Health Science Reviews

By Amy Stubbs

When we’re chatting to our friends over a cup of coffee, one doesn’t stop to ponder the
complexity of events that is occurring in the brain as we speak, pause to listen and formulate
responses. However the neural mechanisms behind this are highly complex and are in fact not
very well understood.

The question:
Artificial intelligence. A concept that has changed the world as we know it, and continues to do
so. Could it be used to help us decode the neural mechanics of human conversation?

Researchers aim to discover what is going on inside the brain during speech production,
language comprehension and the transition between speaking and listening during natural
conversations. They draw in deep learning models to test if they can help us understand these
complex neural mechanisms.

The how:
Deep learning models like GPT, don’t rely on the grammar rules like we learned in school.
Instead, they learn language by analysing huge amounts of text and picking up patterns
statistically, kind of like how your brain learns to speak just by hearing people talk.

These models have the ability to turn language into complex vectors which are essentially
numbers that capture meaning, grammar, and context all at once.

Researchers recorded electrical activity from the brains of people engaging in a free flowing
conversation, using a method called sEEG (stereo-electroencephalography). Meanwhile, they
fed the same speech into GPT-2 and captured the model’s internal “neural” activations, called
embeddings, at different layers. Then came the big test: were the patterns from the human brain
and the AI model correlated?

The discovery:
After locating the neural signals, and comparing the vectors from the NLP model with neural
patterns, researchers found that there is shared brain activity in speaking and listening, brain
activity changes as conversations switches between people, dynamic brain organization
supports real conversations and that the model’s representations correlate with neural activity,
suggesting alignment in how both systems structure meaning.

References:
Cai, J., Hadjinicolaou, A.E., Paulk, A.C. et al. Natural language processing models reveal neural
dynamics of human conversation. Nat Commun 16, 3376 (2025). https://doi.org/10.1038/s41467-025-58620-w