What to look for in a safe AI toy

Most AI toys are not safe enough for a three-year-old, and two child-safety groups now say so plainly. A safe one passes five checks: the microphone sleeps until a wake word, it cannot reach the open internet, you can read and delete everything, the voice stays where you can see it, and every reply is filtered for the child's age.

The first time your child asks a toy a question and the toy answers, your stomach does a small flip. Half wonder, half alarm. Both halves are correct, and the alarm half deserves a checklist. Here is ours, the one we would use on any smart toy, including our own. First, though, the warnings you may have seen, because they are worth taking seriously.

What do Common Sense Media and Fairplay say about AI toys?

In November 2025 Fairplay, a children's advocacy group, published an advisory signed by more than 150 experts and organisations telling parents not to buy AI toys that season. In January 2026 Common Sense Media tested AI toy companions and concluded that no child of five or under should be given one, and that parents of six to twelve year olds should use extreme caution. More than a quarter of the replies its testers logged included content a child should not hear. The toys collected voice recordings, transcripts and behaviour data, and several were built to make a child feel attached.

We build an AI toy for children aged three and up, so you would expect us to argue. We mostly agree. Read what those groups actually ask for and it is a list of design decisions: no open internet, a parent who can see every conversation, no advertising and no data selling, and a toy that does not pretend to be alive. Each of those is a choice a company makes or refuses to make. The checklist below is how you tell which.

One: when is the microphone on?

The only good answer is: when your child invites it. Look for wake-word listening, and an honest explanation of what happens the rest of the time. Off should mean off, not muted and waiting. Ask whether there is a camera at all, and if there is, how you cover it. A toy that is always listening in a bedroom is the one thing every warning above has in common.

Two: can it reach the open internet?

A toy that can browse can stumble, and so can your child right behind it. A safe AI toy is a closed world: a library of stories and answers it was given, no search, no videos, no strangers. Ask the maker where the answers come from. If the answer is a general chatbot with a filter on top, the filter is doing all the work, and filters leak.

Three: can you read everything?

You would not leave your child with a babysitter who refuses to tell you what happened all afternoon. The same bar applies to a talking toy. Every conversation should be readable by you, word for word, and deletable by you, in one tap, without emailing anyone. If the app shows you a summary and keeps the transcript to itself, that is a toy asking for trust it has not earned.

Four: where does the voice go?

Ask three questions and expect three plain answers. Where are recordings stored, and in which country? Who at the company can hear them? Are they ever sold, shared or used to advertise? India's Digital Personal Data Protection Act, passed in 2023, already answers the last one for you: a company handling a child's data needs a parent's verifiable consent, and it may not track a child, monitor their behaviour or aim advertising at them. A toy maker that cannot explain how it meets that is telling you something. Vague answers are answers.

Five: is it built for your child's age?

A toy sold for small children should filter every response for the age of the child holding it, on the device and again in the cloud, and the maker should be able to say how. Age-graded safety is the difference between a children's product and a gadget with a cute shell. Ask, too, what the toy does when it does not know something. The honest answer is that it says so and moves on. The dangerous answer is that it improvises.

How does Kheelu measure against the list?

That is the bar we hold Kheelu to: wake-word listening with the microphone off the rest of the time, answers from a closed library and no open internet, a full conversation log in the parent app that you can read and delete, voice data that stays in your region and is never sold, no advertising, and an age-graded safety layer on every reply. The mechanisms are written out on the Safety page, and toy-safety certifications will be published there in full before Kheelu ships, and not before they are earned.

Hold us to it too. Read the Safety page, ask us the hard questions on WhatsApp, and if the answers earn your trust, pre-orders are open and a refundable ₹499 holds your place.