a child's device time, the tablet piece covers the same trade-off for the device most families actually buy, and what the safety labels cover handles the microphone and account questions in more detail.