Scientists Built a Test to See If AI Can Read Any Brain
Why teaching AI to understand brains could speed up treatments for disease.
Researchers built BrainWideBench, a unified benchmark for testing whether AI models pretrained on brain recordings can transfer to animals they've never seen. It's built on the International Brain Laboratory's Brainwide Map dataset: neural and behavioral recordings from 139 mice, spanning 276 brain regions, all performing a sensory-guided decision-making task.
The benchmark has three task suites, checking whether learned representations can decode behavior, predict masked or future neural activity, and recover biologically meaningful anatomical organization. Testing pretraining methods across finetuning and zero-shot transfer to unseen animals, the team found pretraining beats matched single-session baselines. But transfer gains depended heavily on how well pretraining objectives aligned with downstream tasks. No single approach performed uniformly well across all three suites, and most methods were designed to tackle only a subset. As the authors put it, learning representations that jointly generalize across behavior, dynamics, and anatomy remains an open challenge.
- A team of 40+ researchers built a shared exam, called BrainWideBench, to compare AI models that read brain activity — using data from 139 mice across 276 brain regions.
- AI that 'studies first' (pretraining) beats AI trained from scratch on one animal, but only when the practice matches the actual job.
- No single method passed all three tests, so a true general-purpose brain model — the kind that could one day power medical devices — doesn't exist yet.
Why It Matters
Better brain-reading AI could lead to medical devices and treatments that work reliably across different people, not just one patient.