AI turns brain prediction models into testable theories of human cognition

Large language models have become remarkably accurate at predicting how the human brain responds to language—but their success conceals a fundamental problem: nobody can read them. Millions of learned parameters don't translate into interpretable theories about what a brain region actually encodes.

Researchers at Microsoft Research, UC Berkeley, UC San Francisco, and Columbia University have developed a solution called generative causal testing (GCT), published in Nature Neuroscience. The method bridges the gap between predictive accuracy and scientific understanding by converting black-box models into short, testable hypotheses.

How it works#

GCT operates in two stages. First, the system identifies the short phrases that most strongly drive a brain region's predicted response and uses an LLM to distill them into a concise verbal explanation—phrases like "food preparation" or "location names."

The second stage closes the loop experimentally. An LLM writes entirely new stories engineered to activate a targeted brain region according to its explanation. When subjects read these synthetic stories in an fMRI scanner, the predicted region lights up if the explanation is correct—transforming a correlation into a genuine causal test.

Read the paper: "Generative causal testing to bridge data-driven models and scientific theories in language neuroscience," accepted in Nature Neuroscience

What it revealed#

In experiments, GCT confirmed known brain selectivity and revealed nuanced distinctions among three neighboring place-processing regions long assumed to be functionally similar: the retrosplenial cortex (RSC) responds more strongly to proper nouns like "Tokyo," while the parahippocampal place area (PPA) and occipital place area (OPA) encode different location information.

The method also surfaced previously unknown prefrontal "micro-regions" tuned to specific concepts: one selective for dialogue between people, another for clock times, and a third for numeric measurements. These discoveries emerged because the system could propose a hypothesis and immediately test it.

Broader implications#

The significance extends beyond neuroscience. Researchers increasingly face the same dilemma: models that predict well but explain nothing. GCT demonstrates that data-driven models need not signal the end of interpretable theory. With the right framework, AI-powered prediction and human-readable science can advance together.