I was mooching around Meta’s website which showcases recent research that it’s funded [‘We’re advancing AI for a more connected world.’] when I came across an interesting study into how the brain acquires language.
A team of researchers in Paris invited children, teens & adults to listen to The Little Prince -while over 7,000 electrodes recorded their brain activity in real time. The participants were patients undergoing epilepsy treatment, which allowed doctors to safely implant intracranial electrodes for clinical monitoring.
It was a rare opportunity. These recordings -called stereotactic EEG- give researchers a detailed, high-resolution look at language processing that simply isn’t possible with fMRI or standard EEG, especially in very young children who can’t sit still for long. So instead of lab tasks, participants listened to a story. In doing so, they helped uncover how the brain develops the capacity for language.
While the study offers valuable insight into how language develops in children, its core aim was to use that understanding to improve Large Language Models (LLMs). By comparing brain activity with the internal workings of AI systems, the researchers hoped to reverse-engineer some of the learning strategies that make human language acquisition so efficient.
What we know about how kids learn language
We’ve long known that language doesn’t all arrive at once. From about 6 months, infants start tuning in to the sounds of their native language (Kuhl et al., 1992). By 9 months, they can distinguish familiar word forms. But understanding words as meaningful units, and combining them into grammatically correct sentences, is something that develops more gradually over the next decade.
Some theories, like Friederici’s neurocognitive model (2012), suggest a bottom-up process: children first acquire lower-level perceptual features like phonemes, then gradually build the ability to process syntax & semantics as their brains mature. This new study backs that up -with direct neural evidence.
What did the researchers find?
Young brains handle sound early on. Children as young as 2–5 showed robust brain responses to phonetic features like voicing or nasality -especially in the superior temporal gyrus (STG), a key area for speech perception.
But words & meaning come later. Word-level features (e.g. frequency or part of speech) weren’t strongly represented in the youngest children. These only became consistently detectable in older participants -especially in higher-order areas like the inferior parietal lobule & lateral frontal cortex.
Language spreads as the brain matures. Over time, more brain regions get involved. Where toddlers showed focused activity, teens & adults had more distributed responses -mirroring the expanding cortical network seen in studies of executive function & literacy (Gogtay et al., 2004).
AI shows similar patterns -with more training. When the researchers ran the same audiobook through LLMs, the results got even more interesting. Untrained models barely aligned with human brain responses. But after training, they showed patterns increasingly similar to the brain -especially in deeper layers, which seemed to match higher-level linguistic features.
Just like children, in other words, the models needed input & time to develop more structured representations of language.
Teacher Takeaways?
- Language development isn’t all-or-nothing. While even toddlers are tuned into the sounds of speech, meaning and structure develop more slowly. This supports teaching approaches that revisit & recycle language patterns, especially at early stages.
- Vocabulary depth takes time. Word-level features like part of speech & semantic category are effortful for the brain to acquire. Don’t underestimate the value of storytelling, context clues & discussion-based input to support lexical growth.
- Rich input matters. AI models needed massive exposure to learn what human children grasp through everyday conversation. We can’t (& shouldn’t) replicate that scale -but it’s a reminder that regular, meaningful interaction really is the best model.
Clearly this research isn’t aimed at the classroom -but it does give us a powerful glimpse into how language development unfolds at the neural level, & how AI might one day help us model or even predict learning progress.
How do you support vocabulary development beyond the initial introduction? Do you return to words across time, or expect them to stick right away?
Leave a Reply