Interoceptive inference
Introduced/crystallized by Seth (with Critchley, Friston, Hohwy) and given its canonical statement in Seth (2013). The claim: emotions are the integrated content of hierarchical top-down interoceptive predictions that counterflow with bottom-up interoceptive prediction errors, computed within a salience network anchored on the anterior insular cortex (AIC) and anterior cingulate.
Relation to appraisal theory
Interoceptive inference is explicitly framed as the predictive-coding generalization of cognitive-appraisal theories. Appraisal says: emotion = physiological change + cognitive interpretation of it. Interoceptive inference reframes “interpretation” as Bayesian inference on the hidden causes of interoceptive signals. So James–Lange, Schachter–Singer, and Damasio/Craig become special or partial cases within a single predictive scheme.
Two ways to reduce interoceptive prediction error
- Perceptual/model updating — change the generative model to fit incoming visceral evidence. Farb et al. (2015) elaborate this route into perceptual-inference, giving it a full clinical/contemplative treatment (mapped onto acceptance and equanimity) that Seth’s original paper left comparatively underspecified.
- active-inference — engage autonomic reflexes to change bodily state so it matches the prediction. This requires transient low precision (attenuated attention) on interoceptive prediction errors, otherwise they would revise the model instead of driving autonomic action.
Applied extension: the simulation map
Farb et al. (2015) build a clinical/contemplative superstructure on top of this framework, proposing the simulation-map as the operative construct closest to interoceptive awareness (distinct from raw sensation), and extending Seth’s presence-and-agency constructs so that presence is reachable via perceptual inference and not only active inference. See mindfulness-interoceptive-training and interoceptive-training-clinical for the resulting wellbeing and clinical applications.
Empirical footholds
No direct confirmation yet, but: AIC encodes anticipated and prediction-error signals for pain, risk, itch, and social “feeling” outcomes; false cardiac feedback increases emotional salience and AIC activity; cardio-visual synchrony modulates the experience-of-body-ownership. Applied to psychopathology: anxiety, psychosis, alexithymia, and depersonalization (imprecise interoceptive predictions) as failures of interoceptive inference.
Contrast with predecessors
Damasio, Craig, and Critchley also ground selfhood in bodily representation, but treat the interoceptive hierarchy as largely bottom-up (an ascending posterior→anterior insular re-representation with a comparator at the apex); none formalize emotion as a top-down generative model of interoceptive causes. That is the distinction at the heart of feedforward-vs-predictive-interoception — where Craig’s own gesture toward “predictions of future feelings” (Craig 2009) marks how close, yet architecturally distinct, the two accounts are.
The interocept, and what “interoceptive” is actually indexing (Petzschner et al. 2021)
Petzschner et al. (2021) supply two pieces of terminology this page has been managing without.
The interocept. The inferred internal state — the posterior — “is what is typically called the percept (or in the case of internal states, the interocept).” It is an estimate of the real internal state, never the state itself. That word is useful precisely because it keeps the estimate and the estimand apart, which most discussion here does implicitly: the accuracy literature (interoceptive-taxonomy, is-the-heartbeat-counting-task-valid) is entirely about how far the interocept tracks the state.
A cleaner internal/external criterion. “Interoception and exteroception can thus be distinguished based on the type of state that is being inferred — internal or external — rather than the specific sensory channels that contributed to the inference.” Exteroceptive channels count as interoceptive input whenever they carry information about an internal state. This is a real improvement on the channel-based definition the wiki mostly uses, and it retroactively softens a caveat carried on several pages: Farb (2011)‘s protective sensory pole is visual, and has been flagged everywhere as “exteroceptive, not interoceptive.” On Petzschner et al.’s criterion the question is not which cortex lit up but which state was being inferred — which does not dissolve the caveat, but reframes it as a question about the study’s design rather than a category error.
Formalized and extended: Seth & Friston (2016)
Seth & Friston (2016) restate this model with the full free-energy/active-inference apparatus and add: (1) an anatomical argument that the AIC/ACC/OFC complex — named visceromotor areas (VMAs) — is agranular (lacking layer IV), structurally suited to issuing descending predictions rather than receiving ascending error; (2) a distinction between epistemic active inference (information-seeking action) and instrumental active inference (regulatory action, per perceptual control theory); and (3) clinical extensions to autism (failure to attenuate interoceptive prediction-error precision; see interoceptive-trait-prediction-error) and depression/fatigue (chronic dyshomeostasis from noisy interoceptive afferents, or metacognitive beliefs about low allostatic self-efficacy).