Understanding the brain with AI-driven explanations and experiments

2026-06-25 · Source: Microsoft Research · Field: Science & Research — Life Sciences & Biology, Research Methodology & Innovation, Artificial Intelligence & Machine Learning · Depth: Expert, medium

Summary

Generative causal testing (GCT), a framework developed by Microsoft Research in collaboration with the University of California, Berkeley, University of California, San Francisco, and Columbia University, addresses the explainability problem in language neuroscience. Published in Nature Neuroscience, GCT distills complex LLM-based brain-prediction models, which accurately forecast human brain responses to language but are unreadable, into concise verbal explanations. The method involves two steps: first, an LLM summarizes phrases strongly driving a brain region's predicted response into a short explanation like "food preparation." Second, an LLM generates new stories designed to activate that specific brain area, and subjects' fMRI responses are measured to causally verify the explanation. Experiments confirmed known selectivity, differentiated neighboring place-processing regions (RSC, PPA, OPA), and identified new prefrontal "micro-regions" tuned to concepts like dialogue, clock times, and measurements.

Key takeaway

For research scientists grappling with black-box predictive models, GCT offers a framework to transform opaque AI outputs into verifiable scientific theories. You can apply this generate-and-verify philosophy to accelerate hypothesis formation and experimental design in domains beyond neuroscience. Consider integrating LLM-driven explanation and causal testing to validate your data-driven models, ensuring your predictions lead to deeper, interpretable understanding.

Key insights

Generative causal testing (GCT) translates inscrutable AI brain models into testable scientific hypotheses through LLM-driven explanation and verification.

Principles

• Black-box models can yield testable theories.
• Causal testing validates correlational predictions.
• AI can accelerate hypothesis generation.

Method

GCT identifies phrases driving a brain region, summarizes them into an explanation using an LLM, then uses another LLM to generate stories for causal fMRI verification.

In practice

• Map cortical regions with AI-proposed hypotheses.
• Differentiate functionally similar brain areas.
• Discover new concept-specific micro-regions.

Topics

• Generative Causal Testing
• Language Neuroscience
• Explainable AI
• fMRI
• Large Language Models
• Brain Mapping

Code references

• microsoft/automated-brain-explanations

Best for: AI Scientist, Research Scientist

Related on AIssential

• New course on generative AI for behavioral science — Generative AI offers both opportunities and significant methodological challenges for social science research.
• Causal Learning with Neural Assemblies — Neural assemblies can learn causal directionality using local plasticity, offering an auditable, explainable-by-design framework.
• Causality is Key for Interpretability Claims to Generalise — Causal inference is essential for ensuring interpretability claims about LLMs are valid and generalizable.
• I built a tool that shows you what GPT-2 is "thinking" in real-time as it generates 3D graph of concept activations per token [R] — AXON visualizes LLM internal states via real-time 3D graphs of SAE feature activations, revealing pre-token generation dynamics.
• Models of the Mind: How physics, engineering, and mathematics have shaped our understanding of the brain — Mathematical models are crucial for explicit hypothesis testing and understanding the brain's complex, evolved systems.
• Enhancing Brain MRI Anomaly Detection and Reasoning with ROI Rethink and Synthetic Data — Explicit hypothesis-verification grounding with ROI marking enhances trustworthy brain MRI anomaly detection and reduces hallucination.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Microsoft Research.