By the age of 3, a child already knows something that took engineers years to articulate: the eyes are not just organs of vision. They are organs of intention. A toddler watching an adult glance toward a bowl of strawberries will immediately grasp, without being told, that this person wants the strawberries. The inference is fast, automatic, and so reliable that developmental scientists have spent decades mapping its emergence. But new research from an Italian-Japanese collaboration has added a wrinkle that robot designers may find uncomfortable. Swap the adult out for a humanoid robot, one with eyes, a face, a head that turns, all the right equipment, and the inference evaporates entirely.
The finding comes from a study of 58 preschool children in Milan, aged between 3 and 5, who watched short videos of either a person or a robot named Robovie gazing at one of two objects. Robovie, for the record, is not some cartoonish machine with blinking LED eyes. It occupies what researchers describe as the mid-range of the mechanical-to-human continuum: genuine facial features, an approximation of a mouth, eyes that move. Not a vacuum cleaner. Not quite a person either.
The children were asked a simple question after each video: which object does the gazer prefer? When the gazer was human, children answered reliably and correctly. They had, apparently without effort, converted the direction of someone’s gaze into a model of that person’s inner desires. Theory of mind researchers have a technical name for this skill: preference attribution via referential gaze. Children develop it early and exercise it constantly, reading adults’ eyes for information about the social world. What the Milan study reveals is that this skill, perhaps unexpectedly, does not transfer to humanoid robots. The children were not confused by Robovie. They simply did not treat its gaze as meaningful in the same way. The robot looked at an object; the children shrugged, or the cognitive equivalent thereof.
Presence, Not Just Plumbing
The instinct might be to say the children were simply not fooled. That they could tell a machine from a person and withheld the usual social interpretation accordingly. That would be a tidy explanation. But the data complicate it. The researchers also measured how many mental states children were willing to attribute to each agent, using a validated questionnaire that probes whether children think robots can feel, decide, imagine, or desire. Children gave robots some credit. Robovie was not treated as a toaster. The difference in mental state attribution between human and robot was real but not total, and yet the difference in gaze-based preference attribution was stark. The children could, in some sense, believe the robot had an inner life, while simultaneously refusing to read its eyes for evidence of it.
Crucially, neither the human gaze nor the robot gaze changed what the children themselves wanted. If a child preferred the broom over the flag before the experiment, watching an adult stare at the flag for six seconds did not shift that preference. Gaze appears to function, at least at preschool age, as a social inference tool rather than a persuasion mechanism. Children use it to model what others are thinking, not to update their own desires. The human eyes were informative; they were just not contagious.
“This does not mean robots cannot play an educational or social role,” said Antonella Marchetti, Director of the Department of Psychology at Università Cattolica and one of the study’s lead researchers. “However, it suggests that simply imitating a single human signal, such as gaze, in a robotic artifact is not enough to make it truly communicative in a child’s eyes. Designing robots and intelligent technologies for children requires richer, more natural, and developmentally appropriate interactions: made up of words, gestures, reciprocity, context, and shared presence.”
What Adults Can Do That Children Cannot
There is a developmental wrinkle here that the researchers flag carefully. A companion study with adult participants, using the same paradigm, found that adults could read preference from both human and robot gaze. Given enough life experience with artificial agents, people learn to extend the gaze-reading reflex outward, past the boundary of the human face. The preschoolers in Milan have not yet crossed that threshold. Whether this happens gradually, through exposure to devices that have faces but not persons behind them, or whether it requires some more specific cognitive shift, is an open question. Preschoolers can follow a robot’s gaze, tracking the head movement, looking at the object. But they do not, apparently, complete the next inferential step of concluding that this gaze reveals something the robot wants.
One plausible account is that the children lack not the cognitive machinery but the accumulated social evidence. They have spent years around people whose eyes consistently deliver reliable information. They have, in effect, learned to trust the human gaze through thousands of interactions. Robovie has given them no such track record. Its eyes move, but they have never been proven to mean anything. Familiarity, in this account, matters as much as anthropomorphism. The researchers note that unfamiliar humans may also exert weaker gaze effects on children than familiar adults would. Add novelty and non-humanity simultaneously, and you get the result they found: a social signal that goes unread.
The study also found something about the relationship between gaze comprehension and theory of mind that runs against intuition. Passing the classic false-belief task, the standard measure of whether a child can represent that another person holds a mistaken belief, was not associated with better gaze-based preference attribution. Children who understood false beliefs were not better at reading human gaze for preferences than those who did not. What did predict preference attribution was a broader willingness to attribute mental states in general, emotions, intentions, epistemic states, not just beliefs. This suggests that reading another person’s gaze for preference is less a theory-of-mind achievement than a more diffuse form of social attunement, one that is developing throughout this age range rather than arriving in a single cognitive step.
Designing for How Children Actually Think
The practical stakes are larger than they might seem. Humanoid robots are increasingly proposed for educational settings, language learning, and, perhaps most sensitively, therapy for children with autism spectrum disorder. Gaze and shared attention are among the skills most vulnerable in autistic development, and robots have been studied as a lower-pressure context in which to practice them. If preschool children do not treat robot gaze as communicatively meaningful, designing interventions around robotic gaze cues alone would be working against children’s natural social cognition. The team behind this study is involved in a forthcoming project, launching in June 2026 and funded by the Italian Ministry of Health, that will use humanoid robots with autistic children to develop imitation and gaze-following skills. Marchetti’s findings feed directly into how that work will be framed.
The broader implication, as Marchetti puts it, is about what communication actually requires. Language models now speak, respond, recommend. They do so fluently and at scale. But communication, at least for young children, seems to be something more than the correct deployment of signals. It depends on a kind of shared presence, a sense that the signals are being produced by something that intends them, and that this intention can be trusted because it has been demonstrated before. Whether robots can accumulate that kind of trust, or whether children must simply grow old enough to extend it on thinner evidence, is a question the next generation of child-robot interaction research will have to answer.
https://doi.org/10.1016/j.ijcci.2026.100822
Frequently Asked Questions
Why can children read a human’s gaze but not a robot’s gaze?
Children appear to treat human gaze as a reliable signal of inner mental states because years of social experience have confirmed it as such. A robot like Robovie moves its eyes and head in similar ways, but children have no comparable track record of its gaze meaning anything. The gap may be less about recognizing the robot as non-human and more about having no learned reason to trust its eyes as windows to intention.
Could this change as children grow older and spend more time around robots?
The evidence suggests yes. A companion study found that adults could infer preference from both human and robot gaze, suggesting the capacity does develop over time. Whether that shift comes from sheer familiarity with artificial agents or from some broader cognitive change in how we extend social interpretation is not yet clear, and it is one of the key questions researchers plan to pursue.
Does this mean robots are useless for teaching young children?
Not at all, but it does mean that relying on gaze alone is insufficient. Children appear to need richer, more multimodal interaction from robots, including words, gestures, contingent responses, and repeated interaction, before a robot’s nonverbal cues carry real communicative weight. Robots that speak, gesture, and respond dynamically are a different proposition from one that simply looks at an object.
Why doesn’t watching someone gaze at something change what a child wants?
Gaze, at preschool age, appears to work as a social inference tool rather than a persuasion mechanism. Children use another person’s eyes to model what that person wants, but this does not automatically update their own preferences. Desire, it seems, requires more than observation; it needs something closer to endorsement from someone the child has reason to trust.
ScienceBlog.com has no paywalls, no sponsored content, and no agenda beyond getting the science right. Every story here is written to inform, not to impress an advertiser or push a point of view.
Good science journalism takes time — reading the papers, checking the claims, finding researchers who can put findings in context. We do that work because we think it matters.
If you find this site useful, consider supporting it with a donation. Even a few dollars a month helps keep the coverage independent and free for everyone.