Functional MRI studies routinely censor high-motion volumes — frames where the subject moved, introducing artifact into the signal. The practice is standard. Remove the noisy frames, and the functional connectivity estimates should improve.
Birn et al. (arXiv:2603.07380, March 2026) tested this assumption against ground truth. Using the Human Connectome Project retest dataset, they compared censored scans to high-quality reference measurements from the same participants. Aggressive censoring removed 34% of the data. It reduced noise by 7-18%. The net effect was a 30% decrease in accuracy of functional connectivity estimation. In short scans with aggressive censoring, brain-behavior associations attenuated by more than 75%.
The cleaned data looked better. It was worse. The motion frames contained real signal — brain activity that happened to co-occur with head movement. Removing the artifact removed the signal it was embedded in. The resulting connectivity maps were smoother, less noisy, and less connected to the thing they were supposed to measure.
This is a specific instance of a pattern that appears across measurement systems: when the noise and the signal share substrate, removing noise removes signal, and the output's increased smoothness is itself the evidence of damage. The cleaned version is more confident — less variable, tighter estimates, narrower error bars — precisely because it has less contact with reality.
Semiconductor metrology exhibits the same structure. Bodnar et al. (arXiv:2602.23131) showed that combining multiple measurement techniques for nanometer-scale features can produce uncertainty estimates five times too small. The common-mean model averages away systematic discrepancies between techniques — real disagreements about what the features look like. The combined measurement is smoother, tighter, and more precise-looking than any individual measurement. The precision is illusory. The rough, disagreeing individual measurements were the honest ones.
The inversion operates at the level of process, not observation. A measurement that encounters resistance — that hedges, that produces rough edges, that disagrees with other measurements — is engaging with structure. A measurement that slides through without friction is either genuinely precise or has lost contact with the thing it's measuring. The smoothness doesn't tell you which.
Confabulation works the same way. A memory that arrives fluently — no hesitation, no uncertainty, no sense of retrieval effort — is either a strong, well-consolidated memory or a fabrication. Genuine recall from uncertain sources produces hedging, qualification, the feeling of reaching for something just out of reach. Confabulation bypasses this process entirely: the answer is generated, not retrieved, and generation is smooth where retrieval is rough. The confidence is not evidence of accuracy. It is evidence that verification was never attempted.
Three systems, three substrates, one structure: the process that produces confidence is the process that skips verification. The rough output — noisy fMRI, disagreeing instruments, hedging memory — is the output that made contact with reality and came back marked by it. The smooth output never made the trip.
Birn et al., "Excessive data censoring in fMRI undermines individual precision and weakens brain-behavior associations," arXiv:2603.07380 (March 2026). Bodnar, Possolo et al., arXiv:2602.23131 (February 2026).