State vs trait interoception

Almost everything on this wiki is written as though interoception were a trait. interoceptive-sensitivity opens by calling it “a characterological trait.” All four partitions on interoceptive-taxonomy — Farb’s seven, Garfinkel’s three, Khalsa’s eight, Berntson’s seven — name properties of a person. The heartbeat-detection-task is administered once and the score is carried around as an individual difference. Correlations are computed between one number per participant and one outcome per participant.

Wallman-Jones et al. (2023) measured self-reported interoception ten times a day for a week in 70 people and found an intra-class correlation of 0.51: only about half the variance is between people. The other half is within them, moving from prompt to prompt with what they happen to be doing.

That is not a small methodological footnote. A construct with an ICC near 0.5 measured once is measured with a systematic error the size of the signal, and every single-session estimate on this wiki inherits it.

What moves it

In the one study that has looked, the state moves with ordinary behaviour on a fifteen-minute timescale (physical-activity-and-interoception): up with movement, down with sitting, up most during exercise, down most during sleep and screen time. Nothing exotic was required — no induction, no drug, no clinical event. The construct is responsive to what a person is doing with their afternoon.

The evidence was already here

The finding reframes several items the wiki had filed separately as anomalies of the heartbeat task:

• Van der Does et al. (2000): of 17 panic patients accurate at baseline, only 8 were still accurate after treatment. Exercise manufactures accurate perceivers above ~100 bpm and they vanish by ~95. The wiki recorded this as “the trait may not be a trait” and read it mainly as a cardiodynamic confound — a fact about signal amplitude contaminating a perceptual measure. It reads equally well as a fact about state: the quantity really does move, and moving it is not obviously an artefact.
• Zoellner & Craske (1999): within a person, heartbeat-counting error fell as state anxiety rose. A within-person effect on the field’s canonical trait measure, reported before anyone was looking for one.
• Banellis et al. (2026): bodily attention correlates with worse momentary affect and lower trait symptom burden in the same 536 people. The wiki read the split as “better for what?” It is also “over how long?”
• Wallman-Jones et al. themselves: momentary activity raises self-reported interoception while habitual activity lowers it, in one model.

Four sources, four different instruments, one shape: the within-person and between-person versions of a measure need not agree, and where both have been estimated they have disagreed.

Why this cuts deeper than the taxonomy quarrel

interoceptive-taxonomy records that the field has four incompatible partitions of interoception, and Banellis et al. (2026) add that the partitions are channel-specific — cardiac sensitivity does not predict respiratory sensitivity, so the words do not name person-level properties across organs.

This page adds the orthogonal failure. Even holding the construct and the channel fixed, the words may not name person-level properties across time. Both critiques attack the same assumption from different sides, and neither taxonomy anticipated either: all four lists are written as though naming the construct precisely enough would make it stable.

The practical consequence for reading this wiki: wherever a page says “individuals with higher interoceptive accuracy,” substitute “individuals who scored higher on one occasion” and ask whether the claim survives. Sometimes it will — the age gradient is too large and too replicated to be session noise. Sometimes it plainly will not: the Dunn disagreement turns on correlations of r ≈ .08–.34 estimated from single sessions, and a measure with an ICC of 0.5 attenuates such correlations by an amount nobody in that dispute has computed.

What is missing

Everything, essentially. The one estimate is of self-reported interoception (sensibility) on an instrument built for something else, in 70 young healthy adults. There is no state estimate for accuracy — nobody has run a heartbeat or filter-detection task ten times a day for a week and partitioned the variance, though the test-retest and state-anxiety results above suggest the answer would not be reassuring. And there is no state instrument anyone designed on purpose: Wallman-Jones et al. reached for a mindfulness scale and then wrote the request for a proper one into their limitations.

Until that exists, the state/trait split is a warning about the existing literature rather than a research programme, which is the honest place to leave it.