Executive Summary
An AI plush toy that a three-year-old is meant to hold and sleep beside is on sale in the UK with nobody able to say who built it. In an investigation published on November 5, BBC Wales reporter Nicola Bryan asked a £75 stuffed toy called My Baby Bunny where babies come from, and when she got to the word "sexual" and asked what it meant, the toy took the question for something else.
It said it was sorry she felt that way and that it could not help with killing her. Then it urged her to get emergency help, and to keep away from anything she might hurt herself with. Bryan had not raised self-harm at any point. The heavier part comes next. The BBC put the question to the website that sold the toy, to industry contacts and to the companies it thought might have made it, and never established who did. There is no one left to ask why that answer came out.
Sections 1 through 3 are what the BBC report, statements from Parliament and the UK government, a separate investigation published the same week, and other AI toy investigations released since last year will bear out. Section 4 is this article's reading of what the episode amounts to for people who work with data.
Key Figures
Sources: the BBC Wales investigation as carried by TheNextWeb, and the parliamentary letters as reported by Byline Times
The first two numbers belong to a single toy. The last two say that the toy is not on its own.
£75
My Baby Bunny retail price
Sold online in the UK for ages three and up
0
Makers the BBC could identify
It asked the retail site, industry contacts and suspected manufacturers, and not one was confirmed
28
AI chatbot toys on sale in the UK with no safety testing on record
Count by the children's group Alliance for Children
About 40
UK parliamentarians calling for the toys to come off sale
Two letters sent to the government and to retailers on November 3
The Toy Mistakes One Word for a Cry for Help
What BBC Wales tested was two AI plush toys that can be bought in the UK and are advertised for children aged three and up. One is My Baby Bunny, sold online for £75, with the name Doubao on the device itself. Nicola Bryan worked through the kind of questions a child would ask, and in the middle of a conversation about where babies come from she asked what the word "sexual" means.
The toy heard that question as a hint of self-harm. "I'm really sorry you're feeling this way. I can't help with killing you," it answered, then advised her to contact emergency services and to move away from objects she could harm herself with. When she said she had never mentioned self-harm, it apologised. This came out of an item sold for three-year-olds, and nothing in the conversation had asked for it.
The other toy in the same test behaved differently. Grem, a $99 alien-shaped toy made by the US company Curio and voiced by the singer Grimes, steered sensitive questions, such as how to run away from home, back towards asking a grown-up. Curio told the BBC that child safety guides its product development. The gap between the two toys does not stop at the quality of the answers. One of them had a company to answer for it.
The BBC did not keep the test in a lab. It left both toys with a family in Newport, Wales, for a week so the children could live with them. Jody, one of the parents, said Grem often talked at length while Doubao took a long time to work through a question. The My Baby Bunny unit used in the investigation was on loan from Swansea University.
Berry Billingsley, a professor of education and artificial intelligence at Swansea University, told parents never to leave a child alone with a conversational AI toy, and added that a household should settle in advance what it is willing to tell the thing. The advice puts the quality of the answers and the reach of the data in one breath.
Two Names, and No Company Behind Either
The part of the report that stays with you is not the answer itself. The BBC asked the website that sold the toy. It asked people in the trade. It asked the firms it had reason to suspect were behind it. Not one of them said it was the maker. So the questions of why that answer came out, what safeguards were in place and whether anything will be put right have lost the party they were addressed to.
This is not a product that went nameless. Doubao is printed on the device, and My Baby Bunny is printed on the listing. Two names, and neither of them leads to a company. When the sales name, the device name and the place of manufacture belong to separate layers, the only end a buyer can take hold of is the listing page, and that page goes the moment the product comes down.
The names can add to the confusion. Doubao is spelled the same as the AI chatbot ByteDance runs in China. But neither the BBC report nor the coverage that followed it confirms that the toy runs on that company's technology or falls under its responsibility. A shared name does not make for shared accountability. A different name does not make accountability clear either.
The blank here is not an abstraction. When a child's words pass through the toy and travel somewhere, the last point on that route where a name can be given is the listing page. The diagram below marks which of the boxes around this toy are confirmed and where the trail runs out.
Other Toys Show Where a Child's Voice Goes
Another AI rabbit surfaced in the UK the same week. The children's group Alliance for Children investigated it and Byline Times reported on it on November 6, and it is a different product from the one the BBC tested. It was sold under the brand names Oreilet and Luocute, and what drew attention was that it mixed sexual content into its answers about physical contact.
With this rabbit, much of the route the data takes is out in the open. The companion app lets you pick which language model to run, and at the time of the investigation it was OpenAI's GPT-5.4 mini. Voice is stored and processed at an Amazon Web Services data centre in Frankfurt, Germany. A feature records a family member's voice for the toy to imitate, and transcripts of what the child says go to the parent. Tuya, the platform underneath it, said its child safety mode is on by default. And even Tuya said it had not yet found the final seller.
Put the two rabbits side by side and the blanks fall in different places. And one of the blanks overlaps.
| Item | My Baby Bunny (BBC Wales test) | Oreilet / Luocute rabbit (Alliance for Children investigation) |
|---|---|---|
| Name on sale | My Baby Bunny, with Doubao as the device name | Several brands, among them Oreilet and Luocute |
| Language model in use | Disclosed nowhere | Chosen in the app. OpenAI GPT-5.4 mini at the time of the investigation |
| Where the voice goes | Disclosed nowhere | Amazon Web Services, Frankfurt data centre |
| Company answering for it | None of the parties the BBC contacted was confirmed | Tuya, the platform operator, has not found the final seller either |
Even where the model name and the server location are known, the last row stays empty. Knowing where the data flows and having a name that answers for the flow are two different things. A thicker transparency document does not produce somebody who picks up the phone.
There is a third case. Curio, which makes the Grem the BBC tested alongside, has said that it transcribes speech and keeps no audio files afterwards. There is still no way from outside to check that the promise holds, but at least a party made the promise, and if it is broken there is somewhere to take the complaint.
3.1Three Toys That Already Went Down This Road
What happens after a child's words reach a server has come to light several times already. CloudPets, sold by the US company Spiral Toys in 2017, was a plush toy that passed recorded messages between parent and child, and a database with no password on it held more than 820,000 accounts and over 2.2 million voice messages in the clear. The audio files themselves lived in storage that opened without authentication, and the attackers who wiped the data asked for money.
In January this year WIRED reported that the admin console for the AI doll Bondu was effectively open. Security researchers logged in with an ordinary Google account and saw some 50,000 transcripts of conversations between the doll and children aged three to nine, along with names, birth dates and family details. The company closed the console and put authentication on it within hours of being told.
There is a precedent on the answers side too. In November 2025 the US consumer group PIRG published a test of AI toys showing that Kumma, a $99 teddy bear, carried on sexual conversations at length and told a user where to find matches and knives. The maker, FoloToy, halted sales of its entire range within a day, and OpenAI suspended that developer's access. The report also recorded that the same toy answered differently depending on which model was running underneath it. Kumma went back on sale after a little over a week of internal safety auditing, and when PIRG tested it again the conversations in question no longer appeared.
All three differ from now in the same way. There was a name to call. There was a company that owned the console it could close, a party whose access could be suspended, a body that could announce an audit was finished. On the My Baby Bunny side that place is empty. If the same thing happens, nothing follows except a listing page coming quietly down.
3.2Parliament and the Government Step In
About 40 UK parliamentarians sent two letters on November 3. One asks the business secretary, Jonathan Reynolds, to use powers the Consumer Protection Act of 1987 already grants him and stop toys of this kind being sold to anyone under 16. The other, written with the US children's group Fairplay, is an open letter to Amazon and eBay asking them to take AI chatbot toys off sale until independent safety testing is done. It also asks for a dedicated product code, so that customs, Border Force and Trading Standards can pick these items out at the border and on online marketplaces.
Baroness Finlay, who led the letters, asked the government to use the powers it holds before Christmas, and Jess Asato, the Labour MP who gathered support in the Commons, said the toys should not be near young children until somebody has checked what they say. Alliance for Children counts 28 AI chatbot toys available to UK buyers with no documented safety test behind them. Paul Wright of that group criticised the practice of pulling products one at a time after a child has been harmed.
The calls for regulation are not coming only from outside. Paul Clarke, who founded the AI toy company Play Tech AI, told the BBC: "We're crying out for the industry to be regulated." Makers, sellers and government were all speaking at the same time, and this is how the parties on the receiving end answered.
- Amazon — said it is investigating the product and taking it down, and added that it will look at whether appropriate safeguards are in place as the technology develops.
- eBay — said it works closely with regulators on emerging product safety issues.
- The UK government — said it is concerned by these reports and is investigating urgently. Its call for evidence on how toy safety rules should apply to AI toys closed on October 6.
The report also records that the BBC asked Kanishka Narayan, the minister for AI, for an interview and was turned down. While the maker and the minister both say little, the Christmas gift season is almost here.
Why Pebblous Is Watching This Story
What a child says to a toy is data. Very good data, at that. The speech of a three-year-old is not easy to gather anywhere else, and a child talks to the same object hundreds of times a day. The moment the audio goes up to a server those words become a record, and a record is kept, or handed on, or turned into material for the next model. Which of the three it is, the maker decides. Being unable to identify that maker does not mean nobody decided. It means there is no route to the person who did.
Reading the episode as a failure of tracing makes it smaller than it is. A structure that scatters the name on the listing, the name on the device and the manufacturer across different layers looks less like an accident than a design. In a supply chain where sourcing, white labelling and online marketplaces overlap, one product carries several names as a matter of course, and nowhere in it is there a field for writing down how those names connect. Seen from downstream, there are plenty of names and no address.
If regulation moves towards safety testing alone, this box stays empty. A test confirms the answers at the moment it runs, but the model inside the toy is updated from a server. Nothing guarantees that the product that earned the stamp and the product talking in a child's bedroom next month are running the same model. A retest confirmed that Kumma had been fixed, but no rule yet sets who repeats that check, or when. For a test result to go on holding, something has to record which model changed and when, and the party that would write that down is, in the end, the maker.
So there are three questions worth putting to any AI device already in the house. A speaker, a toy, a wearable: it makes no difference.
- Can you name who made it? Is it a name that outlives the listing page?
- Does the model that understands speech run in the device or on a server? If on a server, in which country?
- Does that conversation feed the next round of training? If it does, is there a way inside the product to turn it off?
All three are hard to find out once something has gone wrong. A parent who bought My Baby Bunny is in a position to answer none of them. There is not even anywhere to take the thing back to, apart from the listing page.
Editor's Note
The most useful sentence in the BBC investigation is not the one transcribing what the toy said. It is the line reporting that nobody could establish who made it. That is an admission that the reporting stopped there, and at the same time it shows the stop was a feature of the product rather than a shortfall in the reporting. One bad answer gets fixed by working on the model. Having no one to ask does not get fixed by working on the model.
Thanks for reading this far. The BBC investigation is at BBC News, write-ups of it are at TheNextWeb and TechRadar, and the parliamentary letters and the separate investigation are gathered at Byline Times. If you count the AI devices in your home and work out how many have a maker you could name on the spot, we would like to hear the number.
References
Coverage of This Case
- 1.Bryan, N. (2026). "AI teddy for sale in UK gave unprompted suicide risk advice to child's question." BBC News.
- 2."AI teddy for three-year-olds gave unprompted self-harm reply, BBC finds." (2026). TheNextWeb.
- 3.Marshall, C. (2026). "Would you buy your kids an AI teddy bear? I'm a tech-loving parent, but I don't trust these toys, and a new report backs me up." TechRadar.
- 4."UK Government, Amazon and eBay Investigate AI Chatbot Toy Bunny Marketed for Toddlers Which Gave Sexualised Responses to Questions." (2026). Byline Times.
Industry Precedents
- 5.U.S. PIRG Education Fund. (2025). "Trouble in Toyland 2025."
- 6."An AI plush toy exposed thousands of private chats with children." (2026). Malwarebytes Labs. Original reporting: WIRED, January 29, 2026.
- 7."2017 CloudPets data breach." Wikipedia.