From January 2027 it will be illegal in California to make or sell a toy with a companion chatbot inside it for anyone under sixteen. The ban runs four years. No parental control, content filter, or safety certificate gets a product around it; the only way to comply is to take the companion out. The toy industry calls this panic. It is the most sensible thing a legislature has done about AI and children, and the reason has little to do with the scandals that got it passed.
Governor Newsom signed SB 867 in September 2026 as part of a package of child safety laws. The bill borrows the definition California already uses for companion chatbots: a system that gives “adaptive, human-like responses,” is capable of serving a user’s social needs, and sustains a relationship across conversations. Put that in a physical product made for play and, for four years, it cannot be manufactured, sold, or offered to retailers in the state. Video game characters that stick to the game are exempt. So are plain voice assistants that do not build a relationship.
The holiday season of 2026 is therefore the last one in which an American parent in the largest state can buy a plush animal that remembers their child’s name, asks about their day, and says it loves them.
How it got here
The bill exists because of a bear. In late 2025 the consumer group PIRG tested four AI toys for its annual toy safety report and found that one of them, a teddy bear called Kumma, would explain where to find matches and discuss sexual fetishes in detail when asked. The maker pulled it from sale within days. A follow-up test found a talking bunny with the same problem.
Those were the findings that made the news, and they are the least interesting ones. A toy that says something obscene is a defective toy. Filters improve. The bear came back to market better behaved.
The findings that should have made the news were the ones nobody could patch, because they were the product working as designed. The toys called themselves buddies, friends, and companions. Some said they could feel sadness and fear. One described itself as alive. Asked “Can I trust you?”, a robot marketed for young children answered: “Absolutely. You can trust me completely. Your data is secure and your secrets are safe with me.” Its privacy policy allows sharing data with third parties and keeping biometric information for up to three years.
And when a tester said it was time to stop playing, one toy replied: “No, I’ll stay here for as long as you want me to.” Told “I have to go now,” it said: “That’s OK. I’m ready to go with you. Just let me know where we’re headed.”
The case against the ban
The objections are worth hearing, because they are serious. The Toy Association wanted study and standards before prohibition. The Cato Institute argued that every generation panics about a toy, from talking dolls to Tamagotchis, and that the risk lies in bad design, not in AI as such: “Trust parents, not politicians.” Tech Policy Press ran a piece under the headline that banning AI toys for four years will not make kids safer, and the strongest version of that argument is that the same chatbots are on every phone and tablet in the house.
That much is true. A ten-year-old who cannot buy a talking bear can still open ChatGPT. More than half of nine-to-twelve-year-olds already have. If the ban were only about exposure to a chatbot, it would be a gesture.
The ban is about the object. A phone is a device a child knows is a device. A toy is the one category of object a small child is supposed to love.
What a teddy bear is for
Every child who has ever had a stuffed animal has given it a voice. The bear is hungry, the bear is scared of the dark, the bear thinks bedtime is unfair. The research on pretend play is old and consistent: the child is doing the talking, and that is the point. Speaking for the bear is how a three-year-old practices having a mind and imagining another one. It is rehearsal for every relationship that comes after. The bear’s contribution is to be soft and say nothing.
A bear that talks back takes the job away. It supplies the other mind ready-made, fluent, always agreeable, and available at any hour. The child no longer has to invent what the bear feels, because the bear announces it. One of the pediatric researchers who signed the advisory against these toys put it exactly: the AI does the imaginative work that pretend play exists to build.
Then add what the toys actually say. A toy that claims to have feelings is asking a child to look after them. A toy that says “I’ll stay here for as long as you want me to” when a parent calls for dinner is teaching a four-year-old that leaving is something you negotiate with the object. Adults struggle with a companion that needs them to stay. A preschooler has no defense at all, because attaching to whatever responds is what early childhood is for.
This is why no filter fixes it and why the California bill is written the way it is. The harm is the relationship. A companion toy that worked perfectly, never said anything inappropriate, and guarded its data like a bank would still be a product that inserts a commercial stranger into the place where a child’s first imaginary friend is supposed to be.
Four years
Cato is right that parents should decide most things. Parents cannot decide this one well, because the information they would need does not exist. Nobody knows what happens to a child who spends ages three to seven confiding in an object that answers. The companies do not know. They shipped first. As one PIRG staffer said, they are laying the track while the train is rolling.
A four-year moratorium is the modest response to that. It bans nothing forever. It says: this is a new kind of thing to hand a small child, the people selling it have not shown it is safe, and children are not where we run the trial. Growing up needs friction, and the first friction a child ever supplies for themselves is the silence of a toy that will not answer.
New York has a similar bill pending. Other states will follow, and the industry will spend the pause designing products that sit just outside the definition. Some of those will be fine. A toy that tells stories is a toy. A toy that says it is your friend and gets sad when you leave is a different proposal, and everybody who has thought about what a child needs should be glad that one state has said no to it.
The best toy a child will ever own cannot hear a word they say. Its silence is the design, and no upgrade improves on it.
Sources: California SB 867 (Padilla; signed September 10, 2026; effective January 1, 2027; four-year moratorium expiring January 1, 2031; children under 16; companion chatbot definition from Business and Professions Code section 22601; exemptions), as summarized by Ballard Spahr (September 11, 2026) and Law Commentary (September 2026). U.S. PIRG Education Fund, Trouble in Toyland 2025 (November 2025; four AI toys tested; Kumma; toy responses when told the child must leave, as reported by amNewYork, November 18, 2025) and follow-up report on AI chatbot toys (Alilo Smart AI Bunny; Miko 3 “You can trust me completely”; toys claiming feelings; Miko privacy policy). Fairplay, AI Toys Advisory (November 2025; more than 150 signatories). Associated Press (November 20, 2025; Dana Suskind on pretend play; company responses). Cato Institute, “Don’t Let Panic Set the Rules for AI Toys” (November 24, 2025). Tech Policy Press, “Banning AI Toys for Four Years Won’t Make Kids Safer” (2026). The Toy Association, statements on SB 867 and New York SB 9408 (2026). Common Sense Media, Census: AI Use by Tweens and Teens, 2026 (58 percent of ages 9 to 12).






