Understanding the brain with AI-driven explanations and experiments
TL;DR
Microsoft Research introduces Generative Causal Testing: an LLM turns black-box brain models into short hypotheses such as „food preparation” or „location names”. The loop is then tested in scanners: an LLM writes stories designed to activate a target brain region, and three participants hear them during fMRI scans. GCT confirmed known language selectivity, separated nearby place-processing regions such as RSC, PPA and OPA, and surfaced tiny prefrontal micro-regions.
Nauti's Take
Ignore the Microsoft gloss and look at the role shift: the model moves from explainer to experiment designer. For AI builders, that is the useful pattern: systems that form hypotheses, generate stimuli, and collide with real measurements beat yet another benchmark champion.
Briefingshow
Many AI models can predict brain activity surprisingly well, but they do not explain what a region is actually responding to. GCT tries to close that gap by turning model patterns into testable verbal hypotheses. The important part is the closed loop: generate a hypothesis, create a stimulus, test it in the scanner.
Sources