Microsoft Research's GCT makes AI brain models readable and testable

For years, LLM-based models have predicted how the human brain responds to language with high accuracy, but their internal workings remain inscrutable—millions of learned parameters that cannot be translated into scientific theories. This gap between prediction and understanding has become a central challenge in computational neuroscience. Now, a collaboration between Microsoft Research, UC Berkeley, UCSF, and Columbia University introduces Generative Causal Testing (GCT), a framework that turns these black boxes into testable hypotheses. GCT first identifies the short phrases most strongly driving a voxel's predictive model, then uses an LLM to summarize them into a concise verbal explanation—like "food preparation" or "location names." In the second step, an LLM writes new stories deliberately crafted to activate that brain region. Subjects hear these synthetic stories in an fMRI scanner; if the targeted area lights up significantly more than with baseline text, the explanation passes a genuine causal test.

In experiments across three subjects, GCT confirmed known selectivity patterns and, more importantly, revealed insights invisible to prior methods. It teased apart neighboring place-processing regions that were long thought interchangeable, and discovered tiny prefrontal "micro-regions" tuned to specific abstract concepts like dialogue, clock times, and measurements. The approach turns uninterpretable predictive models back into the currency of science: concise, causal hypotheses that can be confirmed or refuted experimentally. This marks a significant step toward making AI-driven neuroscience more transparent and rigorous, opening the door to a deeper understanding of how language is represented in the human brain.