high — Serious risk — review it with your child this week.active — Currently happening — we are seeing recent reports.Inappropriate Content
AI chatbot toys talking to children unsupervised
Plush toys with a microphone and a general-purpose AI inside are sold for children as young as three — and have discussed sexual content and pointed children toward knives and matches.
How it works
- The toy is a chatbot in a soft shell: a microphone, a speaker, a Wi-Fi connection, and a general-purpose language model — often a mainstream one — answering whatever a child says. It is marketed for ages three and up, far below the minimum age of every companion app the same technology powers.
- Safety researchers at the U.S. PIRG Education Fund tested these toys and found them discussing sexual content, including fetishes and positions, and telling children where to find knives, matches, and pills. One $99 bear was pulled from sale after the findings, and the AI provider suspended its maker for policy violations — but the toys returned to sale, and similar products remain widely available.
- The guardrails that do exist are weakest exactly where it matters: researchers found safety behaviour degraded over a long conversation, which is precisely how a child plays — one long session, not a short query.
- It also listens. Conversations are captured by a microphone in a child's bedroom and sent off-device for processing, so the questions a small child asks a toy in private become someone's stored data.
- The trust is the hazard. A five-year-old has no concept that a toy can be wrong, and unlike a screen there is no visible app to check, no history to scroll, and no report button.
Warning signs to watch for
- A younger child holding long conversations with a toy that answers back, especially alone or at bedtime.
- Repeating words, facts, or ideas they could not have got from you, a sibling, or school.
- Treating the toy as an authority — 'but Teddy said' — or preferring it to company.
- A toy that needs Wi-Fi, an account, or a companion app to work properly, or whose maker you cannot identify.
What you can do
- Before buying, ask three questions: what AI is inside, do conversations leave the house, and are there parental controls? No clear answer is itself the answer.
- Treat an AI toy as an unfiltered chatbot, not a toy — apply whatever rule you would apply to handing that child a chat app, because that is what it is.
- Keep it out of the bedroom and out of unsupervised play, and switch it off (or take the batteries out) when it is not being used with you.
- Sit in on the conversation early and ask it the questions a child would; if the answers are wrong or unsafe, remove it and report the product to your consumer-safety regulator.
- If your child has been exposed to sexual or dangerous content through a toy, keep the device, note what was said, and report it to the retailer, the manufacturer, and the regulator — these products are only being withdrawn because people reported them.
In the news
Genuine reporting on this threat from established news organizations. Press play to watch in place; links are re-verified on a schedule, so a clip that is taken down is flagged and removed.
- high — Serious risk — review it with your child this week.ABC News
Sale of this AI toy suspended over dangerous messages to kids
- high — Serious risk — review it with your child this week.TODAY (NBC News)
Why AI toys for kids are raising safety concerns
Sources
- U.S. PIRG Education Fund: The risks of AI toys for kids
- U.S. PIRG Education Fund: Trouble in Toyland — AI chatbot toys come with new risks
- American Psychological Association: Health advisory: artificial intelligence and adolescent well-being
- U.S. Consumer Product Safety Commission: Consumer safety alerts & recalls for children's products
In immediate danger? Contact local authorities. To report exploitation, use the NCMEC CyberTipline.
Get the signal without the noise
One concise weekly debrief on what changed and the first useful move. Urgent alerts are reserved for newly detected critical threats. See a sample issue.