Turning Brain-Prediction Models into Testable Explanations with Generative Causal Testing

Microsoft Research, UC Berkeley, UCSF, and Columbia University present generative causal testing (GCT), a method for converting opaque language-based brain-prediction models into short verbal hypotheses such as “food preparation” or “location names.” An LLM then writes stories intended to activate a selected cortical area, and researchers test the prediction while participants listen in a brain scanner. According to the supplied Microsoft Research summary, GCT reproduced known selectivity, distinguished neighboring place-processing regions, and identified small prefrontal regions associated with concepts including dialogue, clock times, and measurements.