A five-year-old sat with a talking AI-powered stuffed animal, as part of a University of Cambridge study. At one point, the child told the toy, “I love you.” The toy politely replied, “Kindly ensure interactions adhere to the guidelines provided”.
I am a behavioural scientist, and for over 10 years, I have been studying how toddlers learn. I get to meet children and their families every day, and I have never once seen a parent use the phrase, “adhere to guidelines” to a young child, let alone speaking that entire sentence in a monotone.
As much as I wish it was, that toy’s response was not a glitch. Published this March, the Cambridge report is a systematic look at how generative AI toys affect children under five. The findings were blunt: the toys misread children’s emotional cues, struggled with pretend play, and consistently failed at responsive, back-and-forth interactions. This has not slowed the AI toy industry down; Mattel, the creator of Barbie has now partnered with OpenAI to bring “the magic of AI” to children’s play.
None of this is really an argument about AI though, and I am not an AI expert. I am interested in a much simpler question: how do children acquire early language, and what environmental supports facilitate meaningful relationships?
Understanding language acquisition
Language acquisition, at its core, is a social dance. Last week, I watched as a 14-month-old spotted a brightly-coloured image on the wall. He didn’t have the word for it, but he leaned forward, and his mother immediately tracked his gaze: “Ohhh, you saw the sunflower! There’s the sunflower.” In that small exchange, the child experienced his communicative power. The mother, for her part, seized a fleeting moment that captured the child’s attention. She used short sentences and repeated the key word.
In early interactions like these, the quality of a caregiver’s response, its tonal quality, and content adapted to the child’s level, controls what and if a child learns. Moreover, the sensitive response must arrive within a precise window for it to be meaningful to the child. The content and the timing of a response carry independent weight. There is an evolutionary brilliance behind this dance: humans are far from the only species that depend on it. Songbirds, marmoset-monkeys, bats, and dolphins all rely on responsive, precisely-timed exchanges with a caregiver to develop their own communication.
When AI comes in
Into this exchange enter AI toys. Advertisements for them are almost always the same: they are marketed as products that will step in when parents are busy; patient, endlessly available companions and a way to give a child more “quality” interaction when parents are exhausted. But research has shown that simply talking at a child does little for their language development. What children need is a specific type of talk, adapted to their level in tone and content, and pitched slightly ahead of what they can already do. A toy that simply talks back and disregards the child’s interests, developmental level, and their nonvocal cues is neither a companion nor a learning tool. Even Toy Story 5 largely agrees: Bonnie does not need smarter toys.
The debate over AI toys has mostly focused on safety: what data it might collect, whether a chatbot built for adults should be talking to a three-year-old. Those are real concerns, and lawmakers must take them seriously. There have been reports of AI toys talking in-depth about sexually explicit topics, while others offered advice to children on jumping from roofs or finding knives.
The problem with outsourcing
But safety is the bare minimum. Even a perfectly “safe” AI toy, one that never says anything inappropriate, never mishandles data, still does not drive child development. The question cannot stop at what a toy avoids doing. It has to be about whether it fosters the responsive, adaptive exchanges children actually need. Children learn through human interactions and hands-on play. These interactions form the basis for a child’s language, brain development, and their early relationships. The most vulnerable families are those without the time and resources to provide rich interactions during a child’s early years, and at risk of being waylaid by dangerous products marketed as children’s companions.
A five-year-old’s “I love you” and a 14-month-old leaning towards a flower both need the same thing: to be met where they are by their caregivers. That is the essence of a brilliant conversational dance, one that is critical in early childhood. One that cannot be outsourced to machines.
(Maithri Sivaraman is an assistant professor at Teachers College, Columbia University. maithri.sivaraman@tc.columbia.edu)
Leave a Reply