Why a robot’s stare means nothing to a 3-year-old—and what that means for AI
A new international study shows that children as young as 3 can infer a person’s desires from their gaze, yet apply no such intuition to humanoid robots. The findings challenge how AI systems for children are designed.
Children are remarkably good at reading people. By age 3, a child can watch an adult glance at an object and silently conclude: that person wants it. But point a robot’s eyes at the same object, and that inference disappears entirely. This is the central finding of a study published in the International Journal of Child-Computer Interaction, led by Professor Antonella Marchetti, Director of the Department of Psychology at Università Cattolica and CERITOM (Research Center on Theory of Mind and Social Competences Across the Lifespan), in collaboration with researchers from Tokyo, Osaka, and colleagues Davide Massaro, Cinzia Di Dio, and Federico Manzi.
The study
As explained here, fifty-eight Italian children aged 3 to 5 were shown videos of either a person or a humanoid robot looking at one of two objects. Researchers then asked the children which object the agent preferred, and separately, which object the child themselves preferred.
The pattern was clear. When a human looked at an object, children consistently inferred that the person liked it — using gaze as a window into mental states like desire and intention. When a robot performed the identical action, children made no such inference. The gaze registered as movement, not meaning.
Notably, neither the human nor the robot gaze changed the children’s own preferences. Gaze helped children understand what another agent might want, but did not nudge them toward wanting the same thing themselves.
Why gaze works for humans but not robots
The underlying mechanism is rooted in how children develop a theory of mind — the understanding that other beings have inner mental lives. Children instinctively search for a mind behind human eyes. When they see a person look at something, they read intentionality into it.
A robot’s stare, even if physically identical, doesn’t trigger the same interpretive process. Children treat it as a mechanical event rather than a communicative act. The study also found that children’s ability to attribute mental states to humans — but not to robots — was a significant predictor of accurate preference attribution, while performance on false-belief tasks showed no such link. This suggests that reading gaze-based cues relies on social-cognitive processes that are distinct from explicit belief reasoning.
What this means for robot design
Professor Marchetti is careful to note that the findings don’t diminish robots’ potential role in education or social development. What they do reveal is a design flaw in how many robotic systems approach communication: “Simply programming a robot to mimic an isolated human signal like gaze is not enough to make it truly communicative in a child’s eyes.”
Effective child-robot interaction, she argues, requires richer, more layered engagement — combining spoken language, physical gesture, reciprocity, shared context, and sustained presence. The same principle applies to AI more broadly. Text-based or voice-only AI may be sufficient for many adult tasks, but for children, communication is deeply embodied. Presence and shared context matter.
This points toward the growing field of embodied AI — artificial intelligence integrated into physical, interactive systems. Humanoid social robots represent one of the most complete expressions of this approach, and may be the necessary vehicle for helping children attribute genuine mental states to technology.
Implications for autism spectrum interventions
The findings carry particular weight for clinical applications. Shared attention and gaze interpretation are core dimensions of early social development — and ones that can be especially vulnerable in children on the autism spectrum.
Researchers are increasingly exploring humanoid robots as rehabilitation tools for precisely these skills. Understanding exactly how — and when — a child interprets a robot’s gaze as intentional is essential for designing effective, sensitive interventions.
Building on this research, the ROBIN project (ROBot-based Neuropsychomotor INtervention to promote imitation skills in young children with autism spectrum disorder), funded by Italy’s Ministry of Health, will launch in June 2026. Led by the Don Carlo Gnocchi Foundation and CERITOM at Università Cattolica, ROBIN will use humanoid robots to support imitation skills and socio-communicative development in young children with autism, with a specific focus on how children interpret and respond to robotic gaze.
There is something quietly profound in a 3-year-old’s refusal to see a mind behind a robot’s eyes. Children, it turns out, are more discerning than we assumed — not fooled by surface mimicry, instinctively holding out for something deeper. For AI developers, that intuition is not an obstacle to overcome, but a standard to meet.
Original research: Manzi F., Ishikawa M., Di Dio C., Itakura S., Kanda T., Ishiguro H., Massaro D., & Marchetti A. (2026). Preschoolers attribute preferences in response to human but not robot gaze. International Journal of Child-Computer Interaction. DOI: 10.1016/j.ijcci.2026.100822. Source: Università Cattolica del Sacro Cuore
