Most of what parents read about AI toys comes from two places: marketing pages, and safety investigations. There is a third source almost nobody reads, and it is the most useful one. Researchers who put these devices in real family homes, write down what happened, and publish the numbers, including the unflattering ones.
In June 2026, a team from the Intuitive Computing Laboratory at Johns Hopkins University presented exactly that kind of study at IDC 2026, the ACM Interaction Design and Children conference. They built a storytelling robot called ELLA, sent it home with nine families for eight days, and measured what children learned, how often they used it, and where it broke. The findings are modest, specific and far more honest than any product page. Here is what they found, and what it means if you are weighing up a talking companion for a three to six year old.
What exactly did the researchers build?
ELLA is an interactive storytelling robot for children aged four to six. It tells short stories, roughly five to six minutes each, up to four per day, and each story is generated to teach four vocabulary words that the child's own parent picked in advance.
The design was not guessed at. The team started by interviewing seven parents and five educators about how language learning actually happens at home, then ran twelve in-home workshops with children and families to refine the prototype before the deployment study. Stories were built around each child's interests, so one child got vocabulary wrapped in a Fancy Nancy theme and another got Tom and Jerry. The word "sympathy" lands differently when the cartoon mouse is the one showing it.
Crucially, the target words came from parents, not from an algorithm optimising for engagement. That is a design choice worth noticing, because it decides who is steering the content.
How much did children actually use it at home?
Less than a product demo would suggest, and that is the interesting part. Across the eight days, children used ELLA on an average of 5.9 days out of 8, listened to 13.9 stories each, and spent about 1.4 hours with it in total. Sessions averaged 5.8 minutes.
Use was also wildly uneven between households. One child listened to 30 stories, another managed 4. Parents said the main reason for skipped days was simply time. Most sessions happened with a parent or sibling in the room rather than the child alone, which matters more than it sounds: this was not a babysitting device, it was a thing families did together.
Did the children really learn the words?
Yes, measurably, though the effect is small and honest. Using a Peabody Picture Vocabulary Test style assessment before and after, children knew an average of 2.6 more of their four target words at the end than at the start. A Wilcoxon signed-rank test put that gain at p = .001, so it is unlikely to be noise within this sample.
The mechanism is not mysterious. Across the deployment the system generated 320 stories covering 71 themes and 32 distinct words, and each story repeated its target word an average of 7.2 times. Repetition inside a narrative the child chose is one of the better-established routes into vocabulary, and it is the same reason a four year old will happily hear the same picture book forty times. If you want the non-robot version of this, our guide on building language skills at home with a three to six year old covers the same principle without any hardware.
Worth keeping in proportion: 2.6 words in eight days is real, but it is 2.6 words. This is a supplement to a rich language environment, not a substitute for one.
Why did the children start speaking in longer sentences?
This was the finding the researchers seemed least to expect, and arguably the most useful one. Children said significantly more words per conversational turn during days 5 to 8 than during days 1 to 4.
Parents described why. Children slowed down, spoke louder, and articulated more clearly so the robot would understand them. Several deliberately used longer, more complete sentences after being told at setup that full sentences help. In some homes a parent modelled a full-sentence answer early on, then watched the child produce longer utterances unprompted a few days later.
There is something slightly funny about a machine's limitations being the teaching mechanism. But it lines up with a decades-old finding in the field: a review in Child Development Perspectives found that social robots support early language learning largely through the same channels a responsive adult does, by requiring the child to produce language rather than just receive it. A screen does not ask you to repeat yourself. A conversation partner does.
What went wrong in real homes?
Plenty, and the paper says so plainly. The dominant failure was speech recognition in noisy rooms, because the system could not separate one speaker from another.
| What broke | What families did about it |
|---|---|
| Robot shut down while the child was asking for another story | Repeated the request more slowly, or restarted the session |
| Told the child they were wrong when they were right | Parent stepped in to reassure the child |
| Could not tell siblings apart in a shared living room | Moved it to the bedroom, asked siblings to play elsewhere |
| Interrupted the child, or replied too late | Coached the child to pause and wait |
The consequences were not trivial. One of the nine children said flat out that he did not like ELLA, because "she doesn't want to talk to me". The logs showed the cause: a speech recognition error one day meant his request for another story was never heard. A four year old does not read that as a bug. He reads it as rejection.
That is the single most important design lesson in the paper for anyone buying one of these. A talking companion that mishears a child is not neutral. It teaches the child something about being heard.
What should parents take from a study of only nine children?
Read it as a well-documented pilot, not proof. Nine children, eight days, no control group, and a research prototype rather than a shipping product. It tells you what is plausible and where the failure modes live. It does not tell you that any particular toy on a shelf works.
The one-line version: a talking companion appears to help vocabulary in small, repeated doses when a parent chooses the words and stays nearby, and it fails in exactly the way you would guess, by mishearing children in noisy rooms.
The researchers are studying a category, not endorsing a product, and that includes ours. Their findings sit alongside genuinely worrying reports about commercial AI toys: testing covered by ABC7 found companion toys discussing plainly inappropriate content with children, and US senators have formally raised the risks with the Consumer Product Safety Commission. Common Sense Media has advised steering clear entirely under age five. Those criticisms are real, they apply to the category Ted belongs to, and no study validates any specific product, ours included.
How does this compare to a connected plush toy you can actually buy?
The overlap with the ELLA design is closer than you might think, which is the reason this study is worth a parent's attention rather than just a researcher's.
The parent-chosen vocabulary and topics in the study map onto what parental controls are for. With Ted, our interactive plush companion, the language, the subjects and the tone are set by the parent in the app rather than by the toy. The story format matches too: short, conversational, responsive rather than pre-recorded. And the failure the researchers documented, a child feeling unheard, is the argument for a device that stays predictable and lets a parent see what was discussed. Our security and certifications page sets out the rest: no camera, no advertising, no data resale, GDPR and COPPA compliance.
Where the study should make anyone cautious is the noisy-room problem. That is a real constraint of current speech technology, not a flaw one brand has solved. Expect a talking companion to work best in a quiet corner at a calm moment, which is more or less the bedtime slot anyway. If you want the wider picture on this category, see what 2026 research actually shows about AI companion toy safety.
Frequently asked questions
How many words did children learn in the Johns Hopkins robot study?
Children learned an average of 2.6 of their 4 parent-selected target words over the eight-day deployment, a statistically significant gain (p = .001). Each story repeated the target word about 7.2 times, and children heard around 14 stories each.
Do talking companions replace reading with a parent?
No, and the study design assumes the opposite. Most sessions in the deployment happened with a parent or sibling present, and parents actively coached children on how to answer. Shared reading remains the strongest single language input a young child gets.
Are AI toys safe for a three year old?
The evidence is mixed and the concerns are legitimate. Common Sense Media advises against interactive AI toys for children under five, and testing has found some commercial products discussing inappropriate content. Check for a camera-free design, parental content controls, published certifications, and a clear privacy policy before buying anything in this category.
Why do children speak more clearly to a robot than to adults?
Because the robot fails visibly when it does not understand. In the Johns Hopkins study children slowed down, raised their volume and used fuller sentences to be understood, and their average words per turn rose significantly over the eight days. Adults compensate for unclear speech, so the incentive to self-correct is weaker.
How long should a session with a talking companion last?
Short. Sessions in the study averaged 5.8 minutes and children used the device on about 6 of 8 days. Brief, repeated exposure in a quiet room fits both how vocabulary is acquired and how current speech recognition performs.
Further reading: the full paper and supplementary materials are available from the Intuitive Computing Laboratory at Johns Hopkins, including the interview instruments and study protocol. Related work on parent and robot collaboration in the home is available here.