GCT: new AI method turns brain-prediction black boxes into testable linguistic theories

LLMs can predict brain activity from language with high accuracy, but why they work remains opaque. Researchers from Microsoft Research, UC Berkeley, UCSF, and Columbia introduced Generative Causal Testing (GCT) to turn these black-box models into readable, testable theories. GCT first extracts short phrases that most drive a brain region's predicted response and summarizes them into an explanation like “food preparation” or “location names.” Then an LLM writes new stories engineered to activate that specific region; subjects hear them in an fMRI scanner, and if the region lights up above baseline, the explanation passes a causal test. In experiments, GCT confirmed known selectivity and distinguished three neighboring place-processing areas: retrosplenial cortex (RSC) responds stronger to proper location names, parahippocampal place area (PPA) to general location, and occipital place area (OPA) to scene layout. It also discovered new prefrontal micro-regions tuned to dialogue (words like “said”), clock times (“one o’clock”), and numeric measurements (“50 feet”). The method proves that data-driven models can be distilled into experimentally testable theories, advancing both neuroscience and broader scientific fields.