Search for a talking teddy bear online and you will find dozens of listings with the same three words in the title: powered by ChatGPT. It reads like a quality guarantee. A ChatGPT toy sounds like it comes with the reliability of a company most parents have heard of, and the safety work that presumably goes with it.
The reality is more complicated, and it was documented in detail on 2 March 2026 by the U.S. PIRG Education Fund in a report called Not for Kids. Found in Toys. The short version: the AI companies whose names appear on these toys frequently do not allow children to use those same models directly. This article explains how that gap exists, what it does and does not tell you about a given toy, and what is actually worth checking before you buy.
What is a ChatGPT toy, exactly?
It is a plush toy or robot with a microphone and a speaker that sends what your child says to a large language model over the internet and plays back the reply. The toy itself contains almost no intelligence. It is a shell around an API call to a model running in a data centre somewhere else.
That matters because it means the phrase "powered by ChatGPT" describes a supplier relationship, not a product. Two toys can both use the same OpenAI model and behave completely differently, because everything a parent actually cares about, what topics are allowed, what the toy refuses, how long a session runs, what happens to the recording, is decided by the toy maker in the layer they build around the model. We walked through that architecture in more detail in how an AI toy understands what your child says.
Are the models inside these toys approved for children?
Usually not, at least not for direct use. This is the core finding of the PIRG report, and it comes down to a split between two sets of rules at the same company: one for consumers, one for developers.
OpenAI, xAI and Meta all bar users under 13 from using their chatbots directly. Anthropic sets its floor at 18. Yet a developer can pay for API access to those same models and ship them inside a product aimed at four-year-olds. Google is the outlier in the other direction: the Gemini API terms explicitly prohibit developers from using its models in products for under-18s, and PIRG still found at least five toys claiming to run on Google models.
| Company | Minimum age for direct consumer use | Developer products aimed at children |
|---|---|---|
| OpenAI | 13 | Permitted, subject to a ban on exploiting, endangering or sexualizing under-18s |
| Anthropic | 18 | Permitted, and the only company that asked PIRG about intended use |
| Meta | 13 | Permitted |
| xAI | 13 | Permitted |
| 13 (Gemini) | Prohibited for under-18 products by the API terms |
The primary documents are worth reading rather than taking on trust: OpenAI's terms of use and Anthropic's consumer terms both set out the consumer age floors clearly.
How closely are toy makers vetted before they get access?
Barely, according to PIRG's own test. Researchers signed up for developer access at five leading AI companies while calling themselves "PIRG AI Toy Inc", to see whether the name alone would prompt any scrutiny. Four of the five asked no substantive vetting questions at all, requiring only an email address, a credit card and a checked box.
The researchers then built working chatbots simulating an AI teddy bear on three of the platforms. Each one took under fifteen minutes. You can read the full PIRG findings on how AI models not meant for kids end up in toys, and the complete report as published. NBC News covered the findings when they were released.
What happens when a toy breaks the rules?
Enforcement exists, but PIRG's follow-up suggests it is harder to make stick than it looks. In December 2025, PIRG's first report on the risks of AI toys found that a bear called Kumma, made by FoloToy and claiming to run on GPT-4o, would give step-by-step instructions for lighting a match and would discuss sexual topics with a user it believed to be a child. OpenAI suspended FoloToy's developer access the following day.
Three months later, PIRG reported that their Kumma still appeared able to run on OpenAI models, showing GPT-5.1 in its user portal. OpenAI told them the ban remained in place. PIRG lays out several possible explanations: the toy may be using a different model and mislabelling it, the company may have opened a new account, or it may be running one of OpenAI's open weight models, which cannot be switched off remotely at all. None of these has been confirmed.
The honest reading is that nobody outside those two companies knows which explanation is correct. What the episode does establish is that a ban announced in a press statement is not the same as a ban you can verify as a parent.
Does the model brand tell you anything about safety?
Very little on its own, and this is the part most product listings depend on you not knowing. A well-built toy and a badly-built toy can sit on top of the same model. What separates them is the work in between: the system instructions that define what the toy will and will not discuss, the filtering applied to replies before they are spoken aloud, the age-appropriateness testing, and whether the company has any process for handling a failure once it is found.
None of that is visible from the words on the box. And a model brand is not a certification: as PIRG's testing showed, the same underlying model produced acceptable behaviour in some toys and clearly unacceptable behaviour in others.
Where we stand on this. PIRG's reports are critical of the entire category of AI toys, including products like ours. They validate no specific toy, and we are not going to pretend otherwise. What we would argue is narrower: the failures they documented are the result of design choices that were avoidable. Selling a toy for four-year-olds with no topic restrictions and no session limits is a decision someone made, not an inherent property of the technology.
What should a parent check before buying a talking plush toy?
Ignore the model name and ask about the layer built on top of it. Six questions that a manufacturer should be able to answer in writing:
- Can I restrict what the toy talks about? If topic control lives only in the seller's own settings and not in a parent app, you have no control.
- Is there a camera? A microphone and a speaker are enough for conversation. A camera in a child's bedroom is a separate category of risk.
- What happens to the audio? Ask specifically whether recordings are stored, for how long, and whether they are ever used to train models or shared with third parties.
- Is there a subscription? A monthly fee means the toy stops working if the company folds or you stop paying, and it usually means a data relationship that continues indefinitely.
- Which certifications does it hold? Physical toy safety marks such as CE, EN71 and ASTM F963 are separate from data rules such as GDPR and the FTC's COPPA rule. A serious manufacturer will list both.
- Can I see the conversation history? If a parent cannot review what was said, no amount of policy language substitutes for it.
The same checklist applies to the question of what a toy does when the connection drops, which we covered in whether AI toys work without internet. For the broader buying decision, our guide to AI plush toys works through the trade-offs in order.
How does Ted&Co handle this?
We would rather be judged on the answers to the six questions above than on a logo. Ted is an interactive plush toy for children aged 3 to 12. It has a microphone and a speaker and no camera. Setup happens in a parent app, where the adult sets the Wi-Fi, the language, the permitted topics, the tone and the session limits, and chooses between push-to-talk and automatic listening. Language is changed manually by the parent in that app and never by the toy on its own during a conversation.
It is a one-time purchase at 129 euros with no subscription and no recurring fee. It holds CE, RED, EN62115, EN71, RoHS, REACH, ASTM F963, CPSIA, CPC and FCC Part 15 certification, and is built to GDPR and COPPA requirements. Our full position on data handling, retention and parental controls is set out on our security page, which is the document worth arguing with if you disagree with us.
Frequently asked questions
Is a ChatGPT toy safe for a 5-year-old?
That depends entirely on the toy, not on ChatGPT. The model brand tells you which supplier the manufacturer buys from and nothing about the topic restrictions, filtering or parental controls built around it. Judge the manufacturer's safety design, not the logo.
Why does OpenAI let toys use ChatGPT if children cannot use it directly?
Because the consumer terms and the developer terms are separate documents. The consumer terms set a minimum age of 13 for using ChatGPT directly; the developer policy prohibits products that exploit, endanger or sexualize under-18s but otherwise permits building products for children.
Do these toys record my child's voice?
Most send audio to a server to be processed, because the model does not run on the toy. Whether that audio is retained afterwards, and for how long, is a manufacturer decision. Ask for it in writing before buying.
Are AI toys banned anywhere?
Not as of September 2026, though several legislatures are considering restrictions, including a California bill that would pause sales of toys with companion chatbots for under-16s. Rules are changing quickly, so check the current position rather than relying on an article.
What is the difference between a ChatGPT toy and a toy with pre-recorded phrases?
A pre-recorded toy plays a fixed library of audio and cannot say anything its makers did not record. A model-driven toy generates each reply as it goes, which is what makes real conversation possible and also what makes the safety layer around it necessary.
Policy positions and the PIRG findings described here reflect the state of the record as of 6 September 2026.
