They're two years old, they point out the window and say "Gog!". You correct them without thinking: "Yes, a dog. Did you see how big it is?". They repeat it, laugh, point again. A few months later they no longer say "gog", and one day, without anyone having explained it to them, they come out with a whole sentence, subject, verb and even a plural in the right place.
It seems like the most normal thing in the world. And yet it's one of the things that most puzzles the people building artificial intelligence today. Because that child has learned to talk with a ridiculously small amount of data compared to what a machine needs to do something similar.
Twenty Meters of Paper Versus a Tower to Space
Let's start with the numbers, because they're what sparked researchers' curiosity. According to a recent MIT Technology Review feature, young children usually start building grammatically correct sentences after hearing about 10 million words, 30 million at the high end. A preteen who has grown up in a talkative home might be around 100 million.
Now, the machine. A model like Meta's Llama 3.1 was trained on about 15 trillion tokens (the word fragments these systems work with), and the most advanced models are estimated to use around ten times more. The feature's visual comparison makes it clear: if we printed a preteen's 100 million words, the stack would be about 20 meters tall. The stack for a current language model would shoot past the International Space Station.
And it's not that the child gets by because what they learn is a small thing. Michael C. Frank, a psychologist at Stanford, puts it with humor: if you train a model like GPT-2 on 30 million words, "you get a nonsense generator; you don't get a kid." Ethan Wilcox, from Georgetown, puts it the other way around: a model like Claude has seen the amount of language an entire city experiences in a generation.
Your child isn't a small language model: it's a different way of learning.
Training an AI as if It Were a Baby
That difference has given rise to a lovely line of research: trying to get machines to learn from the data a child would have, to understand what they're missing.
The best-known initiative is BabyLM, an annual competition proposed by researcher Alex Warstadt in 2022. The rule is simple and demanding: train language models on about 100 million words, a "plausible" amount for a child, or on just 10 million in the category for the youngest. The texts resemble what surrounds a kid: stories, dialogues, subtitles, child-directed conversation. And there are striking results: the 2024 winner, a model called GPT-BERT, beat Llama 2 70B, which had been trained on some 15,000 times more text, on one of the competition's tests.
Others have gone even further and put cameras on real babies' heads. The SAYCam project recorded three babies for about two hours a week, from six months old to two and a half years. Using 61 hours of those recordings from a single child, Brenden Lake's team at New York University trained a model that learned to associate words with objects from what that child saw and heard, and even to recognize new objects. "It turns out you can get a real start on language learning using a lot less than what a number of theories suggested," Lake explains.
And at Princeton, Uri Hasson's team has recorded the first 1,000 days of life of 17 children, with cameras and microphones in almost every room of their homes, about 12 hours a day. For the first time, Hasson says, "we have the input": everything that actually reaches a child while they learn to talk.
Children Choose Their Own Data
If data doesn't explain everything, what does a child have that the machine doesn't? The honest answer is that we still don't fully know. But the clues researchers are working with are very revealing for any parent.
The first is that children don't learn by watching, but by doing. Psychologist Alison Gopnik, of the University of California, Berkeley, explains it this way: "Children are actively exploring, which means that they're actively choosing their own data." The kid who throws the spoon on the floor for the tenth time isn't being annoying: they're running an experiment. The one who asks "but why?" until you're exhausted is looking for exactly the piece they're missing.
The second is that they learn from people, not just from information. Children interpret what they're told differently when they know someone is teaching them: according to the work of researcher Elizabeth Bonawitz cited in the feature, they reason about what the teacher knows and about why they're being shown this and not something else.
The third is the one that contrasts most with the machine: a language model learns passively, alone, swallowing text. A child learns with their body, their eyes, their hands and, above all, with someone in front of them. Researchers suspect that models that seek out information, experiment and interact could learn better, but so far they haven't managed it: the latest edition of BabyLM allowed models that learned by conversing with other models, and they didn't beat the regular ones.
What AI struggles to imitate is exactly what happens in a real conversation.
Talking With Your Child, Not Just to Them
There's one study worth keeping at hand, because it brings all of the above down to something you can do this very afternoon. In 2018, an MIT team led by Rachel Romeo and John Gabrieli asked families with children aged 4 to 6 to have them wear a small recorder for two days, from the moment they got up until they went to bed. A program counted three things: the words the child said, the words said to them, and conversational turns, that is, the back-and-forth exchanges between the child and an adult.
Then they measured the children's brain activity while they listened to stories. What turned out to be most strongly linked to activity in the brain areas tied to language, and to their language skills, wasn't the number of words they had heard, but the number of conversational turns. Romeo sums it up in a sentence worth writing down: "The important thing is not just to talk to your child, but to talk with your child."
An honest caveat: it's a study that shows a relationship, not that conversation by itself causes those changes. But it fits with everything else. A child doesn't learn like a machine fed tons of text; they learn through back-and-forth.
What to Do With All This at Home
None of the above requires apps, courses or gadgets. Above all, it requires sending the ball back.
A. Look for turns, not words. It's less about talking to them a lot and more about talking with them: ask an open question, really wait for them to answer, and respond to what they said, not to what you were going to say. With the little ones, mirroring their sounds or faces counts too: that's also a turn.
B. Let them choose their data. Exploring, touching, taking things apart, asking "why?" ten times in a row. It's their way of seeking out exactly the information they're missing, and it's far more effective than any content we simply put in front of them.
C. Getting it wrong is part of the method. The "gog" that becomes "dog" only happens if someone listens, gives them the word back said correctly and lets them try again. Correcting without cutting the conversation short is exactly that.
D. Use the comparison to explain AI. For a child or a teenager, this idea can be very useful: a chatbot has read more words than an entire city would hear in a whole generation, but it has never touched a ball or asked "why?" while waiting for your answer. It talks very well, but it doesn't learn the way they do. If you want more ideas for that conversation, we gathered them in Cómo explicar la inteligencia artificial a los niños pequeños (in Spanish).
E. Don't turn it into a chore. There's no need to keep count or measure anything. Dinner, the drive to school or the evening bath are already conversations; the difference is in leaving room for them to answer.
The Library and the Kitchen Table
The biggest artificial intelligence in the world has read a library that would reach space. Your child learns with something much smaller: a question, an answer, a laugh and another question. And, for now, science hasn't managed to make the library learn the way the kitchen table does.
Maybe that's the best news in all of this. At a time when machines talk better and better, what makes a child learn to talk is still, to a large extent, in our hands. Technology can be a bridge, but conversation is the path.
How many real conversational turns have you had at home today?