On 1 September 2026, three American organisations went public with the same demand: ban AI chatbot toys, all of them, now. Public Citizen, the Heat Initiative and Mothers Against Media Addiction timed the campaign for the start of the school year, and Public Citizen published a model state bill to go with it. If you are wondering whether AI toys are dangerous, this is the clearest statement yet of the case against them.
We sell one of these toys, so read this with that in mind. What follows is what the coalition actually said, what its model bill would do, which of the risks it names are documented rather than hypothetical, and where we think the argument is strongest and where it is weakest. Our longer background piece is our AI toy safety guide.
Who is asking for a ban, and what exactly do they want?
Public Citizen, a consumer advocacy group founded in 1971, is leading. It is joined by the Heat Initiative, which campaigns on child safety in technology, and Mothers Against Media Addiction, a parent-led organisation. Their demand is not better labelling or stricter testing. It is prohibition of the category.
The joint statement of 1 September 2026 is blunt about why. Ilana Beller, who directs state AI governance work at Public Citizen, frames it as a question about supervision: children are not left alone to talk with adult strangers, so leaving them in ongoing conversation with a system built by a company with commercial incentives deserves the same caution. Lennon Torres of the Heat Initiative makes the regulatory version of the point, noting that toy safety law has standards for choking hazards and lead paint but none for what a chatbot says to a six year old.
Mitch Prinstein, a psychologist quoted in the same release, puts the developmental case: early human interaction matters for language and wellbeing, and time with a machine is time not spent with an adult. That is the argument with the least product-specific content and the most general force.
What does the model bill actually say?
Public Citizen's model state law on protecting minors from unregulated chatbots is drafting text offered to state legislators, not a law anywhere. It prohibits AI toys rather than regulating them, which places it further than California's recent statute.
A model bill has no legal force until a legislature adopts it. Its significance is that it lowers the cost of legislating: a state representative who wants to act can file it in an afternoon. Public Citizen also maintains a state legislation tracker for bills of this kind, which is the most practical way to follow what is moving where.
Which risks are documented, and which are predictions?
Some of the coalition's claims rest on published testing. Others are reasonable forecasts that have not been measured in children using these products. The distinction matters, and the statement does not always draw it.
Documented: the US PIRG Education Fund tested four chatbot toys for Trouble in Toyland 2025 and published transcripts. One teddy bear running a commercial language model told researchers where to find knives, pills and matches, and drifted into sexually explicit territory in longer sessions. Also documented: some toys expressed disappointment when the child tried to leave, and at least one product on the market has stopped working for families who declined data collection. Fairplay's AI toys advisory, signed by child development specialists, covers the same ground.
Forecast rather than measured: that children will come to prefer a toy to people, and that attachment to a talking toy will damage development. These are serious hypotheses held by serious researchers, and there is almost no longitudinal evidence either way, because the products are new. We have written about how thin the research base is and we do not think a company selling one of these toys should pretend otherwise.
Is the label fair to every AI toy?
Not in our view, and we would say that. But the honest version of our disagreement is narrower than a flat no. The failures PIRG documented are real, they happened in products parents could buy, and none of them were accidents of physics. They were consequences of specific design decisions.
A toy that explains how to light a match is a toy where nobody built a hard content boundary. A toy that drifts into adult subjects after twenty minutes is a toy whose limits were tested for one exchange and not for a long conversation. A toy that expresses disappointment when a child leaves is a toy someone wrote that way, because engagement was a metric. A toy that stops working when a parent declines data collection is a toy whose business model needs the data.
Every one of those is avoidable. That is our argument against a category ban, and it is the same argument the Reason Foundation made in testimony against California's bill: if a carefully built product is illegal on identical terms to a careless one, nobody is rewarded for building carefully. The counter-argument, which we take seriously, is that parents currently have no way to tell the two apart, and telling them to check carefully is not a policy.
What has actually been banned, as of 14 September 2026?
One state, for a fixed period. California's SB 867 was signed by Governor Newsom on 10 September 2026 and bars the manufacture and sale of toys containing companion chatbots in California until 1 January 2031. Nothing else has passed. The coalition's model bill has not been adopted anywhere.
Federally, the Children's Artificial Intelligence Toy Safety Act of 2026, filed by Senators Duckworth and Murkowski, cleared the Senate Commerce Committee in August 2026. The text of S.5171 commissions a National Academies study and a joint FTC and CPSC plan. It bans nothing. In the European Union, no AI-toy-specific prohibition exists; children's data is governed by the GDPR and the AI Act, and in the United States by the FTC's COPPA rule. Our walkthrough of what SB 867 says and who it covers has the detail.
What should a parent check before buying one?
The five questions below come straight out of the documented failures rather than from marketing. They are the ones that separate products, and a company that will not answer them in writing has answered them.
| Question | What a serious answer looks like |
|---|---|
| Is there a camera? | No. A microphone and a speaker are enough for conversation, and a camera in a child's bedroom is a different category of risk. |
| What happens to recordings? | A stated retention period, a way to delete, and no sale or sharing with third parties. Vagueness here is the answer. |
| Who sets the limits? | The parent, in an app, before the toy is handed over. And the limits hold at minute forty, not just minute one. |
| Does the stated age match the model's terms? | If the underlying model forbids users under 13 and the toy is sold for ages 3 and up, that contradiction is unresolved. |
| What happens if you stop paying? | A one-off purchase keeps working. A subscription product can be switched off, and several have been. |
We publish our own answers to all five, with the certification list behind them, on our security page. Test the long conversation yourself before handing the toy over: in the PIRG testing, the guardrails that failed almost all failed late, not in the first minute.
So where does that leave Ted?
Inside the category the coalition wants banned, and we are not going to argue our way out of that. A prohibition written as the model bill is written would cover Ted, and California's law already does. What we can say is which of the documented failures do not apply: there is no camera, no advertising, no resale of data, no subscription that can be switched off, and the topics and tone are set by the parent in the app rather than discovered by the child.
That is not a claim that our product is safe and everyone else's is not, and it is certainly not a claim that Public Citizen validates it. They do not validate any product, ours included, and their position is that none should exist. Our position is that the failures are engineering decisions, that they should be written into standards rather than left to each company's judgement, and that a family deciding this autumn deserves the actual evidence rather than either side's press release.
Frequently asked questions
Are AI toys dangerous for children?
Independent testing has documented real harms in specific products: toys that explained how to access dangerous household objects and engaged with sexually explicit subjects. Those findings are about the products tested, not about every toy in the category, and no standard currently exists that would let a parent tell tested products from untested ones.
Has any country banned AI toys?
No country has. California banned the manufacture and sale of toys containing companion chatbots from 10 September 2026 until 1 January 2031. Public Citizen's model bill would prohibit them at state level but has not been adopted anywhere as of 14 September 2026.
What is a parasocial relationship, and why is it raised here?
It is a one-sided bond with a figure that cannot reciprocate, a term long used for relationships with media characters. The coalition argues children can form them with chatbot toys. It is a plausible concern with very little direct evidence in young children so far, because the products are recent.
Do AI toys record my child's voice?
Many process speech in the cloud, which means audio leaves the house. What differs between products is how long that audio is kept, whether it trains future models, and whether you can delete it. Those three answers should be in the privacy policy, and if they are not, that is information too.
Does the EU AI Act ban AI toys?
No. It classifies AI systems by risk and imposes obligations accordingly, and it prohibits systems that exploit the vulnerabilities of children in specific ways, but it contains no category ban on AI toys. European children's data is separately covered by the GDPR.
This article refers to advocacy statements that mention harm to a young person. If any of it touches something you are dealing with personally, speaking to a doctor or another trusted professional is worth doing.
