Children are much better at learning language than AI

In raw processing power, a large language model will outclass toddlers. Yet when a team of researchers did the math, they found a startling gap: if humans learned language at the same rate as ChatGPT, it would take them 92,000 years.

The calculation highlights a long-standing mystery. Children master speech and grammar after only a few years of everyday experience, whereas state-of-the-art AI, trained on billions of words, still trips over basic nuance.

A new study conducted at the Max Planck Institute for Psycholinguistics offers the most comprehensive explanation to date.

Led by developmental psychologist Caroline Rowland and colleagues at the UK-based ESRC LuCiD Center, the research proposes that the decisive advantage is not the amount of data children receive, but the way they interact with it.

Their active, embodied, and socially embedded learning engine may also hold lessons for the next generation of artificial intelligence.

How toddlers power research

The past decade has produced a multitude of tools – head-mounted eye trackers, wearable microphones, machine-vision scene analyzers – that let scientists capture moment-by-moment snapshots of childhood.

A typical modern dataset might include everything a two-year-old sees, hears, grasps, and babbles during an ordinary afternoon.

What has been missing, Rowland and her co-authors argue, is an integrated theory of how these multisensory streams translate into grammar and vocabulary.

How senses shape speech

The framework they propose rests on several mutually reinforcing principles.

First is multisensory integration. Unlike text-bound chatbots, babies process language against a rich backdrop of sights, sounds, tastes, smells, and textures.

A spoonful of mashed banana, for instance, comes with a color, a smell, a shape, a temperature, and a spoken label from a caregiver.

Over time, these correlated cues help infants crack linguistic codes that would baffle a purely textual learner.

Moving bodies, active minds

Second is embodiment. Children’s bodies are constantly in motion – rolling, crawling, pointing, mouthing objects.

Each action changes the incoming data stream, generating fresh correlations between words and physical experiences, like associating the word cup with the feel of a plastic rim.

This kind of “closed-loop” activity allows children to test hypotheses on the fly, using movement as part of the learning process.

Social immersion plays a third critical role. Caregivers instinctively tailor their speech in real time, repeating words, exaggerating intonation, or shifting topics based on the child’s attention.

Whereas AI systems read from static datasets, children receive a dynamic, personalized curriculum designed by human minds evolved for teaching.

Curiosity powers toddlers’ learning

A fourth principle is incremental plasticity. The young brain reorganizes rapidly, strengthening and pruning neural connections in response to experience.

This adaptability allows children to shift their learning priorities as they grow- starting with mastering sounds, then words, then grammar – without needing to discard earlier knowledge.

Finally, there is motivation and curiosity. Perhaps most crucially, toddlers want to decode the world around them. They actively seek novelty, request clarification, and show visible delight when they succeed, sustaining hours of daily language practice without explicit instruction.

“AI systems process data… but children really live it,” Rowland explained. “Their learning is embodied, interactive, and deeply embedded in social and sensory contexts.”

“They seek out experiences and dynamically adapt their learning in response – exploring objects with their hands and mouths, crawling toward new and exciting toys, or pointing at objects they find interesting. That’s what enables them to master language so quickly.”

What children can teach AI

The authors believe these insights could reshape machine-learning strategy. Current large language models ingest terabytes of written text, a modality children barely encounter until school age.

To narrow the performance gap, engineers might endow robots with multisensory inputs, motor exploration, and real-time social feedback loops – essentially giving silicon learners a simulated childhood.

“AI researchers could learn a lot from babies,” Rowland said. “If we want machines to learn language as well as humans, perhaps we need to rethink how we design them – from the ground up.”

Helping language thrive again

Beyond technology, the framework could illuminate adult second-language acquisition, evolutionary linguistics, and educational practice.

For example, it suggests that immersive, interactive classrooms may outperform rote drills, and that endangered languages could be revitalized by recreating full sensory contexts for young speakers.

Simulated toddlers in labs

Rowland’s group is already testing the model with longitudinal recordings from multilingual families. Meanwhile, cognitive neuroscientists plan brain-imaging studies to chart how sensory–motor loops sculpt language circuits.

On the AI front, several labs are experimenting with embodied agents that crawl through virtual nurseries, manipulating objects while linking words to experience.

Whether those synthetic toddlers can close a 92,000-year deficit remains to be seen. What is clear is that humanity’s smallest linguists still wield secrets that even the largest neural networks have yet to crack.

The study is published in the journal Trends in Cognitive Sciences.

—–

Like what you read? Subscribe to our newsletter for engaging articles, exclusive content, and the latest updates.

Check us out on EarthSnap, a free app brought to you by Eric Ralls and Earth.com.

—–