Babies Learn Faster Than the World's Most Powerful AI. Scientists Want to Know Why.

Key points

• Researchers from Meta, Stanford University, the University of Tokyo, and France's École Normale Supérieure launched the EgoBabyVLM Challenge in 2025 to test AI against infant learning.
• Today's best vision language models, AI systems that understand both images and text, failed badly when shown about a thousand hours of real footage recorded from cameras on babies' heads.
• A separate 2023 benchmark called BabyLM found that AI can match a 10-year-old's language exposure using tens of millions of words, far fewer than the trillions most models consume.
• In 2024, a basic vision language model learned to recognise simple objects like a ball using footage from just one infant's head-camera.
• Stanford researchers published findings in early 2025 showing a new model design learned physical cause-and-effect from the same baby-head video data far more effectively than standard AI.

A one-year-old spots a dog, hears the word once or twice, and remembers it. A leading AI system ingests billions of written sentences and millions of images before it can do the same thing reliably. That gap is the puzzle driving a growing corner of AI research.

The EgoBabyVLM Challenge, developed by researchers at Meta, Stanford, Tokyo, and Paris's École Normale Supérieure, gives vision language models, AI systems trained on both images and written text, about a thousand hours of video filmed from cameras strapped to the heads of real infants and toddlers. Every top model tested so far has struggled badly.

The footage is the reason. Baby-cam video is messy. A parent talks about a toy that has already left the frame. An adult points at something with their eyes, not their finger. Conversations jump between past and future events rather than whatever is happening right now. Babies absorb all of this through sight and touch and sound simultaneously, while current AI mostly learns from tidy, curated datasets. The gap shows.

"It's clear that there's more than just language that's needed," says Michael Frank, a cognitive scientist at Stanford who studies language learning and helped design the challenge.

Why does this matter for ordinary people?

Smaller, more efficient AI would cost less to run and consume less electricity, meaning lower prices and a lighter environmental footprint for the products that rely on it. AI-powered robots that learn the way babies learn could navigate homes and hospitals far more reliably than today's machines.

The work builds on earlier research. Our 13 July report on world models and physical reasoning is a useful companion here: the drive to make AI understand cause-and-effect, not just language, is now arriving in infant-cognition labs. BabyLM, a benchmark introduced in 2023 by linguist Ryan Cotterell of ETH Zurich, showed that transformer models, systems that understand language by tracking relationships between words across a piece of text, can learn the rules of grammar from roughly the same amount of text a 10-year-old has encountered. Trillions of words are apparently not required for that particular task.

Physical common sense is a different story. "There isn't going to be a large corpus of human interactions," Cotterell says, meaning no giant dataset of real-world experience exists to train on the way written text does.

Joshua Tenenbaum, a cognitive scientist at MIT, is direct. Transformer models find patterns in data extremely well. Pattern-finding alone does not seem to be enough to build the instinctive understanding of physics and social dynamics that a two-year-old already has.

Should we expect a baby-brained AI anytime soon?

Probably not soon. The open question is how much of that understanding is built into human brains by evolution, and how much any learning system could acquire on its own. "There is a lot of debate in cognitive science and neuroscience about how much is built into the brain evolutionarily," Tenenbaum says. The EgoBabyVLM researchers argue that borrowing ideas from brain science, such as helping models track longer stretches of time and read social cues, could point toward an answer. What to watch: whether the new Stanford model architecture that learned object dynamics from baby-head video in early 2025 can be scaled without losing that efficiency advantage.