“AI toy” is not a product category. It is a description of where the processing happens. A toy with the label contains a microphone, a network connection, and a subscription to somebody else’s servers; the intelligence is not in the toy, it is in a data centre that the toy is allowed to talk to for as long as the manufacturer keeps paying for it.
That is not automatically bad. Cloud processing is why a $90 toy can understand a five-year-old’s speech, which on-device hardware at that price cannot do well. But it changes what you are buying from a possession into an arrangement, and the terms of the arrangement are worth reading before the money leaves, because the category has already produced a very expensive demonstration of what happens when they change.
What “AI” is doing in the sentence
In practice the label covers four different things, and only one of them is what parents think it means.
Speech recognition in the cloud. The toy streams audio to a server, the server transcribes and answers, the answer comes back. This is the most common implementation and the one with the most consequential privacy profile, because it means a microphone in a child’s bedroom connected to a company’s infrastructure.
Scripted responses with a keyword layer. A fixed set of answers selected by matching words. Sold as AI, is a lookup table. Harmless, cheap, and often indistinguishable to a child under six.
On-device processing. Rare, and the only version that keeps working without a network. Where it exists it is worth paying for, and vendors who have it say so loudly, which is a useful signal: the ones who do not have it use the word “AI” without qualification.
A chatbot in a plush body. A language model behind a speaker. The most capable and the least predictable, and the version where the content risk is genuinely unknown in advance because the responses are generated.
The four questions to ask before buying any AI toy
Does it work with the network off? Not “does it have offline mode”, which usually means a reduced demo. Ask what functions survive. If the answer is “it still lights up”, the toy is a terminal.
What happens to the audio? Is it processed and discarded, stored, or stored and reviewed by humans for quality? The last one is common and rarely stated on the box. It should be in the privacy policy, in a sentence that names retention.
What is the subscription, and what does lapsing do? Many AI toys include a year of service and then bill annually. The question that matters is not the price but the behaviour after non-payment: some toys keep their basic functions, some become silent, and the difference is not advertised.
Who is the counterparty? A company with a funding round and no revenue is a different risk from a toy division of a company that has existed for forty years. Neither guarantees the server stays up, but it changes the odds.
What a shutdown actually looks like
The clearest evidence in this section is a failure, and it is worth reading in full: the Moxie timeline. A $1,499 companion robot with a monthly subscription stopped working when the company behind it wound down; owners were left with hardware that could not do the one thing it was bought for, and a community effort was needed to keep any of it running.
Three things generalise from it. The first is that the failure is not gradual — the toy works on Tuesday and does not on Wednesday. The second is that the price of the hardware has no bearing on the risk; the most expensive toys in this category are the most exposed, because they are the ones with the most server-side intelligence. The third is that “lifetime” in a marketing sentence means the lifetime of the company, not of the child.
What the toy collects, and who it goes to
Voice is the obvious item and not the only one. A connected toy with an app typically also collects the child’s name or nickname, age band, usage times, and the device identifiers of the phone it is paired with. Under COPPA in the United States and the Age Appropriate Design Code in the UK, that combination is regulated, and the regulation is about verifiable parental consent rather than prohibition.
What that means practically: a toy that asks for a child’s full name, school and photograph to “personalise the experience” is collecting more than the experience needs, and the excess is the part that becomes a liability in a breach. The minimum viable profile is a nickname and an age band, and a vendor who asks for less is usually a vendor who has thought about it.
The broader checks — magnets, batteries, small parts — are in the safety guide, and they apply to connected toys as much as to any other.
What we have reviewed in this section
Moxie — the shutdown, the timeline, and what to buy instead if the appeal was companionship rather than robotics.
Are smart toys safe? — the three rulebooks that apply to a connected toy, what the marks actually certify, and which categories carry the most risk.
Coding robots sits next door and is often the better purchase: the ones worth buying do their work locally and will still switch on in five years.
FAQ
Are AI toys safe for children?
The physical risks are the same as any toy and are covered by the same marks. The risks that are new are a microphone in a private space, a data profile that outlives the toy, and content generated by a model nobody has audited for a six-year-old. The first two are manageable by reading the privacy policy; the third is not fully manageable by any parent, which is the honest answer.
Do AI toys stop working when the company closes?
Usually, and abruptly. Anything where the intelligence lives on a server stops when the server does. The exceptions are toys that do their processing on-device, and those are worth identifying before purchase rather than after — the test is to disconnect the network and see what still responds.
Is a subscription required?
On most of the category, yes, either at purchase or after an included period. The number to look for is not the monthly price but the behaviour on lapse: whether the toy degrades to a smaller set of functions or stops entirely. A toy that degrades is a toy you can stop paying for without throwing it away.
How we put this together
From published specifications, privacy policies, regulatory text, company filings and statements, and the pattern across user reports. We have not tested these toys and do not claim to. Where a date matters — a shutdown, a policy change, a price — the date is stated in the article, because in this category the timeline is the story.
Corrections are welcome and are logged: contact.
On the limits of this page. No toy in this section has been run with a child, and no model has been audited by us. Everything here is built from published specifications, privacy policies, regulatory text, company filings and statements, and the pattern across user reports. If a claim below depends on a number, the number and the date it was checked are stated. How the checks work.