A conversational AI toy is a plush that holds an unscripted back and forth with a child instead of playing recorded lines. On 2 September 2026, researchers at the University of Washington published what happened when eight children actually played with three of them, and the answer is not flattering to the category: curiosity turned into frustration, and frustration turned into hostility.
We sell one of these toys, so read this with that in mind. We are covering the study anyway, because the failures it describes are specific and most of them are design choices rather than laws of nature. Below: what the researchers did, why the children turned on the toys, what the study says about imagination, and the checks worth making before you buy anything in this category. If you want the background first, start with how AI plush toys work.
What did the University of Washington study actually do?
Eight children aged six to eleven played with three commercially available Curio AI plush toys and then discussed the experience with researchers. The work was run through KidsTeam UW, a group at the university's Information School that designs technology together with children rather than testing it on them. It was published as Toys that listen, talk, and play: Understanding Children's Sensemaking and Interactions with AI Toys in the proceedings of the 25th ACM Interaction Design and Children conference.
The design matters for how much weight to put on it. Eight children is a qualitative sample, observed in a lab on a university campus over a summer, not a representative survey and not a controlled trial. It tells you what can go wrong and why, in detail, and it does not tell you how often it goes wrong in ordinary homes. The University of Washington's own write-up and the paper in the ACM Digital Library are both worth reading directly.
Why did the children go from curious to frustrated?
Because the toys promised more than they delivered, and children test promises hard. The pattern the researchers describe runs in a fixed order: children opened with easy questions, hit the limits of the toy within minutes, and then started playing against it rather than with it.
Two failures did most of the damage. The toys could not handle complex questions, and they could not read physical cues, so a child tickling a toy's feet or holding it up to something got nothing back. One participant complained that a toy had not listened to them, in the way an eight year old counts, roughly twenty six million times. That is a latency and turn taking problem as much as a language problem.
What is a conversational AI toy doing when it fails like that?
It is usually mishandling the turn rather than misunderstanding the words. Speech recognition, model response and speech synthesis each add delay, and a child who gets no acknowledgement within a second or two assumes they were not heard and speaks again, which restarts the whole chain.
The other class of failure is the model itself. Aayushi Dangol, the paper's co-lead author and now a researcher at Foundry10, warned that generative systems introduce risks traditional toys never had: they can state things that are wrong, flatter a child excessively, or push a child toward attachment. That last one is a product design decision, not an inevitability, and it is the reason we do not describe Ted as anyone's friend. We went through the plumbing in detail in how a talking teddy bear actually understands a child.
Does an AI toy take over a child's imagination?
The researchers think it changes who is doing the imagining, and they are honest that nobody knows the consequences. For generations a child supplied the voice and the personality for an inanimate toy. A toy that talks back arrives with its own.
Jason Yip, co-author and director of KidsTeam UW, put it as the script being flipped, with the toy now holding an imitation of imagination, and noted that we have no precedent for it and no idea how long children will even want to play with these things. His practical recommendation was not to ban anything but to talk with children about the technology they are being handed. That is a low cost thing for any parent to do and it does not depend on which toy is in the room.
It is also worth noticing what the study did not find. There is no claim here that the toys harmed the children, and the report from the University of Cambridge on AI toys and play, which is critical of the category overall, still recorded early years practitioners saying these toys could support language and communication given time, and could matter for non-verbal children. Neither report validates any product, including ours.
What did the children say about bedtime stories?
They rejected the idea, bluntly. Asked whether they would want an AI toy to read them a bedtime story, one child said it just sounded awful and another said it would destroy their dreams. That is worth sitting with if you are considering one of these toys as a substitute for a parent at bedtime.
Our reading is that children were reacting to a specific experience, a toy that had already failed to listen to them for twenty minutes, rather than to the abstract idea of a story. But the direction of the answer is a useful corrective to the marketing in this category, our own included. A talking toy is a thing a child plays with. It is not a stand-in for the adult who sits on the edge of the bed.
Which of these failures are design choices rather than inevitabilities?
Most of them. Latency, memory, what the toy is allowed to say, whether it flatters, whether it encourages attachment and what happens to the recordings are all decisions made by the company, and they are the decisions worth interrogating before you buy.
| What went wrong in the study | Is it inevitable? | What to ask the manufacturer |
|---|---|---|
| The toy did not respond to being touched or held | No, but it needs sensors and a push to talk mode | How does the child start a turn? Is there a button? |
| Long pauses, children repeating themselves | No. This is latency and turn taking design | Does it acknowledge the child before it answers? |
| Wrong answers stated confidently | Partly. All generative systems do this | Can a parent restrict topics? Does it say when it does not know? |
| Excessive flattery, pull toward attachment | No. This is a tuning and framing choice | Does the product call itself the child's friend? |
| Unclear what happens to recordings | No. This is a policy choice | Where is audio stored, for how long, and is it used for training? |
On the last row, the consumer testing published by the PIRG Education Fund on AI toys and the advisory from Fairplay and a coalition of child development experts are the two documents to read, and both are critical of the category as a whole. The legal floor in the United States is the FTC's Children's Online Privacy Protection Rule, which governs what a connected toy may collect from a child under thirteen and what parental consent it needs first. Our own answers on storage and retention are on our security page.
What should parents check before buying a conversational AI toy?
Check the four things the study exposed: how a turn starts, how fast it answers, what a parent can restrict, and what happens to the audio. A toy that fails any of these will produce exactly the arc the researchers documented, whatever is printed on the box. Our longer version of this is in seven questions to ask before you buy an interactive plush toy.
For our part, Ted is built around those answers rather than around them being unaskable. Conversations happen in real time with no pre-recorded lines. Parents choose push to talk or automatic listening, set the language, the topics, the limits and the tone from the app, and the language only changes when a parent changes it. There is no camera, only a microphone and a speaker, and it is a single purchase at 129 euros with no subscription. None of that makes the underlying technology perfect, and the honest position is the one the researchers landed on: talk to your child about the toy, and keep the adults in the room.
Frequently asked questions
How many children were in the University of Washington study?
Eight, aged six to eleven, playing with three commercially available Curio AI toys at KidsTeam UW. It is a qualitative study designed to surface what happens and why, so it describes failure modes richly and cannot tell you how common they are.
Did the study find that AI toys are dangerous?
No. It found that children became frustrated and then hostile when the toys failed to listen, and its authors raised risks around wrong answers, flattery and attachment. Safety and privacy questions are covered more directly by the PIRG and Fairplay reports.
Should I let my child use an AI toy at bedtime?
The children in this study did not want one reading to them, and there is no evidence that a talking toy improves sleep. If bedtime matters to you, the reliable ingredient is a consistent routine with an adult present, and a toy is at best an addition to that.
Do all AI toys record conversations?
They all process speech, but what is kept, for how long, and whether it trains a model differs by company. Ask for those three answers specifically, in writing, and check whether the product claims compliance with COPPA and the GDPR.
What age is a conversational AI toy suitable for?
The children in the study were six to eleven and the frustration came from unmet expectations, which younger children have fewer of and older children have more of. Ted is designed for children aged three to twelve, with the parent setting the limits from the app.
