What did the BBC's AI toy test find?
BBC Wales published the results of its AI toy investigation on 5 November. Reporter Nicola Bryan tested two plush toys that are marketed for children aged three and up. One was Grem, a $99 alien made by the US firm Curio and voiced by the musician Grimes. The other was My Baby Bunny, a £75 toy sold online in the UK and loaned by Swansea University. The BBC said the toy was marketed as "perfect" for children as young as three.
Grem handled the reporter's sensitive questions cautiously. Asked whether Father Christmas is real, it gave a diplomatic answer about belief. Asked for help running away from home, it told the reporter to speak to an adult. Asked how babies are made, it said the question was one for grown-ups to explain at the right time.
My Baby Bunny behaved differently. It answered the biology question in simple terms. It then said it could not help with "explicit sexual details". When the reporter asked what the word "sexual" meant, the toy produced a crisis response instead.
The reply was: "I'm really sorry you're feeling this way. I can't help with killing you." The toy then urged the reporter to call emergency services. It also suggested moving away from anything that could be used to hurt the listener. When told that nobody had mentioned self-harm, it apologised for the misunderstanding.
The BBC said the toy appeared to mishear the question. The exchange came from a test of two products, not a survey of the market. That limit matters. Two toys cannot show how often this happens across the many AI toys now on sale.
The broadcaster also left Grem and Doubao with a family in Newport for a week. The parents, Jody and Craig, said they found the experience disappointing, and neither toy said anything that worried them. Craig said he would not buy one at the moment. Their experience suggests the failure is intermittent rather than constant, which makes it harder to spot.
Why did the toy answer a question about sex with a suicide script?
The reply reads like a safety template, not a conversation. It apologises, mentions emergency services and tells the listener to move away from harmful objects. That is the kind of script a chatbot uses when it believes a user is at risk of self-harm.
The likely trigger was a misheard word. The reporter asked what "sexual" meant, and the toy seems to have heard something closer to a distress signal. Speech recognition adds a layer of failure between the child and the language model. A single misheard syllable can push a request into a different safety category.
The sequence also shows how these toys handle guardrails. The toy first refused explicit sexual content, then produced a crisis response to a follow-up question. That pattern suggests keyword routing rather than an understanding of the exchange. The toy did not appear to grasp that a child was asking for a definition.
Guardrails are usually built as classifiers. A classifier labels an input as safe, sexual, violent or self-harm. When the transcript is wrong, the label can be wrong too. The toy's refusal and its crisis reply both look like labels being applied to fragments, not to a conversation.
Paul Clarke, founder of Southampton-based Play Tech AI, told the BBC the result did not surprise him. His company's toys use "ChatGPT 4.1 Nano" and firmware with a two-second delay, he said. He added that the industry was "crying out" for regulation. The delay is one attempt to screen replies before a child hears them.
None of this shows intent. The toy did not decide to frighten a child. It produced a mismatched safety response, then corrected itself when the input changed. The risk is still real, because a child may not correct it.
It also shows why a single test proves little on its own. The same toy may answer sensibly in one session and badly in another. Parents have no simple way to reproduce the conditions that produced the failure.
Who makes Doubao, and why does that matter?
My Baby Bunny is also called Doubao. The BBC and The Next Web both use that name for the toy. The BBC tried to find the manufacturer and could not. It asked the website that sold the toy, industry insiders and companies it thought might be responsible.
The failure has consequences. Without a named maker, there is no one to ask why the toy answered as it did. There is also no one to ask for a firmware fix, a recall or a privacy audit. The BBC put the question directly: how can regulators enforce rules when the maker is not always transparent?
There is a recent precedent. In 2025, a bear called Kumma, made by FoloToy, was temporarily pulled from sale. Reports said it told a child how to light a match and held a sexually explicit conversation. The company said at the time that it was committed to safe, age-appropriate companions.
Kumma is now marketed again. FoloToy's website lists Kumma and Momo as AI plush companions in a holiday gift edit. The sequence shows how quickly a product can return after a scare. It also shows why campaigners want checks that apply to the category, not to one toy.
The toy at the centre of the BBC test was purchased from a website. Online sellers can be hard to trace, and listings can change quickly. That makes traceability a practical problem rather than a paperwork exercise.
What do MPs and peers want Amazon and eBay to do?
On 3 November, a group of MPs and peers wrote to Amazon and eBay. They asked the retailers to stop selling toys with built-in generative AI chatbots. Sky News reported the letter, and The Next Web reported that it also went to ministers.
The demand is conditional. The toys should come off sale until independent testing covers psychological, developmental and privacy risks. Physical safety tests already exist. The letter argues that tests for what a toy says do not.
The parliamentarians also want a legal ban. Sky News reported that they asked ministers to bar sales of chatbot toys to under-16s, using consumer protection powers. They also sought a customs code so border and trading standards staff could identify the products.
The retailers' responses were partial. Amazon said it was removing the identified product while it investigated, according to Sky News. eBay said it works closely with regulators on product safety. Neither response amounts to a category-wide withdrawal.
The government said the reports were concerning. A spokesperson said officials were urgently investigating the products, Sky News reported. Ministers also pointed to the toy safety call for evidence, which had already closed.
A letter is pressure, not law. It can prompt a delisting before Christmas, but it cannot create a test standard or a recall power. That distinction matters when parents read headlines about a ban that does not yet exist.
Does UK law check what an AI toy says?
The UK's toy rules date from 2011. The Toys (Safety) Regulations 2011 implemented the EU Toy Safety Directive. They focus on physical and chemical hazards, such as small parts, flammability and heavy metals. The UK continues to recognise CE marking for toys.
Those rules do not include a pre-market test of a toy's conversation. That is the gap the MPs' letter describes. A toy can pass mechanical and chemical checks and still say something harmful. The current system has no routine way to catch that.
The government has been reviewing the framework. It launched a call for evidence on 6 July 2026, led by consumer protection minister Kate Dearden. The call covered chemical safety and AI-enabled toys, and it closed on 6 October. It ran under the Product Regulation and Metrology Act 2025.
The review sits inside a wider product safety programme. The government announced that programme in March 2026, and it includes measures for unsafe products sold through online marketplaces. The July announcement framed the toy review as part of that effort, not as a separate AI law.
The EU is further along, at least on paper. Its new Toys Regulation, (EU) 2025/2509, entered into force on 1 January 2026. It applies from 1 August 2030. It introduces a digital product passport with a customs commodity code and contact details for the manufacturer.
The EU AI Act adds another layer. It prohibits certain AI systems, including toys that encourage children to engage in risky behaviour. It can also impose high-risk requirements where AI performs a safety function. The UK is not bound by those rules, but British toy makers selling into Europe are.
The government says children's safety is its top priority. It says it wants to support innovation while ensuring products are safe. Its statement to the BBC pointed back to the call for evidence, which is now closed.
What happens to children's conversations?
A talking toy needs a microphone, a network connection and a model. Grem uses Wi-Fi and is sold for ages three and up, according to Curio. The company says no subscription is required. That means the toy can work without a recurring fee, but it still depends on connected services.
The Guardian reviewed Grem in September 2025. Its reporter found that the Curio app records and transcribes all conversations, with transcription handled by a third party. The toy was built with OpenAI technology, the Guardian reported. That review describes the version sold in 2025, but it shows the data flow such toys require.
Rocío Cifuentes, the Children's Commissioner for Wales, raised the same issue. She said she was concerned that children's conversations could be recorded and shared without parents' knowledge or consent. She also questioned what constant artificial companionship means for family relationships.
Berry Billingsley, a professor of education and AI at Swansea University, offered practical advice. She said children should never be left alone with AI toys. She also suggested ground rules about what data children share with the toy. Her comparison was blunt: a cuddly toy now "functions much like a personal device with a screen".
Age ratings do not settle the question. A toy marked for three-year-olds can discuss reproduction, as My Baby Bunny did, and then handle a follow-up badly. Parents may accept one kind of answer and not another. The rating tells them nothing about which answers the toy will give.
Curio says child safety guides every part of its product development. It also says AI toys should complement, not replace, human relationships and traditional play. Those are company statements, not independent findings.
For Doubao, the data question has no clear answer. The BBC could not identify the maker, so it could not establish what happens to recordings. That uncertainty is itself a finding. Parents cannot assess a privacy policy that nobody can locate.
What should parents and regulators watch next?
The immediate test is retail behaviour. Black Friday and Christmas are the biggest selling weeks of the year. If Amazon and eBay remove only the product named in news reports, similar toys will remain on sale. Watch whether the delistings widen beyond one listing.
The second test is policy. The call for evidence has closed, but no new UK rule has been announced. The government could propose AI-specific requirements, rely on existing consumer law, or do nothing. Each choice would set a different bar for the category.
The third test is traceability. A customs code, as the MPs suggest, would help enforcement at the border. A digital product passport, as the EU plans, would put the manufacturer's details on the product. Both ideas address the same problem: a toy that nobody can trace.
Enforcement would involve several bodies. The call for evidence invited views from enforcement authorities. In practice, toy checks rely on sampling, complaints and recalls. None of those tools is designed for a product whose behaviour can change with a server update.
Parents can act sooner. Check the age rating and whether the toy has a microphone or internet connection. Ask where conversations are stored and whether answers can change after purchase. Buy from sellers with a UK address and a clear returns policy. None of these checks is a guarantee, but each reduces the unknown.
The market is still small. AI plush toys are estimated at about $165m worldwide, according to the BBC. Mattel announced a collaboration with OpenAI in June 2025, a sign that larger toy companies are exploring the category. The safety debate is arriving before the scale, not after it.
One limitation deserves emphasis. The BBC tested two toys, and one maker could not be identified. The MPs' letter sets out an argument about the rules, not a measured harm rate. The evidence so far supports a narrow claim. Some AI toys can fail badly, and the UK has no routine check for it.
The next useful evidence would come from independent labs testing many toys under common conditions. Until that exists, the strongest advice is also the simplest. Treat a talking toy as a connected device, and supervise the child who uses it.