Israeli Study Finds Striking Similarity Between Human Brain and AI Language Processing
A new Israeli study reveals that the sequence by which the human brain processes spoken language closely mirrors the layered processing performed by large AI language models. Published in Nature Communications, the research shows that both systems progress through similar stages, from raw acoustic signals to deep semantic understanding.
Led by Ariel Goldstein of the Hebrew University of Jerusalem, the study analyzed neural activity from nine epilepsy patients implanted with electrodes for clinical reasons. While listening to a 30-minute podcast, their brain responses were recorded with high precision. Researchers then compared the timing and location of brain activity to the internal workings of a large AI language model processing the same words layer by layer. The early layers of the AI model, which handle basic linguistic features, corresponded to early brain responses, while deeper AI layers related to context and meaning matched later brain activity in regions like Broca’s area, known for language processing.
Despite the remarkable alignment, the scientists caution that this does not prove the brain operates like a computer or runs the same algorithms as AI models. The brain and AI are fundamentally different systems, one shaped by evolution and lifelong experience, the other trained on massive text datasets using statistical methods. The shared "solution shape", a layered progression from sound to meaning, likely arises because language comprehension inherently requires hierarchical processing. Thus, two distinct systems independently converged on a similar multi-stage approach.
This finding is significant as it provides neuroscientists with a concrete model to study how language understanding develops in the brain. It also challenges older theories that viewed comprehension as rule-based grammar application, instead supporting a gradual, statistical construction of meaning. As AI models increasingly predict brain activity, they become valuable tools for brain research regardless of whether they perfectly replicate brain function.
