Princeton researchers train an AI model to learn language the way children do

Princeton Neuroscience Institute

10 min read Original article ↗

Nov. 5, 2026

Every parent has watched it happen, and no scientist can fully explain how.

A baby enters the world unable to understand a single word. Within a few years, that child recognizes thousands of words, follows complex sentences, and communicates their own thoughts.

What makes the achievement remarkable is how children do it. They do not study grammar, and unlike today's large language models, they do not consume much of the written internet. They learn from a small world: the voices of parents and siblings, conversations over breakfast, play on the living-room floor, a story before bed. Their world is local, social and embodied, and often noisy and chaotic.

From that small world, children extract more than any artificial system can from vastly more data. And they never stop. Today's AI models are trained once and then frozen before release; a child keeps learning every day, folding each new experience into what came before. Understanding how a mind does that is a question about human intelligence, not artificial intelligence, and it has proved surprisingly hard to answer.

For decades, scientists have debated how children accomplish this feat. Some theories emphasize knowledge that humans are born with; others emphasize the developing brain's ability to discover structure in experience. One obstacle has made these ideas hard to test: no one has ever had a near-complete record of what a child actually hears while learning to talk.

A team at Princeton University set out to change that. In the 1,000 Days Project, cameras and microphones in 15 homes recorded children's daily lives from birth through roughly age three, capturing most of each waking day and creating an exceptionally dense record of the sights, sounds and interactions surrounding each child.

Now, in a study published in the journal Science, the researchers have put those recordings to a new use. They built artificial learners and asked a deceptively simple question: if you gave a learner the speech that surrounded one child, day after day, could it begin to learn language the way that child did?

The work was led by Uri Hasson, a professor of psychology and neuroscience at Princeton and the principal investigator of the project; Dr. Hadas Raviv, a computational cognitive scientist, led the analysis infrastructure development and computational modeling; Dr. Liat Hasenfratz led the research operations, developed the project’s behavioral assessment battery in collaboration with Casey Lew-Williams, and designed its ethical and governance framework; Casey Lew-Williams, a professor of psychology, director of the Princeton Baby Lab, and co-principal investigator of the project, provided developmental-science leadership and guided the behavioral assessment strategy. 

Recording the first 1,000 days of life, and then building a learner that can learn from it, is not a job for a single laboratory or a single discipline. It took engineers who developed recording systems running in family homes for years, staff who visited those homes, annotators who turned an enormous archive of audio into usable data, and machine-learning researchers who built the learning models. "Tens of people with different professions and expertise worked on this project over the years to make it happen," Hasson said. "But none of it would exist without the families. They opened their homes and their daily lives to us for three years, and we are deeply grateful to them." "The challenges of this project were immense,” Lew-Williams said. “It took us years to rethink the basics of ethical considerations, data collection, data storage, and analysis techniques. In the end, this dataset will help us rethink children’s development and learning.”

Giving an AI a childhood

The study focused on eight children from the larger project, each identified by a pseudonym. For each child, the researchers built a separate language-learning model and fed it that child's everyday speech input, drawn from hundreds of days of recordings.

Each model began knowing nothing: no speech sounds, no dictionary, no sense of where one word ended and the next began. It was never shown a text. It had to discover the structure of language from the sound of speech itself, as children do.

Crucially, the model lived through the child's speech as it unfolded over time. It did not receive three years of recordings shuffled into one giant training set. It listened to a child's world in order, day one, then day two, then day three, and as the child's waking hours lengthened, the model's daily dose of speech grew with them.

"We wanted to turn the usual logic of developmental research around," Hasson said. "Instead of observing a child for an hour in the laboratory and trying to infer the learning process, we asked whether we could build a model capable of learning and immerse it in the child's actual experience, day after day."

The team started with one child, Coral.

“We were surprised to see that, despite the immense amount of speech available from a single child, the learning results were discouraging,” says Raviv. “Simply exposing Coral’s model to the rich stream of daily speech was not enough.”  The model retained almost none of what it heard from one day to the next. By the equivalent of Coral's second birthday, it had learned no words at all.

Something was missing.

The missing ingredient: replay

The researchers looked to sleep. A great deal of evidence suggests that during sleep the brain revisits recent experiences, reactivating patterns of activity from the day, a process thought to help new memories settle and stick. Replay lets the brain get more out of the same experience without adding anything new: by returning to what it has already heard, it gets a second chance to notice patterns it missed the first time. Sleep occupies a particularly large share of infants’ lives, and it has been linked to their word learning.

So the team gave their artificial learner something similar. At the end of each simulated day, the model revisited a sample of what it had already heard, mostly from the past few days. Machine-learning systems routinely go back over their training data many times, but they do it all at once, shuffling years of experience together. Here the model could only look back the way a child can: a little at a time, at what happened recently. The team was not trying to reproduce sleep in all its complexity, with its cycles and stages. They borrowed one idea from it, a nightly chance to revisit the day, and asked whether that alone would help the learner consolidate what it had experienced.

The effect was stark. With replay, learning took off.

"Daily experience alone wasn't sufficient," Raviv said. “Once we allowed the model to replay a limited amount of its recent experience, learning changed dramatically. It could hold on to what it had learned while continuing to learn from each new day,” Raviv adds. “It was striking to see the connection between a cognitively plausible mechanism like replay and a familiar idea from machine learning: that revisiting the same data can deepen learning.”

Coral's model first carved the stream of speech into units that matched nearly all English speech sounds (phonemes). Then it began to pick out recurring words. By the equivalent of Coral's second birthday, it had learned thousands of words from everyday speech.

Eight homes, eight different childhoods

The researchers next asked whether the same learner could learn language from the experiences of children from the seven other homes.

The households differed substantially. Some children heard more speech than others; families differed in what they talked about, who spoke to the child, and how conversations unfolded. Yet using the same settings as Coral, the learner succeeded in every household. Each model discovered the sounds of English and built a vocabulary of thousands of words, 

"If a child had been born into another one of these homes, we would still expect that child to learn language," Hasson said. "A plausible learning mechanism should be robust to those differences too".

But then a puzzle emerged. The children developed at very different rates: some parents reported a large vocabulary well before the first birthday while others reported the same milestone nearly a year later. This kind of variation is a normal part of development, yet Coral’s model failed to capture it. When the same learner was placed in each of the eight homes, it progressed at roughly the same pace every time.

"The richness of these recordings allowed us to see both sides of development," Raviv said. "Every child grows up in a distinctive linguistic environment, and every child follows an individual developmental trajectory. The challenge was to capture both at the same time."

So the researchers changed the learner rather than the environment. They let just three things vary from one child's model to the next: how much it replayed each night, how quickly it learned from what it heard, and when it began to learn words.

That was enough to make a difference. Each model could reproduce the timing of the child it represented. "We’ve known for a long time that children’s developmental environments are complex,” said Lew-Willaims. “But now we can actually see this everyday complexity, from language to vision to social interactions to screentime. It’s a mess, and we hope to understand how learning happens through this mess.”

A new way to study development

The researchers are clear that their model is not a child. It has no body. It does not crawl across a room, point to objects, watch a parent's face, or use language to get someone's attention. It has none of the social motivations or biological machinery of a developing brain, and it does not yet know what any word means.

But those gaps are the point. The approach reflects an idea famously captured by the physicist Richard Feynman: "What I cannot create, I do not understand." A theory of how children learn language can now be tested by building it, training it on a real child's experiences, and seeing whether it learns the way the child did. Where it falls short, the shortfall shows what is missing.

"For more than a century, psychology has largely studied development from the outside," Hasson said. "We observe behavior and try to infer the mechanisms that produced it. Now we can take a complementary approach: build a learning system, give it the child's actual experience, and ask whether it develops the way the child does."

"Unlike much of modern AI, which aims at superhuman performance, we are trying to understand human learning and how it gives rise to human intelligence," Raviv said. "The framework is general. Researchers can design learning systems, expose them to the experiences individual children actually have, and compare the results with the children themselves".

"The current model is only the beginning," Hasson said. "These recordings contain an extraordinary amount of information about how children see, hear, interact, and act in the world. Our goal is to progressively give the models more of those capacities and see what emerges."

Authorship & Support

Hadas Raviv, Arkadii Tsyhanov, Kira Gousios, Aja Altenhof, Haocheng Wang, Berlin Chen, Ofri Raviv, Tal Rosenwein, Casey Lew-Williams, Liat Hasenfratz, Uri Hasson. The first 1000 days: Modeling language learning from individual children’s everyday speech. Science. November 05, 2026. DOI: 10.1126/science.aeh9311

This work was supported by Wellcome Leap; the McGregor-Girand Charitable Endowment; Princeton Language and Intelligence; Princeton Precision Health; Data-Driven Social Science and the AI Lab at Princeton University; and a C.V. Starr Fellowship from the Princeton Neuroscience Institute.