AI Models Can Now Track How Language Develops in Children's Brains From Age 2

Cognitive Neuroscience Society

Ask where language lives in the brain, and most people will point to two spots: Broca's area for production and Wernicke's area for comprehension. That two-region model has dominated neuroscience textbooks for over a century. It is also, according to researchers presenting new work at the Cognitive Neuroscience Society's 2026 annual meeting in Vancouver, fundamentally incomplete.

From locations to systems

"Language is not a single 'thing' in the brain. It is a system," said Stephanie Forkel of Radboud University Nijmegen, one of several researchers presenting at a symposium chaired by Tamara Swaab of UC Davis and the University of Birmingham. The symposium brought together three distinct research approaches, AI-based modeling, white matter brain imaging, and large-scale genetics, all converging on the same conclusion: understanding language requires looking at connections and processes, not just locations.

Swaab framed the shift: "We still tend to study language one level at a time, genes, brain pathways, neural activity, behavior, computation, without fully connecting those levels into a coherent mechanistic account. Now, however, we can study those connections at multiple levels and in far more detail."

AI models decode language in toddlers' brains

Jean-Remi King at Meta tackled a question at the intersection of AI and human development: how do children acquire language so efficiently, with vastly less exposure to words than today's large language models, while other species cannot?

Working with the Rothschild Foundation Hospital's pediatric epileptology unit, King and colleagues analyzed neural activity recorded from more than 7,400 electrodes implanted in the brains of 46 children, teenagers, and adults with intractable epilepsy. These patients were temporarily implanted with stereotactic electrodes prior to surgery, providing a rare window into brain activity at high spatial and temporal resolution.

The researchers found that large language models could accurately predict the brain responses of these participants to an audiobook, including children as young as 2 years old. The AI models captured the neural representations of language across ages, but with an important developmental distinction: high-level language features such as grammar continued to mature between ages 2 and 10, while lower-level features like phonetic processing were more stable across ages.

"While the underlying mechanisms remain to be determined, this work offers the first compelling evidence that modern AI systems can provide powerful new insights into how language develops in the human brain," King said.

The myth of left-brained versus right-brained language

Forkel's team used ultra-high-field 7 Tesla diffusion MRI to reconstruct seven major white matter pathways involved in language in 172 individuals. The question: do people fall into distinct "left-brained" or "right-brained" categories for language processing?

The answer was no. "Instead of distinct categories, we found that language is not binary in the brain; it forms a continuum," Forkel said. Some individuals show stronger left-hemisphere involvement, others stronger right-hemisphere involvement, but the distribution is smooth, not clustered into two types. This challenges categorical models of hemispheric dominance that have persisted in both popular culture and scientific literature.

The practical implications extend to clinical settings. Stroke damage to the brain affects language differently depending on where the patient falls along this continuum, which helps explain why two patients with similar-looking brain lesions can have vastly different language outcomes. Forkel's team now has funding for a five-year project to understand how language emerges from its biological foundations, with the goal of both understanding language creation and developing better strategies for protecting or restoring it after injury.

Rhythm problems as a genetic window into reading disorders

Reyna Gordon of Vanderbilt University Medical Center approached language from the genomic level. Using large publicly available genetic databases, including data from over 1 million participants through 23andMe, her team identified multiple genes associated with dyslexia and, more intriguingly, found 16 separate regions of the genome that are common to both rhythm impairments and dyslexia.

This shared genetic architecture between musical rhythm and reading ability suggests a biological connection that goes all the way back to the genome. "Rhythm impairments may actually be a risk factor for language problems and reading disorders," Gordon said. The finding could eventually inform earlier screening: difficulties with rhythm might serve as an early marker for children at risk for reading problems, potentially enabling intervention before reading instruction begins.

The genomic approach benefits from datasets of unprecedented scale. "Thanks to publicly funded data resources, we have been able to start studying language genetics at scale and to start to link that to its neural basis in some really innovative ways," Gordon said.

What integrated approaches cannot yet do

Each of these research lines has significant limitations. The AI brain-decoding work relies on data from epilepsy patients, whose brains may not be fully representative of typical language development. The electrode coverage, while extensive, does not capture all brain regions equally, and the sample of 46 patients, while remarkable for this type of research, is small for drawing broad developmental conclusions.

The white matter imaging study provides structural information but cannot directly observe language processing in action. And the genetic studies identify statistical associations across large populations but cannot determine how much of any individual's language ability is genetic versus environmental.

The researchers are also candid about what remains unknown. The AI models predict brain responses accurately but do not explain how the brain produces those responses. The genetic overlaps between rhythm and reading disorders are correlational; the causal pathways remain to be mapped. And connecting all three levels, genes to brain structure to neural computation, remains aspirational rather than achieved.

Still, the convergence of approaches represents a genuine shift. "The human brain is not built from rigid blueprints, but rather from adaptable architectures," Forkel said. Understanding how those architectures create and sustain language will require precisely the kind of multi-level integration these researchers are pursuing.