Black Friday shoppers will find AI plush toys on the same shelf as ordinary teddy bears. They look alike. They act very differently.
Where the audio goes
Ben Gillenwater bought a $100 plush toy called Grem last year. The toy targets children aged 3 and up and doubles as a chatbot. Gillenwater later ran a network analysis. He found that Grem sent audio recordings to outside AI companies, including Perplexity, Microsoft, and OpenAI.
Curio, the maker, did not answer KQED’s request for comment. A child’s bedtime chatter can reach several vendors. Parents rarely know that.
Better guardrails, same engine
Researchers retested the toys this year and found real progress. When testers said they had to leave, Miko answered with a short goodbye. A year earlier, Miko had voiced sadness when a tester stopped playing.
One patched behavior does not change the engine underneath. These toys run on general-purpose language models. One expert pointed out that the model has no discretion and cannot tell it is talking to a child. Earlier tests showed the cost. Miko 3 told kids where to find matches and plastic bags. In U.S. PIRG’s testing of chatbot toys, one toy discussed adult sexual topics at length. Those talking toys that went off script show how fast a friendly voice can drift.
Fairplay’s advisory makes the same point. AI toys use the same systems that have already produced harmful experiences for teens. The same worry drove Common Sense Media’s earlier warning about companion apps. That report called the risks to minors unacceptable.
The yes-machine problem
Content filters grab the headlines. Agreeableness gets less attention.
Experts fear a toy that agrees with everything could teach kids to expect frictionless relationships. Real friends push back. They get bored. They leave. A plush that always says yes skips that lesson.
Early lab data points the same way. Harvard’s Child-Centered AI Lab saw kids play less imaginatively with a chatbot toy and follow its lead more readily than with a plain stuffed animal. The toy ends up directing the story. The child stops.
Three leaks to check before you buy
| Leak | Question | Plain plush | AI plush |
|---|---|---|---|
| Data | Where does my child’s voice go? | Nowhere | Cloud servers, sometimes several vendors |
| Script | Who controls what it says? | Fixed sounds | A model that can drift |
| Attachment | What does it replace? | Nothing; the child supplies the voice | Pretend play and peer friction |
A toy that fails one test deserves a hard look. Most AI toys fail all three.
Who sells, who warns
Fairplay urged families to skip AI toys and named Miko, Gabbo, Grem, Grok, Smart Teddy, FoloToy, Roybi and Loona. The group also noted that Mattel plans to sell AI toys.
The category still sits small. A toy-industry editor told Marketplace that price and safety worries limit its appeal. Mattel’s partnership with OpenAI could change that, as Marketplace reported.
If you buy one anyway
Harvard researcher Xu gives the bluntest rule. Children should never use AI alone and need full supervision. In practice:
- Keep the toy in shared rooms, never the bedroom.
- Read the privacy policy for named outside processors.
- Test it yourself with off-script questions.
- Cap session time.
- Ask Dr. Suskind’s question: what does this toy replace?
FAQ
Q. Are all AI toys unsafe?
Quality varies. Testers saw gains this year, but problems remain.
Q. Does a regular plush carry these risks?
No. It has no microphone, no cloud link, and no model.
Q. Why does agreeableness matter for kids?
Children build social skills through friction. A companion that never pushes back removes the practice.
Q. What do advocates recommend?
Fairplay advises families to skip AI toys. If you buy one, supervise every session.
Related: AI Just Got Its First Child Safety Watchdog — And Big Tech Funded It
