ResearchModels

Microsoft Research: AI explains the brain through generative testing

Researchers at Microsoft Research have developed a method that translates opaque AI models into understandable hypotheses about brain activity and experimentally verifies them.

In detail

  • Generative Causal Testing (GCT) uses LLMs to write new stories designed to activate specific brain regions.
  • Subjects hear these stories in an fMRI scanner; if the explanation is correct, the target region lights up.
  • Method bridges the gap between predictive power and interpretability of AI models.
  • Research published in Nature Neuroscience, collaboration with UC Berkeley, UCSF, and Columbia University.

Why it matters

The method makes black-box AI models scientifically useful by translating them into testable hypotheses—a breakthrough for neuroscience and AI interpretability.

For you Watch this method as a template for explainability in your own AI systems: if you need to make AI predictions that require trust, similar validation steps could make your models more credible.

← All news

Summaries are generated automatically and link to the original source.