Training Data Governance for Brain Foundation Models

Neural data requires stronger protections than text or images: Brain data (EEG, fMRI, etc.) carries "stronger expectations of, and claims to, protection" because it is body-derived and has historically been governed within clinical and research settings with strict oversight—a standard that foundation models now risk undermining through large-scale repurposing and cross-context data sharing.

Foundation models create new governance gaps: The shift toward foundation models makes neural data accessible to a much broader range of actors, including commercial developers, yet governance remains "fragmented and unclear." This mismatch between expanding access and inadequate oversight creates significant risks that current regulatory frameworks were not designed to address.

Consent and data repurposing are fundamentally misaligned: Foundation models subject neural data to "large-scale repurposing, cross-context stitching, and open-ended downstream application"—practices that conflict with traditional informed consent models where participants agreed to specific, limited uses of their brain data in clinical or research contexts.

Multiple ethical concerns require simultaneous attention: The paper identifies five interconnected areas of concern—privacy, consent, bias, benefit sharing, and governance—suggesting that addressing brain foundation models responsibly requires coordinated action across all these dimensions, not piecemeal solutions.

Baseline safeguards are needed now: The authors propose both agenda-setting questions and "baseline safeguards as the field matures," indicating that proactive governance measures should be established early rather than waiting for problems to emerge at scale.