Is interoception feed-forward or predictive?
The theoretical fault line introduced by Seth (2013). Everyone in this debate agrees emotion and selfhood are grounded in the body; they disagree on the direction of processing.
Seth’s argument
Interoception “has remained generally understood along feed-forward lines, similar to classical evidence-accumulation theories of exteroception.” Seth argues this is a mistake given cross-talk between levels of viscerosensory representation and top-down influences down to brainstem/spinal centres. Recasting interoception as predictive-coding yields interoceptive-inference and active-inference (autonomic reflexes as fulfilled predictions).
What’s actually contested
Seth is explicit that Damasio, Craig, and Critchley “emphasize a continuous, dynamic, but largely bottom-up interoceptive representational hierarchy… None identify emotional states with top-down inference of the causes of interoceptive signals.” So the disagreement is not about whether the body matters, but whether:
- feelings are read out from an ascending representation (Craig/Damasio), or
- feelings are generated by descending predictions, with ascending signals serving as prediction errors (Seth).
Craig’s actual position (now anchored in primary sources)
With Craig 2002 and Craig 2009 ingested, the “feed-forward” pole can be stated more precisely than Seth’s secondhand characterization. Craig’s architecture is an ascending integration/re-representation hierarchy: primary interoception (posterior insula) → mid-insula integration → AIC meta-representation, with a comparator/buffer at the apex (the global-emotional-moment). Crucially, Craig 2009 is not purely bottom-up: he states the AIC represents “predictions of future feelings” and frames anxiety/functional somatic disorders as “distorted interoceptive predictions.” So the genuine disagreement is narrower and sharper than “bottom-up vs top-down”:
- Craig: predictions/comparison happen at the top of an ascending hierarchy that reads out integrated afferent signals.
- Seth: prediction is the architecture itself — descending generative models predict interoceptive input, and ascending signals are prediction errors (predictive-coding).
Both invoke a “comparator” in the AIC; they differ on whether the interoceptive representation is generated top-down (Seth) or integrated bottom-up and then compared (Craig).
Connection to the classic theories
This is the modern recurrence of the James–Lange question. James’s “perception of bodily change IS the emotion” is ambiguous between a feed-forward read-out and a predictive-inference reading; Seth claims the predictive reading, and positions cognitive-appraisal (Schachter–Singer) as a special case of top-down prediction. Whether the reframing is a genuine empirical advance or a re-description awaiting decisive evidence keeps this open — Seth concedes there is “not yet any direct confirmatory evidence.”
The predictive camp has clinical evidence it does not know about, from 2000
Added with the Van der Does et al. (2000) ingest. It is the first evidence on this page that is neither anatomical nor theoretical, and it comes from neither camp — from panic research, thirteen years before Seth (2013).
The phenomenon. Pool 709 participants counting their heartbeats. More than 95% produce a count; roughly 80% are wrong by ~30%. And the errors have a shape that matters: participants do not report losing track of beats. They report feeling a regular rhythm somewhat slower than their actual heart rate. Confident, steady, and wrong.
Why that shape is awkward for a read-out account. If interoception is bottom-up accumulation of visceral evidence, poor interoceptors should have noisy percepts — degraded, uncertain, unstable. Instead the majority have percepts that are stable, regular, and systematically displaced from the physiology. A confident percept decoupled from the signal is not what a weak read-out produces. It is what a prior produces.
And the authors say so, in Pennebaker’s vocabulary. Symptom schemata formed from past experience “guide the perception and processing of current symptoms”; once threat activates an anxiety schema, perception becomes “more guided by the schema (that is, by past information) than based upon present physiological status.” That is interoceptive-inference with different nouns.
The precision demonstration. The exercise result is the sharpest part and neither camp has cited it. Raise heart rate by exercise and measured accuracy appears above ~100 bpm; let it decay and accuracy vanishes by ~95 bpm — equally in patients and controls, with exactly one of 60 participants durably improving. Nobody learned anything. The balance between prior and evidence shifted because the evidence got louder. In predictive-coding terms that is precision-weighting, manipulated experimentally, with a threshold. It is a cleaner demonstration of the mechanism than anything Seth (2013) or Seth & Friston (2016) offer, and it was produced by people arguing about whether panic patients feel their hearts.
What it does and does not settle. It does not vindicate Seth over Craig — Craig explicitly allows “distorted interoceptive predictions” in anxiety and functional somatic disorders, and this is arguably the phenomenon he is gesturing at. What it does is move the debate’s evidential base: the page has been theorists reading cytoarchitecture and reviews at each other, with Seth conceding “not yet any direct confirmatory evidence.” Here is a large behavioural dataset whose central puzzle is prior-dominated perception, and whose central manipulation moves the prior/evidence balance by changing signal amplitude.
Hold it loosely, for three reasons. Van der Does et al. offer the schema account as an admittedly “hypothetical” reading in a discussion section, with no model and no test. One piece of their own evidence goes the wrong way — the somatosensory amplification scale failed to correlate with counted beats, which a “misinterpreting vague sensations” story predicts it should, and they report the failure. And “a schema guides perception” names the phenomenon rather than explaining it, which is the standard complaint about predictive coding arriving in a form that is more vulnerable to it, not less. See schema-guided-symptom-perception.
The axis may be the wrong one (Petzschner et al. 2021)
The most useful thing Petzschner et al. (2021) do to this page is decline its framing.
First, the debate conflates two claims. “Interoception is inferential” (the brain combines a noisy likelihood with a prior, precision-weighted, to estimate a hidden bodily state) is a computational-level claim with real support — humans behave close to ideal observers in multisensory integration, in integrating past experience, and under abstract beliefs. “Interoception is predictive coding” is a specific algorithmic bet about how that computation is neurally realized, and it is one of several candidates. The page has been treating these as one position because its principal advocates (Seth, Friston, Barrett) hold both. Van der Does et al.’s prior-dominated percepts, recorded above as evidence for the predictive camp, in fact support only the first — a prior beating weak evidence is Bayesian, and says nothing about prediction-error units.
Second, there is an exit neither pole offers. Both Craig and Seth assume the body must be represented — they disagree about which direction the representation is built. HRL shows a substantial class of body regulation could proceed with no percept of the body at all: internal state signalled directly into reward computation (orexin LH→VTA; ghrelin/leptin/insulin receptors in VTA), drive reduction as reward, no interocept required. If that is right for some regulatory behaviour, then for that behaviour this debate has no subject matter — and the interesting question becomes which bodily regulation needs an interocept, rather than which direction the interocept is built.
Third, the anatomical arguments cut less cleanly than they look. The paper notes that parabrachial→insula/vmPFC projections exist in rats but not monkeys, so rodent circuit evidence may not transfer to the human interoceptive hierarchy that both poles of this debate describe.
None of this resolves the debate; it relocates it. The live questions become (a) which algorithm implements the Bayesian computation, and (b) when inference is involved at all. See sensory-control-loop.
A mouse that predicts, and a warning about the measures (Berntson & Khalsa 2021)
Berntson & Khalsa (2021) contribute evidence rather than a position, and it cuts both ways.
The evidence. Hunger/thirst-dependent activity in mouse insular cortex shifts in anticipation of expected satiety of food or water signals (Livneh et al. 2020) — read by the review as showing that “mice can generate and modify interoceptive predictions in response to ongoing contextual changes in the external environment.” Every other line of argument on this page is human imaging, cytoarchitecture, or theory. This is anticipatory interoceptive representation in a preparation that can be silenced, stimulated and recorded, which is the kind of evidence the debate has been conspicuously short of. It does not by itself favour Seth over Craig — Craig allows predictions at the apex — but it moves the question into a species where the architecture can be interrogated.
The warning. Two methodological notes the page should carry, because much of its evidence is heartbeat-based:
- “Perceptual sensitivity for one signal may not generalize to others” (Ferentzi et al. 2018, whose title is precisely Multichannel investigation of interoception: sensitivity is not a generalizable feature). Claims about “interoceptive accuracy” derived from the cardiac channel may not be claims about interoception. See is-the-heartbeat-counting-task-valid, interoceptive-taxonomy.
- Interoceptive processing can affect exteroceptive perception (Motyka et al. 2019, cardiac activity and conscious somatosensory perception), so the two are not cleanly separable at the measurement stage either.
And an alternative to both poles. Berntson & Khalsa doubt that “a purely reductionistic approach” will suffice given the complexity, and recommend a dynamic systems approach that models components and their interactions (Fried & Robinaugh 2020). This is the first appearance in this wiki of a methodological alternative to both the read-out picture and the predictive-coding picture — not an answer to the question, but a claim that the question as posed (which direction dominates?) may not have a direction-shaped answer in a system of continuously reciprocal afferent/efferent loops. See gary-berntson, central-autonomic-network.
A silent posterior insula, and what it is worth (Haruki & Ogawa 2023)
A small item, recorded because the debate is short of evidence and because it is the kind that gets over-cited.
Haruki & Ogawa (2023) had 31 people attend to their heart or their stomach in a scanner. The posterior insula — the primary interoceptive cortex on Craig’s account — did not activate above rest, and its multivoxel pattern did not distinguish the two organs. Only mid and anterior subdivisions did either.
The authors turn the null into a claim for the Craig pole: posterior insula codes ongoing physical change in the bodily signal (Craig et al. 2000; Meier et al. 2018), and an attention paradigm involves no homeostatic perturbation, so there was nothing to code. Attention alone reaches the mid-anterior insula where awareness is represented; it does not engage the primary map, because the primary map is tracking a signal that did not move.
Why the wiki holds this loosely. It is an argument from a null in a study that recorded no physiology at all — no ECG, no electrogastrography, no respiration — so the premise that nothing changed is asserted rather than measured. And the finding is compatible with the predictive reading too: if attention is precision-weighting, a channel attended without a prediction error to weight is exactly where you would expect to find no error signal. Both poles can absorb it, which is the recurring shape of this page’s evidence.
What it does contribute is a constraint on the paradigm rather than on the theory: the interoceptive-attention-task does not engage primary interoceptive cortex, so results from it are results about the re-representation stages. Anything the wiki cites from that literature is a claim about the mid and anterior insula whether or not the paper says so.
The direct evidence arrives (Harrison et al. 2021)
This page has repeated the same concession four times in different voices: Seth (2013) — “not yet any direct confirmatory evidence”; Seth & Friston (2016) — direct evidence “is still lacking”; Petzschner et al. — “the full interoceptive brain network underlying this implementation has not been identified”; Berntson & Khalsa — an open empirical challenge.
Harrison et al. (2021) is the first study in this wiki that goes at it directly, and the design is the right one. Not “does the insula respond to bodily stimuli” but: fit a generative model to each person’s behaviour, take the model’s own trial-by-trial estimates of the theoretically-postulated quantities, and ask whether the brain tracks them. Eighty trials of learning that a cue predicts an inspiratory resistance, at 7T, with reversals to keep the quantities moving.
What confirms
| model quantity | brain | direction |
|---|---|---|
| prediction certainty | dlPFC, anterior insula, ACC, MFG | deactivation |
| prediction error magnitude | anterior insula, ACC, MFG, PAG | activation |
Both signs are forced by the theory and neither is trivial. Greater certainty (precision of belief) should reduce belief updating; greater error should increase it. Getting both, from independently-fitted per-trial regressors, in the regions the anatomy nominates, is more than the page had before.
The PAG result is the additional item: it tracked error magnitude and not certainty — beliefs in cortex, deviation signals in the midbrain, which is the vertical structure the visceromotor hierarchy specifies and which the page had held only as cytoarchitecture.
What embarrasses
The anterior/posterior insular dissociation did not appear. Both quantities lived in the anterior insula, with opposite signs. The only hint in the predicted direction was a valence effect — unexpectedly receiving a resistance activated left posterior insula more than unexpectedly escaping one.
This is a problem for the Seth & Friston structural argument specifically, which is the strongest thing the predictive pole has on this page. That argument says agranular visceromotor cortex is anatomically unfit to receive ascending prediction error. Here, the agranular anterior insula tracks prediction error as strongly as anything in the slab.
Two honest defences. The functional resolution cannot separate cortical layers, which is where the granular/agranular claim actually lives — the authors say so. And the fitted model was Rescorla-Wagner, chosen by pre-registered fallback after no model won selection, so its “prediction certainty” is a transformed point estimate rather than an inferred precision (interoceptive-precision) — a mis-specified regressor can smear two quantities into one region.
What it changes about how to read this page
The Petzschner reframing above says the page conflates a computational claim (interoception is inferential) with an algorithmic bet (it is implemented as predictive coding). Harrison et al. are the first source here where that distinction has empirical teeth: the computational claim gained support and the algorithmic bet lost some. Model-derived prediction and error quantities are tracked by the brain; the specific laminar/regional architecture predicted for them is not what was found.
So the debate does not resolve, but the concession sentence quoted at the top of this section is now out of date, and what replaces it is more interesting: there is direct evidence, and it is mixed.
A behavioural handle: induce the misperception (Iodice et al. 2019)
Harrison et al. are correlational-with-a-model — fit the theory’s quantities and ask whether the brain tracks them. Iodice et al. (2019) are the page’s first manipulation: hold the body fixed, feed the system false interoceptive evidence, and see whether the percept moves. It does — false (faster) heartbeat feedback makes cycling feel harder (illusion of effort) while the actual heart rate stays put (no entrainment, BF = 0.014). A perceived physiological state decoupled from the physiological state, on demand.
This is the kind of evidence the whole page has been short of, but note carefully what it does and does not do:
- It establishes penetrability, not architecture. That expectation/false evidence can override the actual bodily state shows interoceptive perception is prior-penetrable — which is necessary for the predictive account and awkward for a pure bottom-up read-out. But Craig’s comparator-at-the-apex, which explicitly allows “distorted interoceptive predictions,” absorbs it. So it discriminates against strict feed-forward, not between the two live positions.
- The asymmetry is the discriminating detail. A symmetric sensory-conflict model predicts slower feedback should deflate effort as much as faster inflates it; it does not (Bayesian support for the null). A cost-sensitive prior — resisting downward revision because underestimating effort is metabolically dangerous — predicts exactly this. That is precision-weighting shaped by an asymmetric loss function, which is a more specific claim than “the body is penetrable by expectation,” and it is the sort of thing a read-out account has no natural machinery for.
- It supports the computational claim, and is silent on the algorithm. In the Petzschner framing above, a prior beating sensory evidence is Bayesian inference; it says nothing about prediction-error units. So this joins Van der Does’s prior-dominated percepts on the “inference, algorithm-unspecified” side.
Brakes the page should carry: N = 18, all male, one channel, and the effect depends on a laboratory-induced belief that the sound is one’s own heart — so it is a strong existence proof (interoceptive illusions are real and manipulable) rather than a measurement of how large a role prediction plays in ordinary interoception. See interoceptive-illusion, false-feedback-paradigm.
A structural (not just functional) argument: Seth & Friston (2016)
Seth & Friston (2016) moves the debate slightly by grounding it in cytoarchitecture rather than function alone: the visceromotor areas (VMAs) — AIC, ACC, subgenual cortex, OFC — are agranular/dysgranular, lacking the well-formed granular layer IV that canonically receives ascending, prediction-error-carrying afferents in laminar models of cortical hierarchy. If descending-prediction cortex is structurally distinguishable from ascending-error-receiving cortex, that is independent (anatomical) evidence for reading the interoceptive hierarchy as generative rather than feed-forward — still not “direct confirmatory evidence” for interoceptive prediction/error signals per se, but a stronger anatomical premise than Seth (2013) offered alone. Still concedes direct evidence “is still lacking,” so the debate remains open.