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.

What does exist, on the accuracy side, is manipulation rather than repeated sampling: Schulz et al. (2013) pushed one acute stressor through a pre/post cardiac-task design and watched heartbeat-counting accuracy rise and heartbeat-discrimination accuracy fall within the same session. So the accuracy score is demonstrably state-movable, not just test-retest-unstable — though it comes with the same caveat this page keeps meeting: the mover may be the heart (a stress-amplified cardiac signal) rather than the perceiver, which for a state account is arguably the point (the quantity really moves, and the source being physiological does not make the movement an artefact) but for an accuracy interpretation is fatal.

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.

One place the state and trait versions agree — and it is the accuracy-belief construct (Murphy et al. 2020)

Everything above is about the two versions disagreeing. Murphy et al. (2020) supply the wiki’s one case where they line up, which is worth recording because it narrows what the state/trait warning applies to.

Believed interoceptive accuracy was measured both ways in the same participants: as a trait (IAS, a stable questionnaire — 30-day test–retest r = .754) and as a state (confidence rated after each heartbeat-counting trial). Both correlated with objective task performance; both failed to correlate with the attention-flavoured BPQ. Two timescales, same behaviour.

That does not soften this page’s argument, and it should not be read as doing so — the quantity with the ICC of 0.51 is attention-flavoured self-report (Wallman-Jones used a state mindfulness scale), and no equivalent variance partition exists for the IAS. What it adds is a candidate refinement worth testing: the state/trait gap may itself be construct-specific, with how much attention is on the body right now being genuinely situational while how accurately I believe I read my body behaves more like a disposition. Nobody has run the ten-prompts-a-day design on an accuracy-belief scale, so this is a hypothesis the page can now state, not a finding it can carry.

One inconsistency inside the paper cuts against even that. Study 4’s trait↔state correlation was null (IAS ↔ mean confidence r = −.026) where Study 5’s was r = .409 — the authors attribute the difference to ambiguous confidence instructions in Study 4, which is plausible and unverifiable.

The same question, asked of a construct with no measure at all (Murphy 2022)

Everything above asks the state/trait question of perception — accuracy, sensibility, the quantities the taxonomies already measure. Murphy (2022) poses it for a construct the taxonomies do not even contain: propensity to use interoceptive signals. She leaves the answer open on purpose, and the fork she draws is this page’s own: “if the use of interoceptive signals varies across situations, one’s ability to identify and learn when using interoceptive signals is adaptive — or not — may also be a meaningful individual difference”; but “if interoceptive propensity should be considered a trait, the available evidence suggests it may only be adaptive in certain (likely individual-specific) situations.”

Worth recording because it shows the state/trait problem is not an artefact of the heartbeat task’s instability — it reappears for a construct with no instrument yet built, before any variance has been partitioned. The lesson runs the other way from the sections above: there, a measure existed and turned out to move; here, the construct is defined and the very first question asked of it is whether it is a thing a person has or a thing a situation elicits. See interoceptive-propensity.