Summary: The human brain processes spoken language in a step-by-step sequence that closely matches the way large linguistic models transform text. Using electrocorticography recordings of people listening to a podcast, the researchers found that early brain responses aligned with early layers of the AI, while deeper layers corresponded to later neural activity in regions such as Broca’s area.
The results challenge traditional theories of language that rely on fixed rules, instead emphasizing dynamic and contextual computation. The team also published a rich dataset linking neural signals to linguistic features, providing a powerful resource for future neuroscience research.
Key facts
- Layered alignment: Early brain responses followed the first layers of the AI model, while deeper layers aligned with later neural activity.
- Context rather than rules: AI-derived contextual embeddings predicted brain activity better than traditional linguistic units.
- New resource: Researchers have published a large neurolinguistic dataset to accelerate the neuroscience of language.
Source: Hebrew University of Jerusalem
In a study published in Natural communicationsResearchers led by Dr. Ariel Goldstein of the Hebrew University in collaboration with Dr. Mariano Schain of Google Research as well as Professor Uri Hasson and Eric Ham of Princeton University, have discovered a surprising connection between the way our brains make sense of spoken language and the way advanced AI models analyze text.
Using electrocorticography recordings of participants listening to a thirty-minute podcast, the team showed that the brain processes language in a structured sequence that reflects the layered architecture of large language models such as GPT-2 and Llama 2.
What the study revealed
When we listen to someone speak, our brain processes each incoming word through a cascade of neural calculations. Goldstein’s team found that these transformations unfold over time in a pattern parallel to the hierarchical layers of AI language models.
Early layers of AI track simple word characteristics, while deeper layers incorporate context, tone, and meaning. The study found that human brain activity follows a similar progression: early neural responses align with earlier layers of the pattern, and later neural responses align with deeper layers.
This alignment was particularly clear in high-level language regions such as Broca’s region, where the peak brain response occurred later for deeper AI layers.
According to Dr. Goldstein, “What surprised us most is how well the temporal unfolding of meaning in the brain matches the sequence of transformations within large language models. Even though these systems are built very differently, the two seem to converge on a similar step-by-step build toward understanding.”
Why it matters
The results suggest that artificial intelligence is not just a tool for generating text. It could also open a new window into understanding how the human brain processes meaning. For decades, scientists believed that understanding language relied on symbolic rules and rigid linguistic hierarchies.
This study challenges this view. Instead, it supports a more dynamic and statistical approach to language, in which meaning emerges gradually through layers of contextual processing.
The researchers also found that classic linguistic features such as phonemes and morphemes did not predict real-time brain activity as well as AI-derived contextual integrations. This reinforces the idea that the brain integrates meaning in a more fluid and contextual way than previously thought.
A new reference for neuroscience
To advance the field, the team made public the full dataset of neural recordings associated with linguistic features. This new resource allows scientists around the world to test competing theories about how the brain understands natural language, paving the way for computational models that more closely resemble human cognition.
Key questions answered:
A: The brain transforms spoken language through a sequence of calculations that align with increasingly deeper layers of larger linguistic patterns.
A: It challenges rule-based theories of language, suggesting instead that meaning emerges through dynamic, contextual processing, similar to modern AI systems.
A: A publicly available dataset combining electrocorticography recordings with linguistic features, enabling new tests of competing linguistic theories.
Editorial notes:
- This article was edited by a Neuroscience News editor.
- Journal article revised in its entirety.
- Additional context added by our staff.
About this language and current AI research
Author: Yarden Mills
Source: Hebrew University of Jerusalem
Contact: Yarden Mills – Hebrew University of Jerusalem
Picture: Image is credited to Neuroscience News
Original research: Free access.
“The temporal structure of natural language processing in the human brain corresponds to a layered hierarchy of large language models» by Uri Hasson et al. Natural communications
Abstract
The temporal structure of natural language processing in the human brain corresponds to a layered hierarchy of large language models
Large language models (LLMs) provide a framework for understanding language processing in the human brain. Unlike traditional models, LLMs represent words and context via layered numerical embeddings.
Here, we demonstrate that the layer hierarchy of LLMs aligns with the temporal dynamics of language understanding in the brain.
Using electrocorticography (ECoG) data from participants listening to a 30-minute narrative, we show that deeper LLM layers correspond to later brain activity, particularly in Broca’s area and other language-related regions.
We extract contextual embeddings from GPT-2 XL and Llama-2 and use linear models to predict neuronal responses over time. Our results reveal a strong correlation between the depth of the model and the temporal window of reception of the brain during comprehension.
We also compare LLM-based predictions with symbolic approaches, highlighting the advantages of deep learning models for capturing brain dynamics.
We are releasing our aligned neural and linguistic dataset as a public benchmark for testing competing theories of language processing.