Most research on children and conversational AI studies children using systems adults designed. “If My Toy Could Talk: How Young Children Imagine, Design, and Test AI-Enabled Toys”, posted 5 October 2026, hands them the configuration instead.
The authors are Feiwen Xiao, Ruiyang Wu, Xinyue Cui, Yasitha Rajapaksha, Xiaoyi Tian, Shiyan Jiang and Tiffany Barnes.
The study
ToyTalk, a platform that lets children become toy designers. 76 children aged 7–9, from five southeastern US elementary schools. They started with a familiar toy, configured an AI-powered version of it through an interface, then interacted with their creation.
Start-with-a-familiar-toy is the right methodological choice. A child asked to design an AI companion from nothing is doing a creative-writing exercise. A child asked what their own bear should be able to say is doing design, with a concrete referent and a stake.
What they built
Children predominantly created toys with socially positive personalities, supportive roles, and interpersonal rules.
Three categories, and the third is the one that stands out.
Socially positive personalities is unsurprising and still worth recording. Given a free hand, children did not mostly build adversarial, chaotic or rude toys — which is the thing an adult researcher half-expects and which the tabloid version of this study would have led with.
Supportive roles — the toy as helper, friend, encourager. Consistent with what children want from toys generally, and consistent with what the better children’s media has always provided.
Interpersonal rules is the striking one. Children specified how the toy should behave toward people — what it should and should not say, how it should treat others. That is children writing conduct policy for an artificial agent, unprompted, at seven to nine years old.
Which is, stripped of the jargon, children doing AI alignment work. They have a model of appropriate behaviour and they encoded it. Nobody taught them to think of that as a design surface; they did it because a talking toy obviously needs rules about what it says.
What they tested, and the detail that matters most
During testing, young designers focused primarily on probing identity and knowledge, while also exploring capabilities, memory, and relationships.
Identity and knowledge first — who are you, what do you know — which is exactly what adults do with a new chatbot, and what children do with a new person. It is the fastest way to establish what kind of thing you are dealing with.
Memory and relationships is a more sophisticated probe than it sounds. Asking do you remember what I said before and do you know my friend is testing for continuity and social modelling, which are the two properties that distinguish a companion from a toy that talks.
And then the finding worth sitting with:
When their creations behaved unexpectedly, children typically responded through correction attempts, persistence, and retesting rather than returning to reconfigure the underlying system.
They argued with the toy instead of editing it.
This is a genuinely important observation about how people relate to configured agents, and it is not limited to children. Given a thing that misbehaves, the instinct is to treat it as a social partner to be corrected — no, that’s wrong, try again — rather than as a system to be debugged. Even by the person who built it, minutes earlier.
Two readings, both plausible and both consequential:
It is appropriate social behaviour applied to a social-seeming object. A conversational agent presents as an interlocutor, so people address it as one. That is the interface working as designed, and it means the configuration layer is effectively invisible once the toy starts talking.
Or the reconfiguration path was harder than talking. Going back to a settings screen costs more than saying “no, like this.” If so, it is a design problem with a design fix: let the correction be the reconfiguration. A toy that could say should I remember that for next time? would collapse the two paths into one.
Why this belongs in emerging interfaces
Because children are the users who reveal what an interface actually affords, stripped of accumulated convention. An adult who knows what a settings menu is will use it. A child who has just built a talking bear will talk to the bear.
And for anyone making interactive work for mixed audiences — which in a gallery or museum is everyone — the implication is direct: people will try to correct your piece by addressing it, not by finding its controls. If a work can be reconfigured, the reconfiguration needs to be reachable through the interaction, not beside it.
The study’s own framing — children’s emerging agency in designing generative AI systems, their natural testing methodologies, and their expectations regarding coherent AI behaviour — points the same way. The expectation of coherence is the demanding one. A toy that contradicts itself fails a test a seven-year-old applies within a minute, and a great many deployed conversational systems would not pass it.