Understanding the brain with AI-driven explanations and experiments
TL;DR
Microsoft Research, UC Berkeley, UCSF and Columbia introduce Generative Causal Testing, or GCT. The method turns brain-prediction models into short, testable hypotheses about what specific cortex patches respond to during language. An LLM summarizes the strongest text triggers for a voxel into labels such as food preparation or location names. Then another LLM writes new stories meant to activate exactly that brain region in an fMRI scanner.
Nauti's Take
The strong part is not that an LLM finds patterns again. The strong part is the closed loop: write a hypothesis, build a stimulus, test it in the brain.
That turns explainability into something measurable. The caveat is real: three participants is small, synthetic stories are controlled stimuli, and Microsoft frames the work in polished research-marketing language.
As a research direction, GCT is far more interesting than another dashboard full of colorful activation maps.
Briefingshow
Many AI models can predict brain responses well, but they do not give researchers a readable theory. GCT matters because it converts black-box signals into concrete hypotheses and tests them directly in the scanner. If it scales, AI becomes less of an oracle in science and more of an experiment generator.
Sources