Most of us take for granted that we can now chat with our phones. It feels like magic when an AI answers our questions or writes an essay in seconds. Yet, there is a deep, technical mystery hidden behind that polished surface: while machines are getting better at sounding human, they are doing so in a way that is fundamentally different from how a child learns to speak.
Scientists are investigating a huge divide between human learning and machine learning called the data efficiency gap. To reach a level of basic fluency, an artificial intelligence model—a digital engine designed to predict the next word in a sequence based on training—must process trillions of words of data. A human child, by comparison, learns to speak their native language using only a tiny fraction of that information. By the time a child turns twenty, they have heard only about 300 million words, which is insignificant compared to the gargantuan datasets used to build modern AI.
The mystery of the toddler's brain
To understand this, look at how an AI is actually built. It starts with training, which is the process of showing the computer an massive collection of text—books, websites, and articles—so it can look for patterns. The computer learns by calculating statistics; it essentially plays a game of autocomplete on a massive scale, guessing which word comes next until it gets good at predicting human speech. Because the computer doesn't have a childhood, experiences, or a physical world to inhabit, it has to compensate by reading the entire internet to learn the rules of language.
Children do the opposite. They don't just process text; they exist in the world. They learn through interaction, tone of voice, and physical context. For decades, linguists debated whether children were born with a pre-installed understanding of how language works or if they built it from scratch by listening. Modern AI proves that you can reach impressive results just by using raw statistics, but it also highlights that the AI approach is incredibly wasteful. It is like trying to teach someone how to bake by making them watch ten million videos of baking, whereas a human can learn to bake by being shown once or twice by a parent.
This gap is a massive practical problem for the tech industry. We are running out of high-quality internet text to feed these models. If we want AI to keep improving, we cannot just keep feeding it more data. Researchers hope that by decoding exactly how a toddler’s brain picks up language so quickly, they can build machines that actually understand the world rather than just guessing which words are likely to appear next. Solving this wouldn't just make better chatbots; it would force us to confront what intelligence really is—and whether our current machines are even on the right path to achieving it.
Liked this one? The next lands at breakfast.
Every story in tomorrow's AI news, rebuilt in plain English — five minutes, sources linked, free forever.
By joining you agree to receive Article's daily newsletter — unsubscribe in one click. Privacy