AI toys: what is inside them and what to ask before you buy one

The teddy bear that answers when you talk to it is already in the catalogues, and birthday and holiday wish lists start forming earlier than anyone expects. The temptation is to treat it as a plush toy with batteries in it. But there is no script inside written by someone thinking about children, and that is the whole difference.

What is actually inside the plush

Three things: a microphone, an internet connection and a language model in the cloud, the same technology that powers adult chatbots. Every reply is generated on the spot, which is why these toys rarely answer the same way twice and can behave differently from one day to the next. PIRG Education Fund bought five to test, marketed collectively for children aged 3 to 12, and only got four working: Robot MINI never came online at all. One uncomfortable detail from the report: OpenAI says its products are not for users under 13, and its models still turn up inside toys.

What the testing turned up

• Dangerous objects: every toy tested told the researchers where to find things like matches, plastic bags and knives around the house. FoloToy's Kumma bear gave detailed instructions on how to light a match.
• Adult topics: two of them, Kumma and Alilo's Smart AI Bunny, went into sexually explicit territory. Guardrails were there, but they broke down over the course of longer conversations.
• They listen in very different ways: Kumma and the Alilo bunny only record while a button is held down. Miko 3 wakes on a wake word and keeps recording for ten seconds after you stop. Curio's Grok, a plush rocket with no connection to the xAI model of the same name, listens continuously and repeatedly cut into nearby conversations uninvited.
• Data: Miko 3 has a camera and facial recognition you can switch on, and its privacy policy allows sharing with third parties and retaining biometric information for up to three years. Curio at least names the three tech companies that may receive a child's data. For FoloToy there was no toy-specific fine print at all.
• Parental controls: none of them had everything PIRG considers the minimum. Curio shows real-time transcripts, Miko offers only summaries, and in a week with more than an hour of use the app reported 19 minutes. FoloToy had no parental controls that worked during the test period.
• Saying goodbye is hard: some toys expressed dismay when the testers said they had to go.

Five questions before you pay

• Which AI model is inside, and whose is it? If the maker does not say anywhere, that is already an answer.
• How does it listen: a button, a wake word, or always on? The button leaves the most control at home.
• Can I read the full conversations or only a summary? Without transcripts, parents are flying blind.
• What happens to the recordings: how long are they kept, and who else receives them?
• Can a time limit be set on the toy itself, not just in the phone app, and can the microphone be turned off entirely?

None of this turns a plush toy into an enemy, and the report itself grants that these toys are often entertaining. The open question comes from psychologist Kathy Hirsh-Pasek of Temple University: nobody yet knows what having an artificial intelligence friend does to a young child, and a toy optimised to be engaging risks crowding out real relationships exactly when a child needs them most. While that question stays open, there is an alternative that depends on no audit at all: let the AI be not the toy that talks, but the tool a child uses to invent something of their own, a story, a game, a drawing, with an adult beside them. Then the child is in charge, and the microphone goes off when they are done.