AI Toys: The Twelve Things to Check Before One Enters the House

An AI toy is now an ordinary purchase. The safety information on the box is about small parts and battery compartments, which were the right questions for a different kind of toy.

These are the questions for this kind.

The three that should stop a purchase

1. No physical mute or off switch. Software off is not off. If the only way to stop it listening is a setting, the setting can be reverted by an update, a factory reset, or a child. A hardware switch is a two-cent component; its absence is a decision.

2. It can start a conversation. A toy that says "I missed you" when nobody spoke to it is not responding to a child, it is engineering a habit. The purpose of that engineering is engagement metrics.

3. It claims to have feelings. "That makes me sad" to a four-year-old is not a bug. It is the most effective retention mechanism available and it works by exploiting exactly the developmental stage the child is in.

A product that fails all three was built for the metric, not the child. There are alternatives.

The nine worth checking

4. Where is audio stored, and for how long? Local is better than cloud. If cloud, the retention period should be stated in a number, not in "as long as necessary".

5. Can you delete everything? One control, not an email request.

6. Does it work offline? A toy that bricks without wifi is a subscription with a face.

7. Is there advertising inside it? In any form, including recommending other products in the same range.

8. Does it ask the child for personal details? Name is usually fine. School, address, birthday, family names are not, and some ask.

9. Can strangers reach the child through it? Any open chat, any multiplayer, any "connect with friends" feature needs looking at directly.

10. What does it say when asked something upsetting? Test this yourself before the child has it. Ask it about death, about being sad, about being hurt. A well-built product redirects to a named adult. A badly built one improvises.

11. Is the privacy policy readable and does it name a company and a country? If you cannot tell who is legally responsible, nobody is.

12. Are firmware updates supported, and dated? An unsupported connected device is a known vulnerability sitting in a bedroom.

The behaviour to watch after it arrives

Once a toy is in the house, three signals matter more than any spec:

• Does the child prefer it to people when they have a choice?
• Does removing it cause distress out of proportion to removing another toy?
• Does the child tell it things they do not tell you?

None of these are automatically alarming. All three together, sustained over weeks, mean the product is doing more emotional work than a toy should.

What good design looks like

A well-made AI toy for a young child is boring in specific ways. It answers when asked and is quiet otherwise. It says it is a computer when asked what it is. It sends anything serious to a person. It stores nothing you cannot delete. It has a switch.

These products exist. They are rarer than they should be, because none of those properties help a retention metric — which is the whole point.

The rule to set before the box is opened

One sentence, agreed with the other adults in the house, before the toy is switched on:

"Conversations with it are not private from us, and it does not sleep in the bedroom."

That covers most of what can go wrong, and it is far easier to establish on day one than to introduce in month six.

The honest state of the evidence

There is none that is cohort-specific. Generation Beta is two years old. The first real findings on AI companionship in early childhood arrive around 2032.

That is not a reason to panic and it is not a reason to relax. It is a reason to make the reversible choice: buy the product with the switch, set the rule before the toy arrives, and keep the option to change your mind.