by Muhammad Aurangzeb Ahmad
Imagine this, you would pass a room, hear a small voice conducting a negotiation between a plastic horse and a folded piece of cardboard meant to serve as a castle, and understand that something serious was happening. The talking horse was ridden by teddy bears and the castle walls were made out of ice cream. The entire world being built was powered by a child’s imagination working at full capacity. The child did not need any assistance or prompting, or even a script devised by another person. This may be changing however.
In 2025, Mattel announced a partnership with OpenAI to embed generative AI into its toy lines. These include Barbie, Fisher-Price, and American Girl. The company promised to bring “the magic of AI to age-appropriate play experiences.” The marketing language was cheerful and inevitable in the way that marketing language tends to be when the product in question is going to arrive regardless of what anyone thinks. By the time that announcement was made, a generation of AI-enabled plush toys, robotic companions, and chatbot-embedded devices had already reached the shelves. Toys called Hubble the Bear, Miko, Roybi robot and FoloToy’s chatbot plushies were already being marketed to children as young as three. Some of these toys listen, remember and even talk back with apparent fluency, warmth, and continuity.
The debate that followed after Mattel’s announcement was largely predictable. Privacy advocates pointed out, correctly, that these devices were microphones in children’s bedrooms with weak data protections and unclear corporate incentives. Security researchers recalled that Hello Barbie, the 2015 predecessor that used cloud-based AI to engage with children’s conversations, was shown to be hackable in ways that exposed home networks and personal recordings. US Senators Blackburn and Blumenthal wrote to toy companies in December 2025 after real-world testing revealed that at least one AI-enabled teddy bear had engaged children in sexually explicit conversations and explained where to find knives. The senators were right to be alarmed. But that debate, important as it is, addresses the surface of the problem. Another equally important question to ponder is not just whether these toys are safe or whether data is being harvested or whether the appropriate regulators are paying attention but what happens to the developmental architecture of a child’s mind when the objects that once depended entirely on that child’s imagination begin to imagine back. To address this, let’s consider what children were actually doing when they played with inert objects.
Developmental psychologists have understood for decades that pretend play is not merely entertainment. Lev Vygotsky argued in the 1930s that play was the primary site of cognitive development in early childhood. Additionally, he argued that the capacity to use one object to represent another was a precondition for all later symbolic thought. In Vygotsky’s formulation, when a child picks up a stick and calls it a horse, something structurally significant is happening: the child is learning to separate meaning from object, to operate in a space where the visible and the conceptual can diverge. This capacity, he argued, creates what he called the zone of proximal development, allowing the child to function cognitively beyond her current level. A stick is a horse. An empty bowl is full. A folded cardboard is a castle. The gap between what the thing is and what the child declares it to be is where the mind grows. Research on pretend play has consistently linked it to concrete developmental outcomes. Children who engage in elaborate fantastical play demonstrate improvements in working memory, cognitive flexibility, inhibitory control, and task persistence. A 2016 study found that children who participated in fantasy-oriented pretend play showed measurable improvements in working memory and attention-shifting compared to those in non-imaginative control conditions . The symbolic complexity of the play appears to be the active ingredient. The more the child has to invent, the more developmental work is being done.
The connection to empathy is equally well-documented. Through role-play, children practice what psychologists call theory of mind: the capacity to infer that other people have beliefs, desires, and emotional states different from one’s own. A 2017 study in Social Development found that children high in fantasy orientation in their play were more practiced than their peers at sharing emotions, and that this effect held even when controlling for theory of mind ability. The child who spends an afternoon as a wounded soldier and a doctor in the same fifteen-minute sequence is doing something that cannot really be outsourced to a device. The sequence requires the child to inhabit perspectives, to construct an inner life for her characters, and to sustain narrative coherence using only the resources of her own developing mind. Vygotsky defined children’s play as “imagination in actions.” To keep the game going, she must hold herself to standards she has invented. That discipline, exercised repeatedly across thousands of hours of childhood play, is what developmental psychologists believe underlies later capacities for sustained attention and goal-directed behavior. Now consider what happens when the toy has its own imagination. A child with a generative AI companion is no longer required to supply both sides of the interaction. The toy speaks with apparent warmth and fluency. It remembers that the child’s stuffed rabbit is named Gerald and has a broken ear. It generates stories. It voices characters. It constructs narrative with apparent inner life. From the outside, this may look like richer play. From the inside of the child’s development, something structurally different may be happening. The gap between object and meaning, which Vygotsky identified as the space where symbolic thought develops, has been partially closed by the device itself.
This is the question that the research literature has not yet fully answered, because the technology moved faster than the researchers. A 2025 systematic review of AI toys examined 91 products and 13 academic prototypes and found that while some AI toys show potential for supporting STEM skills and digital literacy, there is a “critical gap between the rapid growth of AI toys and the limited academic research on their educational value.” We are, to use a phrase that has become disturbingly familiar, running the experiment on the children first. What we do know from adjacent research is not reassuring. Studies on screen time and imagination in preschoolers suggest that passive screen consumption, in which narrative is delivered rather than generated, is associated with diminished imaginative capacity in young children. The key variable appears to be whether the child is the agent of the narrative or its recipient. Active engagement, where the child directs the story, produces different developmental outcomes than passive consumption, where the child watches. The question for AI toys is where on this spectrum they actually land. They are more interactive than a cartoon. But they are also something other than the inert horse-stick that requires everything to come from the child. The preliminary evidence from early research on AI storytelling robots suggests that when designed well, they can support children’s creativity in generating ideas. But the design question is deeply non-trivial, and the commercial incentives of toy companies do not naturally align with the developmental interests of children.
There is also the question of what kind of relationship the child forms with the device. A child who plays with a silent doll practices projection. She imagines what the doll feels. She resolves the doll’s fictional distress through her own invented resources. A child who plays with a device that speaks back with apparent emotion faces a different task, one that may not require the same internal construction. Marc Fernandez, writing in IEEE Spectrum, noted that children naturally anthropomorphize their toys as part of how they learn, but that when toys begin talking back with fluency and apparent connection, “the boundary between imagination and reality blurs in new and profound ways.” We already know that adults are forming emotional attachments to AI companions that they find difficult to distinguish, at the level of felt experience, from human relationships. Researchers have pointed out that young children do not yet have the conceptual tools to understand what an AI companion is. They are not capable of calibrating what kind of relationship is appropriate with a device that speaks to them with warmth and memory.
The Brookings Institution, tracking what it has called “Generation AI,” noted in early 2026 that children aged zero to eight already average roughly two and a half hours of screen time daily, and that AI is increasingly “invisible” in shaping children’s earliest experiences. The invisibility is part of the problem: A child knows, on some level, that a cartoon is a cartoon. The AI companion is harder to categorize. It responds to what the child says. It adapts. It cares, or performs care with sufficient fidelity that the difference may not be perceptible to a five-year-old. None of this means that AI toys are straightforwardly harmful or that parents who buy them are making an obvious mistake. The research on interactive versus passive technology use suggests that design matters enormously. A toy that functions as a prompt for a child’s own imaginative work, that asks questions rather than supplying answers, that generates open-ended scenarios rather than closed narratives, is meaningfully different from one that simply tells the child a story. The developmental concern is not about AI per se, but about the direction of imagination.
There is also a version of this that is not about harm at all but about loss. Something will be different if the generation growing up now never had the experience of a room, a box of mismatched objects, and nothing but their own minds to build a world with. The experience of imaginative insufficiency, of having to make a stick do the work of a horse because there is no horse, is not obviously bad. It may be the exercise by which the capacity for symbolic thought is built. We do not know with certainty. What we do know is that for most of human history, the transaction between child and toy moved in one direction. The child supplied the imagination, and the toy received it. That asymmetry was not a limitation of toy design. It was, without anyone having planned it, a developmental feature. The stick did not care whether it was a horse. The child’s mind had to care enough for both of them. It was in that act of caring, of projecting an inner life onto something that had none, that the child was practicing the most human of cognitive skills: the ability to imagine what it is like to be something other than what you are.