Generative Causal Testing (GCT) distills black box neural network predictions of language cortex into short, testable claims like "food preparation" or "location names," then checks them in functional MRI (brain scan) scanners.
The pattern is familiar: an artificial-intelligence model watches a few hundred hours of people listening to a podcast in a brain scanner, learns to predict which patch of cortex will light up for any given word, and a year later it scores higher than the previous state of the art. The inconvenient truth is that the model still cannot tell you why a particular patch of cortex responds to "cinnamon" but not to "cardamom." It is a black box that gets the right answer without holding a theory in any form a scientist could read.
A collaboration between Microsoft Research, UC Berkeley, UCSF, and Columbia is now turning that black box into something the field can argue with. The method, Generative Causal Testing (GCT), pulls short verbal explanations out of a brain-predicting model, then writes fresh stories designed to light up the targeted region, and finally checks, in a real scanner, whether the predicted region actually responds. If the explanation is wrong, the region stays quiet. If it is right, the model has earned something close to a falsifiable scientific claim about what that patch of cortex is doing.
The team frames the work as closing a long-standing gap in computational neuroscience. Predictive models have gotten very good at matching brain activity, but they are not the same thing as theories. A theory predicts things the modeler did not set out to test. GCT tries to force that step into the loop, by having a large language model write the prediction in plain English, then having the experiment check it.
The reported findings, detailed in the arXiv preprint 2410.00812, split into three moves. First, GCT confirmed cortical selectivity that neuroscience already knew about, with the model generating verbal explanations that matched the conventional understanding of which regions handle which kinds of content. The interesting result is not the confirmation. It is what the method did next.
Second, it disambiguated neighboring regions in the parietal cortex that earlier work had lumped together as a single place-related area. The verbal explanations GCT pulled out separated two adjacent patches with different response profiles, and the fMRI follow-up, with subjects hearing freshly generated stories, separated the two in real brain data.
Third, and more speculative, the method surfaced tiny prefrontal "micro-regions" tuned to surprisingly narrow concepts: dialogue, clock times, measurements. These are claims the model made, then the experiment tested, not claims the experimenters set out to find. That distinction matters. The arXiv paper itself flags the result as a preprint-stage finding by the original author group, and replication is the obvious next step.
GCT also borrows from a broader framework: Singh, Antonello, and colleagues' work on evaluating scientific theories as predictive models, which treats a theory as a function you can score. The neuroscience application turns that scoring into an experimental design.
The reproducibility handle is unusually strong for an AI-for-science claim. The full pipeline, including the prompting templates and analysis code, lives at github.com/microsoft/automated-brain-explanations. Anyone with a similar brain-predicting model and access to a scanner can rerun the verbal-explanation step and design their own fMRI tests against the resulting claims.
The most useful frame is methodological, not commercial. The wire-style version of this story is "AI helps brain research." The closed-loop version is more specific: AI generated a candidate theory, the scanner falsified or supported it, and the field gets a new way to argue with its own models. The interesting question is no longer whether a black box can predict a brain. It is whether the field can build a scientific method around asking the black box, in plain English, what it thinks it is doing, and then checking.