New Research Quantifies the Value of Brain Data for Training AI Models

A new theoretical paper from researchers at the intersection of neuroscience and AI (Lewis et al., arXiv 2026) tackles a fundamental question: How valuable is brain data for training machine learning models? Using a linear Gaussian model of task targets and neural recordings, they derive precise scaling laws that show how model performance improves with the number of brain samples and task labels. The key innovation is a formal exchange rate: one brain sample can be worth anywhere from negligible to dozens of task labels, depending on factors like task-brain alignment, neural noise, latent dimensionality, and the amount of brain data already collected. This gives practitioners a quantitative basis for deciding whether to invest in neuroimaging experiments.

The analysis also extends to test distribution shift, where brain-regularized learning can yield substantial robustness gains by forcing models to learn invariances that align with human perception. Under a fixed collection budget, the authors characterize the precise regimes where brain data is cost-effective—typically when task labels are expensive and the brain signals are well-aligned with the task. This work provides a rigorous theoretical framework for NeuroAI, moving beyond anecdotal results to a principled understanding of when and why neural data helps. For fields like brain-computer interfaces and medical AI, these insights could guide resource allocation and experimental design.