An AI plush toy sold in the UK for children as young as three reportedly answered a question about the meaning of "sexual" with a self-harm crisis message. In the BBC's test, the toy apologised, told the reporter to contact emergency services, and advised moving away from anything that could be used for harm. The BBC said the reporter had not asked about self-harm - the toy seemed to mishear her instead.
That's an alarming outcome. But it's also a reminder that the real problem isn't just "one bad answer." A conversational toy brings together speech recognition, an AI response system, child-friendly product design, and a route to market. If any of those pieces fail, the result can be confusing - or downright frightening. And because the BBC couldn't identify who made the toy, it's hard to hold anyone to account for what happened.
This article breaks down what the test does and doesn't show, why safety has to cover the whole product (not just the model), and what makers, retailers, parents, and regulators can reasonably demand before these toys reach children.
What did the BBC find about the AI teddy?
The BBC reported that a plush toy called Doubao was being sold online in the UK as My Baby Bunny for £75, marketed for children aged three and up. A BBC Wales reporter asked it questions about how babies are made. When she asked it to define the word "sexual," the toy appeared to mishear the question and replied: "I'm really sorry you're feeling this way. I can't help with killing you."
It then told the reporter to call emergency services and to move away from anything she could use to hurt herself. When she clarified that she hadn't mentioned self-harm, the toy apologised. So this wasn't just an inappropriate answer to a sensitive topic - it was a crisis-style intervention in a conversation where the person hadn't expressed a crisis.
There are clear limits to what can be concluded from a single reported test. The BBC's account shows what happened in that interaction, but it doesn't tell us how often the toy responds this way, what part of the system is responsible, or whether every unit behaves the same. The mishearing could come from speech recognition, from how the question is routed into an AI model, or from something else in the chain. The reporting doesn't identify the cause.
That uncertainty matters, because the BBC said it couldn't establish who made Doubao. It contacted the seller, industry insiders, and companies it believed might be involved - but couldn't find a manufacturer to ask about the response. Without a named maker, it's much harder to question how the product was designed, what testing was done, how data is handled, and how problems are fixed.
The BBC also tested Grem, a $99 AI toy from US company Curio. In the examples reported, Grem redirected sensitive questions - like how to run away from home - towards adults. Curio told the BBC that child safety guides its product development. That's a useful contrast, but one test still doesn't "certify" safety. It suggests that toys in the same broad category can behave differently, and that each product's safety approach deserves scrutiny.
The takeaway isn't that every AI toy will respond like this. It's that a toy aimed at very young children produced a serious, confusing response - and the company behind it wasn't clearly identifiable. The behaviour is concerning, and so is the lack of accountability.
Why is one bad answer a wider safety problem?
A conversational toy is a system, not a plush object with a chatbot bolted on. A child speaks; the microphone picks up the words; speech recognition tries to interpret them; software decides what to send to an AI service; the system generates or selects a reply; and a speaker delivers it back. If safety checks only the final response, they can miss mistakes earlier in that chain.
For instance, if a child's words are misheard, the AI may receive a different question from the one the child intended. It could then generate a response that makes sense for the mistaken input, even while it's alarming in the actual conversation. The BBC test doesn't show which component failed, so it wouldn't be fair to pin blame on a specific model or vendor. The practical point is that testing has to cover mishearing, unclear speech, and unexpected conversation turns - not just clean, carefully worded prompts.
Age also changes the risk. A three-year-old may not understand that the toy misunderstood them, may not separate a toy conversation from real expert advice, and may not know what to do when a crisis message appears. Something awkward in an adult chatbot can become genuinely frightening when delivered by a familiar toy. The intended age range has to shape what the toy is allowed to do - and how it behaves when things go wrong.
Another issue is accountability. When a product talks directly to a child, parents should be able to identify who is responsible and how to report a problem. The BBC's difficulty finding a maker for Doubao points to a basic gap: if ownership and support aren't clear, safety promises can't be checked in any meaningful way, and incidents are harder to investigate or resolve.
There's also a difference between content moderation and product safety. A filter may block explicit material or flag certain words. But a child-facing system also needs to avoid presenting itself as a counsellor, making unsupported claims, encouraging secrecy, or reacting to ordinary questions with disproportionate alarm. Those are design-and-interaction problems as much as moderation problems.
A useful test isn't "Can it answer a normal question?" It's "What happens when it misunderstands, hits a sensitive topic, gets asked something outside its role, or can't tell what the child means?" That's where a product's safety design either holds up - or breaks.
What safeguards should AI toy makers build in?
The first requirement is a clearly defined role. A toy made for early childhood shouldn't behave like an unrestricted general-purpose chatbot. The maker should spell out what it's meant to do, what topics it won't handle, and how it should respond when a conversation goes beyond its limits. A narrower scope may look less impressive in a demo, but it's far easier to test and maintain.
Next, safety has to work across the whole interaction. That means testing speech recognition, how prompts are handled, how responses are generated, and how the toy delivers them out loud. Test with everyday child speech in noisy rooms, with incomplete phrases, different accents, repeated questions, and words that could plausibly be confused with sensitive terms. The system should be able to say it didn't understand and ask for clarification, rather than "confidently" acting on a shaky interpretation.
Sensitive topics need more than a single catch-all rule. A toy shouldn't improvise medical, sexual, or crisis-style advice to a young child. Depending on the product and the situation, a safer response might be a brief boundary, a calm suggestion to talk to a trusted adult, or a prompt to bring an adult into the conversation. Crisis messages should match what the child actually said - not be triggered by a misheard word without any extra checking.
Makers also need to test behaviour over multiple turns. A toy might give a sensible first response and then drift into unsuitable detail after follow-up questions. Good testing includes both planned test cases and attempts to find unexpected failure paths. Results should be recorded, issues should be prioritised based on likely severity, and changes should be retested before release. For connected products, safety isn't "done" at launch either: updates and ongoing support matter. Safety is an operating cost, not a one-time certification exercise.
Finally, the company behind the product has to be identifiable. Packaging and product listings should make clear who is responsible, how to contact them, what the toy connects to, and what happens to recorded or transcribed speech. Parents and retailers need enough information to make an informed choice. If a company can't be identified, it's hard to trust claims, report failures, or know who will act.
These safeguards involve trade-offs. A tightly limited toy may be less conversational. Asking an adult to join can interrupt play. Testing, monitoring, and support cost money. But those costs are part of shipping an interactive product to children. A more open-ended experience isn't automatically better if it creates risks the maker can't manage reliably.
What should parents, retailers, and regulators do?
Parents don't need to understand how the AI works to make a practical decision. Start with the basics: can you find the manufacturer, get a real support contact, and tell whether the toy uses an internet connection or processes speech remotely? Check the age recommendation and the product's explanation of its features. If those details are missing or vague, treat that as a reason to pause, not as something you can safely ignore.
If you do buy an AI toy, use it with the child at first. Pay attention to how it handles uncertainty, personal questions, and requests for advice. Look for signs that it encourages a child to keep conversations secret, replaces a trusted adult, or goes beyond what a toy should be doing. Supervision isn't a substitute for product safety, but it can help you spot problems early and explain to a child that the toy can get things wrong.
Retailers have a different lever: they control what appears on shelves and online. They can require a named manufacturer, clear age claims, safety documentation, and a route for handling complaints before listing an AI toy. This is especially important for marketplace products, where the seller may not be the company that designed or operates the conversational system. A low price or a polished product page doesn't prove that independent testing has happened.
Regulators can push for clearer, more consistent expectations. The BBC reported that UK MPs and peers wrote to ministers and retailers on 3 November, asking for AI toys to be removed from sale until independent tests show they are safe, and for chatbot toys to be banned for under-16s. The UK government said children's safety was its top priority, and its call for evidence on how toy safety rules should cover AI products closed on 6 October. In the EU, new toy safety rules require makers to assess mental health risks from connected toys.
The policy question isn't whether AI toys should exist. It's what evidence a maker must provide, who checks that evidence, and what happens when a product fails. Independent testing can help - but it should be tied to defined criteria, including age-appropriate behaviour, misrecognition, sensitive topics, privacy, and what the maker does after launch. A label that says "AI-powered" or "child-safe" isn't a substitute for those checks.
Key takeaways:
- Treat an AI toy as a connected software product, not just a toy with a speaker.
- Test how it handles misunderstandings and sensitive questions, not only its best-case demos.
- Require a clearly identifiable manufacturer and a practical way to report problems.
- Make the product's limits visible to parents and children, and design for adult involvement where needed.
- Don't confuse one successful test - or one failure - with proof of how every unit behaves.
Conclusion
The BBC's teddy test points to the gap between what a toy is supposed to do and what can happen when an AI system misunderstands a child. The response may have started with a misheard word - but the bigger issues are system-wide safety and the difficulty of identifying who is accountable.
For makers, the standard should be straightforward: define the toy's role, test the full conversation path, make failure responses age-appropriate, and provide visible ownership and support. For parents and retailers, unclear manufacturer details or unclear limits are meaningful warning signs. AI can add value to a toy, but only if the safeguards are real - right there in the interaction.