By the time they reach the age of 3, children already comprehend something that took engineers many years to express: the eyes function not only as organs for seeing but also as organs of intention. A toddler observing an adult’s glance towards a bowl of strawberries will quickly understand, without any verbal cues, that the person desires the strawberries. This inference happens swiftly, automatically, and is so dependable that developmental scientists have dedicated decades to examining its development. However, new findings from an Italian-Japanese study introduce a complication that may unsettle robot designers. Replace the adult with a humanoid robot—equipped with eyes, a face, and a head capable of turning—and the inference completely disappears.
The discovery stems from research involving 58 preschool children in Milan, aged 3 to 5, who viewed brief videos featuring either a human or a robot named Robovie gazing at one of two objects. For clarity, Robovie is not a whimsical machine with blinking LED eyes. It occupies what researchers categorize as the intermediate range on the mechanical-to-human spectrum: authentic facial characteristics, an approximation of a mouth, and eyes that move. It is neither a vacuum cleaner nor quite a person.
After each video, the children were posed a straightforward question: which object did the gazer prefer? When the gazer was human, the children provided consistent and accurate responses. They had, seemingly without difficulty, interpreted someone’s gaze as a reflection of that person’s inner wants. Researchers studying theory of mind refer to this ability as preference attribution via referential gaze. Children develop this skill early and continuously utilize it, deciphering adults’ eyes for insights about the social realm. What the Milan study uncovers is that this ability, perhaps unexpectedly, does not extend to humanoid robots. The children were not puzzled by Robovie; they simply did not consider its gaze as meaningful in the same context. When the robot looked at an object, the children shrugged it off, or responded in a cognitively similar way.
Presence, Not Just Plumbing
The instinct may be to conclude that the children were merely not deceived. They could distinguish between a machine and a person and adjusted their social interpretations accordingly. That would be a neat explanation. However, the data introduces complexity. Researchers additionally evaluated how many mental states the children were inclined to attribute to each agent, utilizing a validated questionnaire that examines whether children believe robots can feel, make decisions, imagine, or have desires. Children afforded robots some recognition. Robovie was not regarded as a mere kitchen appliance. The disparity in mental state attribution between human and robot was significant but not complete, and yet the contrast in gaze-based preference attribution was pronounced. The children could partially accept that the robot possessed an inner life, while concurrently refusing to interpret its eyes as indicators of that life.
Importantly, neither the human gaze nor the robot gaze altered what the children themselves preferred. If a child favored the broom over the flag prior to the experiment, observing an adult fixate on the flag for six seconds did not change that preference. Gaze appears to function, at least at the preschool age, as a social inferencing mechanism rather than an influence tool. Children deploy it to model what others are considering, not to modify their own preferences. The human gaze was informative; it just wasn’t contagious.
“This does not imply that robots cannot serve an educational or social function,” stated Antonella Marchetti, Director of the Department of Psychology at Università Cattolica and a lead researcher in the study. “Nevertheless, it indicates that merely mimicking a single human signal, such as gaze, in a robotic entity is insufficient to render it genuinely communicative in the eyes of a child. Creating robots and intelligent technologies for children necessitates richer, more natural, and developmentally appropriate interactions that encompass words, gestures, reciprocity, context, and shared presence.”
What Adults Can Do That Children Cannot
There is a developmental nuance that the researchers carefully highlight. A complementing study involving adult participants, employing the same framework, determined that adults could interpret preference from both human and robot gaze. With ample life experience interacting with artificial agents, individuals learn to expand the gaze-reading reflex beyond the limits of the human face. The preschoolers in Milan have not yet reached that milestone. Whether this development occurs gradually through exposure to devices with facial features but no persons behind them or necessitates a more specific cognitive transition remains an open question. Preschoolers can track a robot’s gaze, following its head movements and focusing on the object. Yet, they do not, for the most part, reach the next inferential level of deducing that this gaze signifies something the robot desires.
One plausible explanation is that the children lack not cognitive capabilities but the accumulated social experiences. They have spent years surrounded by individuals whose eyes consistently convey trustworthy information. They have, effectively, learned to rely on the human gaze through countless interactions. Robovie has not provided them with such a history. Its eyes move, but their significance has never been confirmed. Familiarity, in this context