76 children aged between 7 and 9 are designing an AI version of their favourite toy: what they expect from it

Using the ToyTalk tool, pupils set up a chatbot for their toy and then test it: they want caring companions, but rarely reconfigure the chatbot when things go wrong.

Researchers have developed ToyTalk, a technology probe that puts children in the role of toy designers with a language model. Each child starts with a familiar toy, configures its "smart" version via a no-code interface, and then chats with the character to test it. The tool was rolled out to 76 students aged 7 to 9 in five elementary schools in the southeastern United States.

Caring companions, tested relentlessly

The children mostly imagined toys with a positive personality, in support roles, with rules of behavior towards others. In conversation, they mainly questioned the character's identity and knowledge, but also tested his abilities, his memory and the relationship he has with them.

When the toy did not behave as expected, the children corrected, insisted and retested rather than returning to modify the configuration. The authors see this as a way to think about children's ability to act in the face of generative AI, their ways of testing a system and their expectation of a coherent character, all of which are direct issues for the design of products for the youngest.