We published two articles on the same question, and they were competing with each other rather than helping anyone. This one has been folded into the fuller version, which walks through the same chain step by step with the published research attached: how a talking teddy bear actually understands your child.
The short version is below. Everything here is covered in more depth, with sources, in that guide.
How does an AI toy understand what a child says?
A microphone opens, speech recognition turns the child's voice into text, a filtering layer checks it against age rules and the settings a parent chose, a language model writes a reply, and text to speech turns that reply back into a voice. The whole loop takes roughly one to two seconds, and none of it is pre-recorded.
The weak link is the second step. Speech recognition was built on adult voices, and children have shorter vocal tracts and developing pronunciation, so error rates are highest between ages four and seven and improve steadily from there. A toy that asks a short follow-up question when it is unsure is doing better than one that answers confidently and wrongly.
Why does this raise privacy questions?
Because voice recordings count as personal data under child privacy law. In the United States the FTC's COPPA rule treats a voiceprint as a biometric identifier that requires verifiable parental consent, and analysts at Brookings have flagged that some toys keep recording longer than they need to, or store audio when there is no reason to.
The question worth asking is not whether a toy uses AI, since almost every conversational toy does. It is what happens to the recording afterwards: whether it is stored, for how long, whether it is tied to a named child, and whether it trains anything. We go through the answers in what actually happens to your child's voice data, and our own handling is set out on the security and certifications page.
What should I ask before buying?
Five questions, one for each link in the chain: is the microphone opened by a button or a wake word, and can I switch? Is the speech recognition tuned for children's voices? Can I set the topics, and what is refused by default? What happens when the toy does not know an answer? And how long does a reply take, including what happens with no Wi-Fi?
The full version of this article answers all five and explains what can go wrong at each step: read the complete guide.
