Using artificial intelligence (AI), researchers have deciphered the complex brain activity during everyday conversations, which could enhance our understanding of language neuroscience and improve technologies for speech recognition and communication. They utilized an AI model named Whisper, which transcribes audio into text by training on audio files and their corresponding transcripts. Unlike traditional models that rely on linguistic structures like phonemes and parts of speech, Whisper learns purely from statistical correlations and can predict text from new audio. Although its initial design doesn't encode linguistic features, those features emerged during training. The study, published in *Nature Human Behaviour*, involved four epilepsy patients who had electrodes implanted for clinical monitoring. Researchers recorded over 100 hours of real-life conversations, aiming to study brain activity outside controlled lab settings.

They discovered that different brain regions activated during speaking and understanding speech, supporting the idea of a distributed processing approach rather than distinct areas dedicated to specific tasks. For instance, areas for processing sound showed heightened activity during auditory input, while regions related to higher-level thinking were more active in language comprehension. The sequence of activation indicated a progression from hearing words to interpreting their meaning. By training Whisper with 80% of the audio data, researchers achieved accurate predictions of brain activity for untrained conversations, outperforming traditional models that rely on linguistic features. This groundbreaking research establishes a connection between AI language models and brain functions, suggesting that both might process language similarly, although further investigation is necessary to confirm these parallels. Comparative studies of artificial neural networks and human cognition remain vital to understanding the shared mechanisms of language processing.