Stanford launches five-year project to build AI models of brain activity

Stanford University announced a five-year Digital Brain Project on November 2 to build AI foundation models that predict brain activity. The initiative plans to train on extensive non-invasive measurements of people’s brains. Future datasets and models are intended for researchers.

The immediate target is a research model of how brain activity changes with circumstances. Clinical applications are a longer-term ambition. The announcement provides no completed model’s accuracy or patient-treatment results. The cycle connects task design, measurement, training, validation and revised experiments.

The first campaign uses audiovisual scenes. Participants will direct their attention, make memory judgments and describe what they encounter. The project plans to record the same tasks sequentially with 7-tesla functional magnetic resonance imaging, or fMRI, and optically pumped magnetometer magnetoencephalography, or OPM-MEG. These are separate recording sessions using complementary measurements.

The methods capture different signals. fMRI uses blood-oxygenation changes to locate activity. Stanford’s 7-tesla scanner, introduced in 2025, offers high spatial resolution for examining cortical layers and fine structures. The equipment description names cortical layers and crossing fibers among structures it can visualize, providing detail within the brain. The Lucas Center makes four MRI scanners available to researchers within and outside Stanford. The project draws on the university’s shared imaging infrastructure.

OPM-MEG detects tiny magnetic changes generated by brain activity. Stanford’s September description of its system reports millisecond timing and millimeter-scale spatial information. Its current configuration has 144 sensors and operates inside magnetic shielding that reduces environmental interference. The resolution and predictive value of models combining these measurements remain questions for the research.

Repeated measurements and tests on unseen conditions

Co-leader Russ Poldrack outlined the first sampling plan in a Stanford podcast released on November 2: roughly 10 people, with about 100 sessions per person in each modality over a year. He described about four hours weekly, split equally between MRI and MEG. These are planned recordings rather than an already collected cohort.

Making recordings understandable to other researchers is another part of sharing. Stanford says the effort will build on OpenNeuro and the Brain Imaging Data Structure, or BIDS. BIDS documentation describes consistent organization across institutions, with support for fMRI, MEG and behavioral data. Its ecosystem includes format validators and analysis tools. Checking a dataset’s organization serves a different purpose from evaluating a brain model’s predictive accuracy.

The project’s validation plan names held-out conditions, third-party datasets and simpler baseline models. Those comparisons would test whether predictions improve on existing approaches, with further experiments guiding refinement.

Independent researchers distinguish data sufficiency from scientific explanation. In The Transmitter’s own interviews, University of Washington neuroscientist Bing Wen Brunton questioned whether foundation models and their data-scaling approach will work for brain research. She also cautioned that accurate prediction of activity would not necessarily explain its mechanisms.

Datasets will have appropriate public or controlled access, with participant protections. Poldrack said sensitive, potentially identifying material would require restricted sharing. He hopes initial data collection can begin in early 2027, without committing to a date. At this review, the project website said participant recruitment was not yet open.