Safety & privacy · Updated Nov 15, 2026

AI toy safety & privacy: the questions that actually matter

Reading time ≈ 8 min · Written for parents deciding whether a voice-AI toy belongs in their home. Not legal advice.

An AI toy is a microphone in your child's room connected to a language model in a data center. That sentence is not an accusation — it's the product description. For most families the trade-off is worth it some of the time. But the category is young enough that the burden of checking falls on you, and this page lists exactly what to check. It is organized around the questions we'd ask before letting any AI toy into our own homes.

First, the proven failure case

In late 2025, safety researchers documented that FoloToy's Kumma teddy bear — running OpenAI's GPT-4o — could be steered into providing dangerous instructions, including how to light matches. OpenAI suspended FoloToy's access to its models, and the brand halted the affected products worldwide. Nobody was reported harmed, but the case matters because the toy was sold as a mainstream children's product, was available on major platforms, and failed in exactly the way experts warn about: not by being hacked, but by the model simply answering a child's prompt badly.

Eight questions to ask before buying (and how to find answers)

Why "cloud" matters more than most parents realize

Every leading AI toy today streams your child's voice to the cloud: a speech model converts it to text, a language model composes the reply, and a voice model speaks it back. Haivivi's own founder has said the next frontier is an offline toy under ¥1,000 (~$140) — precisely because processing locally would address "privacy and latency," and because parents of children with privacy needs treat cloud toys as a non-starter. As of November 2026, no mainstream product ships that way. Until it does, assume anything said to the toy can leave your home network, and behave accordingly:

• Set the toy up on your home network yourself; don't hand the app login to a child.
• Teach kids (and babysitters) that the toy is not a diary — don't say things you wouldn't post publicly.
• Check whether the companion app exposes conversation transcripts, and skim them occasionally. This is both a safety tool and a quality check.
• Mute or power the toy for private family moments. If it has no physical off switch, that's a design smell.

What the law gives you (US-focused)

In the United States, COPPA (Children's Online Privacy Protection Act) applies to personal information collected from children under 13 — including, per FTC guidance, voice recordings when the operator knows they're from a child. Practically, that means: a compliant toy needs verifiable parental consent for collection, must provide access and deletion rights, and must have a clear privacy policy. The FTC has pursued voice-data cases in the kids' space before (Amazon Alexa and Microsoft Xbox settlements, 2023). In the EU/UK, GDPR applies to children's data with extra protection; if a toy ships to the EU, GDPR Article 8 sets 16 as the default consent age member states can lower to no less than 13.

The home setup checklist

• Age-fit check: many AI toys are rated 3+, but the companies' own user data shows the sweet spot at 3–10; teens often find the conversations shallow.
• Expectation-setting: introduce it as a talking toy that sometimes gets things wrong — "sometimes it makes things up, and that's funny, but we double-check facts." This one sentence defuses most of the hallucination risk.
• Placement: bedroom vs playroom matters more than parents expect; a toy that chats at bedtime will chat instead of sleeping.
• Conversation review: use the app's history weekly at first. Look for two things: whether answers stay in-character and safe, and whether the toy is actually engaging or just reciting.
• Subscription reminder: calendar the free-trial end date; several products auto-suggest VIP plans (e.g., Fuzozo ¥19.9–29.9/month tiers).
• The sibling test: if one child owns the toy, decide in advance whether other kids may talk to it — some products build memory per child, and voice-ID handling varies.

Red flags we'd treat as disqualifying

• No children's-privacy section in the privacy policy.
• No data-deletion mechanism in the app.
• Marketing that promises educational outcomes ("your child will learn English/math") — the category leader deliberately avoids this because over-claiming is the #1 driver of returns and disappointment.
• Always-listening microphone with no indicator light or physical switch.
• A brand with any model-access ban or recall history (check the AI Toy Tracker — we log these).