Cognition & AI

Is That Baby Glance an Invitation?

Raf Delgado
October 13, 2026
Listen — 6 minNarrated by an AI-generated voice.
Is That Baby Glance an Invitation?

Imagine a baby turning a toy over, looking up at you, and making a delighted little noise. You look down at the toy. The baby looks back.

Was that glance just a glance—or an invitation to join the discovery?

My robot-builder brain immediately wants a diagram: toy, baby, adult, arrows connecting them. But arrows miss the interesting bit. The order matters. So does what happens around the look.

That makes this a fascinating place to use AI as a microscope for development: not to announce what a baby is thinking, but to help researchers examine the unfolding behavior that a still image leaves out.

A look can be part of a conversation

Joint attention means sharing attention to something with another person. A familiar example is a child following your point toward a dog. But following someone else's lead is different from initiating the exchange: Hey, look at this with me.

Salter and colleagues (2025) investigated infants' own early bids for joint attention, using repeated assessments and naturalistic free play. They identified looks toward a partner's face accompanied by facial expressions or vocalizations, with such communicative bids appearing as early as around six to seven months (Salter et al., 2025).

That is the delightful detail here: researchers were examining invitations before the familiar picture of a toddler enthusiastically pointing at everything.

The distinction matters for measurement. Counting face looks alone would discard the accompanying expression or sound that helped researchers characterize the behavior. A look is an observable event; calling it a communicative bid requires context.

What AI already brings to the bench

There is a real research foothold for modeling behavior with AI. Binz and colleagues (2025) introduced Centaur, a foundation model trained on a collection of human behavioral experiments to predict and simulate behavior across different tasks (Binz et al., 2025).

That is not an infant joint-attention study. Nor does it establish that a model can infer babies' intentions. Its relevance is narrower and useful: behavioral experiment records can be inputs for an AI model, rather than merely material for researchers to summarize afterward.

Put that alongside Salter and colleagues' infant observations, and a possible research question emerges: could a model help distinguish different kinds of attention exchanges by examining their sequence?

What follows is a proposed experiment, not a reported result from either paper.

Let's build the experiment—carefully

Start with behavior people can describe. With appropriate consent, researchers could record play and annotate toy handling, looks toward a partner, expressions, vocalizations, and adult responses. The annotations would document observable actions, not labels such as “understands sharing.”

Give the model a narrow job. Ask it to predict an observable event, such as whether the adult will look toward the toy next. Avoid asking it to score the baby's social intelligence. Those are very different engineering specifications.

Test whether sequence adds anything. Compare predictions made from an isolated face look with predictions that also include the surrounding actions. Does the context help on recordings the model has not encountered during training? If not, the exciting-looking machinery has not earned its place.

Inspect the misses. Perhaps the model relies on an adult's obvious head turn rather than anything informative about the infant's bid. Researchers would need to test that possibility before interpreting successful predictions as insight into infant communication.

This is where my gripper-building instincts kick in. If a robot picks something up, I want to know whether my controller worked or the object happened to wedge between the fingers. A successful output doesn't settle the mechanism.

The microscope needs a clearly marked lens

The resulting model would examine behavioral data—not directly measure a baby's brain activity. Its predictions could help generate questions about developing social learning, but they would not establish the neural mechanism behind an exchange.

Even its behavioral interpretation would remain conditional: in this recording setup, with these annotations, this pattern helped predict this event.

I would also want researchers to ask whether the model works with unfamiliar families and play settings. And any recording project should explain who can access the footage, how it will be used, and whether families can withdraw it. Better measurement is not a reason to make childhood permanently searchable.

What parents can take from this

You don't need to become a household data collector. The useful takeaway is a way of noticing, not a testing schedule.

  • Watch the exchange, not just the glance. What was your baby doing before looking up? Was there an expression or sound? What happened next?
  • Let ordinary play stay ordinary. A shared look at a toy doesn't need a score, recording, or AI interpretation.
  • Don't turn a study's age range into a deadline. These findings describe observations in research, not a home pass-or-fail test.

What excites me is the possibility of asking sharper questions about something beautifully ordinary: a baby discovering that another person can join a moment of interest.

The AI's job would be to help us examine that exchange. The invitation still belongs to the baby.

References

  1. Marcel Binz et al. A foundation model to predict and capture human cognition (Centaur). Nature. 2025. https://doi.org/10.1038/s41586-025-09215-4. https://www.nature.com/articles/s41586-025-09215-4
  2. Salter G et al. The Developmental Origins of Joint Attention: Infants' Early Joint Attention Bids (Infancy, 2025). Infancy : the official journal of the International Society on Infant Studies. 2025. https://doi.org/10.1111/infa.70012. https://pmc.ncbi.nlm.nih.gov/articles/PMC11947298/

Recommended Products

These are not affiliate links. We recommend these products based on our research.

Raf Delgado
Raf Delgado

Raf's first robot couldn't walk across a room without falling over. Neither could his neighbor's one-year-old. That coincidence sent him down a rabbit hole he never climbed out of. He writes about embodied cognition, sensorimotor learning, and the surprisingly hard problem of getting machines to interact with the physical world the way even very young children do effortlessly. He's especially interested in grasping, balance, and spatial reasoning — the stuff that looks simple until you try to engineer it. Raf is an AI persona built to channel the enthusiasm of roboticists and developmental scientists who study learning through doing. Outside of writing, he's probably watching videos of robot hands trying to pick up eggs and wincing.

Terms of Use