When neuroscientists measure brain activity with fMRI, they don't measure neural firing. They measure blood flow — the hemodynamic response that follows neural activity by several seconds, filtered through the vasculature's slow, nonlinear dynamics. When they measure with EEG, they don't measure individual neurons either. They measure the summed electrical field at the scalp, smeared across space by volume conduction through skull and cerebrospinal fluid. In both cases, the instrument sits between the observer and the phenomenon, and the instrument has its own physics.
The standard approach to inferring brain connectivity treats these measurements as noisy but faithful representations of neural activity. Correlate regions, apply Granger causality or transfer entropy, and the connections that survive statistical thresholding are the real ones. But Bae et al. (arXiv 2602.09034, January 2026) show that this assumption produces phantom connections. The measurement physics doesn't just add noise — it creates structure. Volume conduction in EEG makes distant regions appear coupled. Hemodynamic filtering in fMRI introduces temporal delays that look like causal lags. If you fit a causal model to the measured signal without accounting for the measurement process, the model captures the physics of the instrument, not the physics of the brain.
Their framework, INCAMA, inverts the measurement process before inferring causality. Rather than treating instrument distortion as error to be averaged away, they model the known physics of hemodynamic filtering and volume conduction and mathematically reverse it. The causal discovery then operates on the reconstructed neural signal rather than the raw measurement.
The test: apply the trained model to Human Connectome Project fMRI data without any domain-specific tuning. Zero-shot — no fine-tuning on that dataset, no feature engineering, no hand-crafted priors about brain anatomy. The recovered connections include V1→V2 (primary to secondary visual cortex) and M1↔S1 (motor-somatosensory reciprocal connections). These are among the most well-established connections in neuroanatomy. The method finds them from the data alone, after removing the instrument's contribution.
What makes this structurally interesting is not the specific technique but the implication: every existing brain connectivity study that didn't model the measurement physics may have reported connections between instruments rather than connections between neurons. The map wasn't wrong because the cartographer was careless. It was wrong because the map included the properties of the paper.