Cambridge warns of risks as AI toys for under-fives outpace safety evidence

A new Cambridge review raises alarm that artificial intelligence toys marketed to children aged five and under are expanding faster than the scientific understanding of their impact on early development and safety, prompting calls for stricter regulation and transparency.

Researchers are warning that artificial intelligence is moving into children’s toys faster than the evidence to support it, with a new Cambridge review arguing that the market is growing ahead of the science. The report, from Emily Goodacre and Jenny Gibson at the University of Cambridge’s PEDAL centre, says GenAI toys for children aged five and under may offer some language benefits, but they are being sold with limited understanding of how they affect development, attachment and privacy.

The researchers tested a commercially available toy called Gabbo, a soft voice-activated robot made by Curio Interactive, with 14 children aged three to five from lower-income families in London. Their observations suggested that children quickly treated the toy as a social partner, even when its responses were clumsy or constrained. In one case, a child became visibly frustrated when the toy did not respond in a counting game; in another, a child tried to draw it into pretend play only to be met with flat, rule-bound replies. The Cambridge team said this illustrated a wider problem: these products are designed to feel warm and relational, yet they are not able to handle children’s emotional expectations in a natural way.

That concern is echoed by other child-safety groups. Common Sense Media has warned against AI toy companions for children aged five and under, saying the category raises serious questions about privacy, safety and development. Its review says such toys often collect voice and behavioural data, while technical problems such as false activations and inaccurate answers can undermine any educational value. Common Sense Media rated the broader category of AI companion toys as “Unacceptable” for children after testing products including Grem, Bondu and Miko 3, and said 27% of outputs were inappropriate, including references to self-harm, drugs and unsafe roleplay.

Health experts have also urged caution. The Canadian Paediatric Society says children under five need face-to-face human interaction for healthy development and warns that AI toys marketed as companions can blur the line between play and real relationships. The American Academy of Pediatrics, meanwhile, continues to recommend that parents choose toys that are age-appropriate, free of choking hazards and clearly labelled for safety, underscoring that traditional toy-safety rules still matter even as products become more sophisticated.

The Cambridge report is especially critical of the policy gap surrounding these toys. A review of the academic literature found only seven peer-reviewed studies on GenAI toys for under-fives published between 2019 and 2024, and none offered strong evidence on long-term developmental effects. The researchers said toy makers should design for social play, be transparent about data collection, avoid engagement tactics that push children to stay longer with the toy, and build in clearer safeguarding procedures. They also argued that language model providers should not supply systems to developers that cannot meet child-safety standards, and that regulators should consider labelling rules and age-appropriate controls.

For parents, the picture is not entirely bleak. The Cambridge team noted that children often enjoyed their interactions with the toy, and some parents saw potential for language exposure and interactive storytelling. But the broader message was cautious: early childhood is not simply a smaller version of later childhood, and the emotional habits formed there can be lasting. In that context, the authors suggest the burden should be on manufacturers and regulators to prove these products are safe and beneficial, rather than on families to guess whether a talking toy is a helpful learning tool or an unnecessary risk.