A recent study suggests that the way the human brain processes language mirrors the function of modern artificial intelligence language models, primarily by predicting the probability of word sequences. This finding emerged from research conducted by an interdisciplinary team led by Patrick Krauss from Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) and Achim Schilling from Heidelberg University.
To investigate this, the researchers integrated a natural listening environment, simulating an audiobook, with detailed, high-resolution measurements of brain activity and employed an AI language model as a comparison reference. The study observed that when a word was highly probable within its current context, the corresponding neural response was weaker during its processing. Furthermore, the data indicated increased pre-activation occurring even before the word began, lending support to the idea that the brain operates predictively.
These results strengthen fundamental assumptions in cognitive neuroscience and offer a potential explanation for the robust performance of AI language models. The researchers caution, however, that the similarity in outcomes does not imply that the brain and AI function identically; rather, it points toward shared principles of information processing. The team plans to investigate the next steps, focusing on whether these discovered principles can be successfully applied to concrete applications such as diagnostics or brain-computer interfaces.
