Giving AI Eyes Doesn’t Make It See

A multimodal AI can look at a kitchen photo and identify the bowl, spoon, chair, and suspiciously ripe banana.

A toddler looks at the same kitchen and sees a drum kit, a climbing challenge, and something that may become lunch if it survives being dropped.

The AI may win the labeling contest. The toddler is doing something messier—and possibly more important. They are figuring out what the scene allows them to do.

This is why “AI can see now” deserves a small asterisk. Fine, perhaps a medium one.

A camera feed is not a childhood

Human vision develops inside a loop: look, reach, miss, adjust, try again. An object is not merely a patch of color. It can be grasped, hidden, shared, avoided, or used to make an adult say, “Please stop licking that.”

A recent neuroscience perspective argues that multimodal language models face a grounding problem because they process representations of the world without having the embodied agency and developmental history through which humans build physical understanding. Images plus words may produce excellent pattern recognition, the authors argue, while still falling short of causal, action-based knowledge ("Will Multimodal Large Language Models," 2025).

That distinction matters. A model can associate “glass” with “fragile.” A child can pick up a cup, feel its weight, hear the warning in a caregiver’s voice, and learn that tipping it changes both the table and the emotional weather in the room.

Same concept label. Very different curriculum.

Of course, embodiment is not magic pixie dust. Put a language model in a robot and you have not automatically created understanding. You may simply have created a chatbot that can knock over your lamp.

The useful claim is narrower: action supplies kinds of evidence that passive observation does not. When a learner acts, it can test a prediction. If I push this, will it slide? If I look behind the box, is the toy still there? The world answers back.

Seeing requires remembering what disappeared

Real environments are stingy with information. The important thing is often behind you, under the sofa, or no longer visible. Cognitive scientists call this partial observability: the current view does not reveal the full state of the world.

That sounds technical. Parents call it “Where did you put your other shoe?”

Pedamonti and colleagues tested artificial agents on navigation problems in which relevant information was partly hidden. Agents with recurrent, hippocampal-like circuitry learned across these conditions more successfully than feedforward systems, and their internal activity reflected information about strategy, reward, and time in ways that resembled recordings from rats (Pedamonti et al., 2025).

The interesting lesson is not that researchers have built a tiny synthetic hippocampus. They have not. It is that vision alone was insufficient. The agent needed a memory-informed process that carried useful information forward when the scene stopped providing it.

That is closer to how biological seeing works. Perception is not a sequence of disconnected screenshots. What you see now is interpreted using what just happened, what you expect next, and what you are trying to do.

So when an AI describes a picture beautifully, we should applaud the actual achievement without sneaking in a larger one. Recognition is not automatically physical understanding. Captioning is not agency. And “it noticed the cup” does not mean “it knows what will happen if its elbow hits the cup.”

Children do not need vision benchmarks

What does any of this mean at home? Thankfully, not that you need to turn playtime into a robotics lab. Children already run the relevant experiments, usually on your furniture.

A few useful principles follow:

• Let looking become doing. Stacking, pouring, rolling, opening, carrying, and building connect visual patterns to consequences. A picture of a ramp is informative. Sending a toy down one—and changing the angle—is a causal investigation.

• Ask for predictions before explanations. “What do you think will happen if…?” gives a child a chance to commit to a model, then revise it when reality objects. Reality is an excellent reviewer and has no concern for anyone’s feelings.

• Treat mistakes as data. A toppled block tower is not failed seeing. It reveals something about balance, weight, or an overconfident sibling. The feedback loop is the lesson.

• Be cautious with claims that an app ‘understands’ what your child shows it. The system may identify objects or generate plausible descriptions without sharing your child’s goal, history, or practical knowledge of the scene. Useful? Absolutely. Equivalent to human perception? Not so fast.

The boring middle here is more interesting than either hype extreme. Multimodal AI is not “just autocomplete with pictures.” These systems can combine visual and linguistic patterns in impressive ways. But neither does adding images close the gap between processing records of the world and learning by living in it.

A child’s vision is built through a long negotiation with gravity, memory, other people, and breakable cups. AI has the pixels. It is still working on the negotiation.