Bigger AI Models Better Predict How the Human Brain Processes Language

Summary

Artificial intelligence is getting better at mimicking human language, and scientists are now using these tools to understand the brain itself. In a new study, researchers looked at large language models, the technology behind modern AI chatbots, to see how well they reflect actual human brain activity. They focused on "parameters," which act as the artificial connections within these systems. By comparing smaller models with millions of parameters to massive ones with billions, they wanted to see which systems best mirror our own thought processes.

To do this, the team played a 30-minute audio story to epilepsy patients. These patients already had sensors placed directly on their brains for medical reasons, a technique called electrocorticography. This allowed the scientists to record real-time electrical brain signals as the patients listened. They found that the massive AI models with billions of parameters were much better at predicting the human brain's natural responses than the smaller models. To make sure size was the true factor, the researchers cleverly tested models trained on the exact same dataset, isolating the effect of simply being larger.

The researchers also peeked inside the AI's hidden processing steps, known as layers. They discovered an interesting mathematical relationship: as AI models get bigger, the point where they best match human brain activity shifts to relatively earlier layers. Furthermore, different regions of the human brain synced up best with different layers of the AI. This maps beautifully onto the brain's own organized hierarchy for processing language, from basic sounds to complex meanings.

Why it matters.

This research bridges the gap between computer science and biology. By proving that larger artificial networks process text in ways that closely resemble human neural activity, scientists can increasingly use these virtual models as tools to study the brain. It suggests that scaling up AI does not just make it smarter, but actually makes its internal workings more similar to human biology. This foundational knowledge could eventually help engineers design better brain-computer interfaces or assist doctors in understanding how language is processed in the brain.

However, it is important to note that this is early-stage computational neuroscience. The study relies on a highly specific sample of epilepsy patients who required invasive brain monitors, and it is based on a single 30-minute audio test. While these massive AI models show incredible promise in predicting neural activity, researchers will need more generalized clinical data before these findings can be applied to the wider human population.

From the source

The story in numbers.

Duration of naturalistic audio story stimulus: 30-minute
Parameter scale of top-performing models: Billions

Keep the context

Follow the evidence.

This source is a preprint and has not completed peer review. Its findings are provisional.

Scientists Map Fetal Brain Connections in Largest Study to Date

Before a baby is even born, their brain is busy wiring itself for the outside world. To understand exactly how this happens, researchers have conducted the largest study to date mapping the structural connections of the fetal brain. Using high-quality diffusion MRI, a specialized scanning technique that safely maps nerve pathways, scientists analyzed 198 fetuses between 22 and 37 weeks of gestation. By tracing these pathways, the team constructed individual "connectomes", or comprehensive maps of the brain's internal wiring. As they tracked these developmental trajectories over time, the researchers noticed distinct patterns. The brain networks showed consistent increases in both integration and segregation as the pregnancy progressed. In simple terms, integration means different regions of the brain become better at communicating with one another, while segregation means specific areas develop their own specialized local networks. To make sense of all this data, the research team created a novel mathematical method to build age-specific templates of the growing brain. This new approach proved far superior to older methods of simply averaging images together. In fact, these custom templates are so accurate that scientists can use them to successfully predict the exact gestational age of an individual fetus, with an average margin of error of less than a single week.

bioRxiv2 min read

Charting Male and Female Brain Differences From Womb to Old Age

Scientists have long debated how male and female brains differ. Now, researchers have built the largest roadmap of the human brain to date to find out. By analyzing magnetic resonance imaging (MRI) scans, which use strong magnets to take detailed pictures of internal body structures, the team examined over 100,000 individuals. This massive dataset includes people from before birth all the way up to 99 years old, with females making up just over half of the group. The researchers tracked 241 different brain structures and found significant sex differences in nearly all of them. Crucially, they proved these differences are not simply because men tend to have larger heads overall. Even after mathematically correcting for total brain volume, the structural differences remained. The study also revealed that male brains show more variation from person to person compared to female brains. Furthermore, as people grow older, the developmental paths of male and female brains actually drift further apart.

bioRxiv1 min read

Brain scans reveal two distinct connectivity patterns in autism

Autism is known for being incredibly diverse. Because the condition looks so different from person to person, scientists have long struggled to connect specific genetic risks to how the brain actually functions. To solve this puzzle, researchers recently designed a cross-species study to hunt for hidden biological similarities between mice and humans. The research team used functional magnetic resonance imaging, commonly known as fMRI, which is a type of scan that maps brain activity by tracking blood flow. They looked at 20 different mouse models carrying genetic risks for autism. By mapping these brain networks, the scientists discovered two primary patterns of atypical communication between brain regions. Crucially, each of these two brain wiring signatures was tied to a different underlying molecular process. After finding these signatures in mice, the researchers looked for them in autistic humans and successfully identified the same biological patterns. This establishes a foundational method to categorize autism based on real, measurable biology rather than just outward behavior. However, it is important to note that this is early-stage translational research. The initial findings do not disclose the sample size, demographics, or statistical strength of the human group used to validate the results, meaning more extensive testing will be required before this reaches clinics.

Nature Neuroscience2 min read

Follow your curiosity.

Bigger AI Models Better Predict How the Human Brain Processes Language

Summary

Why it matters.

The story in numbers.

Follow the evidence.

Your next discovery.

Scientists Map Fetal Brain Connections in Largest Study to Date

Charting Male and Female Brain Differences From Womb to Old Age

Brain scans reveal two distinct connectivity patterns in autism

Summary

Why it matters.

The story in numbers.

Follow the evidence.

Your next discovery.

Scientists Map Fetal Brain Connections in Largest Study to Date

Charting Male and Female Brain Differences From Womb to Old Age

Brain scans reveal two distinct connectivity patterns in autism

A little science goes a long way.