NOTE-1000: Seth — Interoceptive inference, emotion, and the embodied self

Contribution

The article applies predictive coding, until then worked out mainly for exteroception and action, to the sense of the body’s internal state. It makes two claims under one mechanism. First, felt emotion is the content of top-down inference about what is causing interoceptive signals, not a bottom-up readout of the body later appraised. Second, the experience of owning a body is the same kind of inference, run across interoceptive and exteroceptive signals at once. This turns the rubber hand illusion from a result about vision, touch and proprioception into one where the heartbeat can also count. The article offers this as a framework with indirect support. It says itself that there is no direct test yet.

Key insight

The brain has to model the body it belongs to, inside and out, as it models anything else. It infers which signals are “most likely to be me” (p. 567, after Apps & Tsakiris). What is felt as emotion, and as a body being one’s own, is the current best guess of that model. Prediction error can be reduced in two ways: by changing the guess, or by making the body fit it. Pushing the viscera towards predicted set points is the interoceptive counterpart of moving a limb to fulfil a proprioceptive prediction.

Assumptions

• Predictive coding is the brain’s general scheme. Perception is “probabilistic, knowledge-driven inference on the external causes of sensory signals” (p. 566). Errors flow up and predictions down, weighted by precision. Attention is “the optimization of precision weighting” (p. 567). Box 1 grants that “unequivocal neural evidence for PC is still lacking” (p. 566).
• The free-energy reading. Under Gaussian assumptions, free energy is prediction error. Active inference is taken from Friston et al. 2010 (ref. 21), and active motor inference from Adams et al. 2013 (ref. 29).
• Selfhood is a “constellation concept” (p. 565). It runs from homeostatic control through body ownership, perspective and agency to the metacognitive and narrative “I”, citing Metzinger’s Being No One (ref. 17). The article treats body ownership and emotion. The narrative levels are called “more speculatively” within reach (p. 570).
• Interoception is not feed-forward. The case rests on evidence of “substantial cross-talk” between levels, including “top-down cortical and behavioural influences to brainstem and spinal centres” (p. 567, citing Critchley & Harrison 2013).
• The anterior insula is the comparator. Most of the evidence “rest[s] on assuming a central role for the anterior insular cortex” (p. 567).

Key results

There are no new data apart from Figure 3, which reproduces the author’s own group’s study. By section:

• Emotion as interoceptive inference (p. 567).
  • “Interoceptive inference involves hierarchically cascading top-down interoceptive predictions that counterflow with bottom-up interoceptive prediction errors”. Felt emotions “are hypothesized to depend on the integrated content of these predictive representations across multiple levels”.
  • The lineage given runs from James and Lange, through “appraisal” theories, to Schachter and Singer’s adrenaline study and their two-factor theory, “emotion as interpreted bodily arousal”. The new model “generalizes” appraisal (p. 565).
• Active interoceptive inference (p. 567, Figure 2).
  • Interoceptive errors can be cancelled “by transcribing these predictions into reference points for autonomic reflexes” (after Gu et al. 2013). This “directly parallels” active inference in motor control.
  • Descending predictions drive reflexes “only if the corresponding error signals have diminished impact … via transiently low precision weighting”. So “active interoceptive inference depends on the selective attenuation of attention to interoceptive prediction errors”.
  • Figure 2 places the comparator in a “salience network” anchored on the anterior insula and anterior cingulate. Autonomic outflow from them is “in the form of interoceptive predictions that enslave autonomic reflexes”.
• Evidence (pp. 567–569). “There is not yet any direct confirmatory evidence for interoceptive inference”, only “supportive data”. These are:
  • fMRI of the anterior insula, which responds to anticipated pain and affective touch, and encodes predicted pain with pain prediction error, and risk with risk prediction error;
  • contagious itch (Holle et al.), read as a conflict between predicted and actual interoceptive state;
  • Bayesian updating of expected feelings in the ultimatum game, with insula activity scaling with “feeling prediction errors” (Xiang et al.);
  • false cardiac feedback, which enhanced emotional ratings (Valins 1966) and raised right-insula responses and the emotional salience of neutral stimuli (Gray et al. 2007). The verdict: “indirect support” (p. 569).
• Box 2 (pp. 568–569). The anterior insula is described as a “limbic sensory area” beside the cingulate’s “limbic motor area”. It has a high density of von Economo neurons, associated “frequently though circumstantially” with awareness and selfhood. Insula activity and morphometry predict heartbeat-detection sensitivity. Its role in conscious access is hedged “(though see [43,74])”, where ref. 43 is Damasio et al. 2013 on feelings persisting after bilateral insula damage.
• Rival models (p. 569). Damasio, Craig and Critchley all ground the self in representations of the body. But “none identify emotional states with top-down inference of the causes of interoceptive signals”; they “emphasize a continuous, dynamic, but largely bottom-up interoceptive representational hierarchy”. Singer et al.’s account of insula error-based learning uses “prediction error” for change detection and salience, not predictive coding.
• Body ownership (p. 569).
  • The rubber hand illusion (Botvinick & Cohen 1998) is described, with face (enfacement) and whole-body versions (Lenggenhager et al. 2007; Ehrsson 2007).
  • The standard multisensory explanation, visual capture from sensory correlations, is “compatible with PC”. Highly precise vision and touch “could overcome prediction errors in a different modality (proprioception)”. The result is a revised model “incorporating the fake hand as part of the self-representation”.
  • “Perceptual correlations are not necessary for the RHI; the expectation of correlated sensory input is sufficient” (Ferri et al. 2013).
  • Lesion and imaging data link body ownership to the anterior insula (Karnath & Baier 2010).
• Interoception and the rubber hand (p. 569).
  • “Participants with lower IS are more susceptible to the RHI” (Tsakiris et al. 2011), and also to shifts of self–other boundaries. Seth reads this as “possibly reflecting lower precision-weighting of interoceptive prediction errors”.
  • During the illusion the real hand cools and its histamine reactivity rises. Cooling the real hand raises susceptibility. Threats to the rubber hand raise skin conductance. All are called “compatible with descending self-related predictions tuning reference points for autonomic reflexes”.
• The cardio-visual rubber hand (pp. 569–570, Figure 3).
  • Suzuki et al. 2013 built an augmented-reality hand whose colour pulsed towards red over 500 ms, in or out of time with the heartbeat. Synchronous feedback “enhanced the experience of ownership of the virtual hand”, both by questionnaire and by proprioceptive drift, with the parenthesis “these two measures do not always correlate”.
  • The same setup replicated the tactile illusion. When finger movements were mapped to the virtual hand, visual–proprioceptive coherence “overshadowed the influence of cardiac (interoceptive) prediction errors”. The caption: the drift effect held in the “cardiac still” but not the “cardiac move” condition, and control questions showed no effect.
  • Aspell et al. 2013 applied the same feedback to a virtual body seen from behind, and synchrony “increased identification with the virtual body”, objectively and subjectively.
  • The inference drawn: “statistical correlations between interoceptive (e.g., cardiac) and exteroceptive (e.g., visual) signals can lead to updating of predictive models of self-related signals … just as may happen for purely exteroceptive multisensory conflicts” (p. 570).
• Conclusion (pp. 570–571). “Emotion and embodied selfhood are grounded in active inference of those signals most likely to be ‘me’”. “Perceptual scenes that evoke interoceptive predictions will always be affectively coloured.”
• Box 3, psychopathology (p. 571).
  • Alexithymia goes with failure to engage the anterior insula.
  • Low interoceptive sensitivity “provides a bridge” to disorders of body ownership.
  • Depersonalization and derealisation, with their “loss of ‘emotional colour’”, may come from “imprecise (as opposed to inaccurate) interoceptive predictions”, after Seth et al. 2011.
  • The step to Cotard’s delusion “may also reflect aberrant high-level inference”.
• Box 4 (p. 571). The open questions include how interoceptive predictive models relate to “similar models of agency”, and whether the narrative self can be accommodated.

Claims

id	claim	strength	support
C1	Felt emotion is the content of top-down predictions of the causes of interoceptive signals	weak: a hypothesis, explicitly without direct test	argument from predictive coding; indirect data, pp. 567–569
C2	Interoceptive prediction errors are also cancelled by autonomic action, which needs attenuated precision on those errors	weak: by analogy with active motor inference	Gu et al. 2013; Adams et al. 2013; p. 567
C3	The anterior insula is the comparator for interoceptive prediction and error	moderate for its anticipatory and error responses; weak for the comparator role	cited fMRI studies, Box 2; pp. 567–569
C4	Manipulated interoceptive feedback changes emotional salience and insula activity	moderate as reported	Valins 1966; Gray et al. 2007; p. 568
C5	The rubber hand illusion is a revision of a multisensory self-model that minimises precision-weighted prediction error	weak as a reinterpretation, compatible with the standard account	argument; Ferri et al. 2013 on expectation; p. 569
C6	Lower interoceptive sensitivity goes with greater susceptibility to the rubber hand illusion	moderate as reported; the precision reading is Seth’s	Tsakiris et al. 2011; p. 569
C7	Heartbeat-synchronous visual feedback increases ownership of a virtual hand and identification with a virtual body	moderate: two studies, one the author’s own	Suzuki et al. 2013 (Figure 3); Aspell et al. 2013; pp. 569–570
C8	Visual–proprioceptive coherence from movement overrides the cardiac cue	weak: one condition of one study	Suzuki et al. 2013, p. 569
C9	Depersonalization reflects imprecise interoceptive prediction	speculative	analogy with precision models of psychosis; insula hypoactivation, Box 3

Concepts

• Interoceptive inference (equivalently here, interoceptive predictive coding): hierarchical top-down prediction of interoceptive signals, compared with bottom-up error. Felt emotion depends on its integrated content.
• Active (interoceptive) inference: cancelling prediction error by acting, here through autonomic reflexes whose reference points are set by descending predictions.
• Precision: the inverse variance of a prediction error, which sets its influence. Attention is the optimisation of precision.
• Experience of body ownership (EBO): “the experience of certain parts of the world as belonging to one’s body”, for a part (a hand) or globally (a whole body) (glossary).
• Interoceptive sensitivity (IS): a trait, “usually operationalized via heartbeat detection tasks” (glossary).
• Selfhood: “the experience of being a distinct, holistic entity, capable of global self-control and attention, possessing a body and a location in space and time”, citing Blanke & Metzinger 2009 (glossary).
• Signals “most likely to be me”: the target of self-related inference, after Apps & Tsakiris.

Connections

The article builds on predictive coding as Clark, Friston and Hohwy state it, and on the author’s earlier interoceptive model of “conscious presence” (Seth et al. 2011, ref. 6). For emotion it builds on Gu et al. 2013. On body ownership it builds on Botvinick & Cohen 1998, Lenggenhager et al. 2007, Blanke 2012 and Blanke & Metzinger 2009, and on Tsakiris’s work on interoceptive sensitivity. It takes the free-energy self of Apps & Tsakiris (cited as 2013, online) as its frame for self-recognition. It sets itself against the bottom-up interoceptive accounts of Damasio (2000; Damasio & Carvalho 2013), Craig and Critchley. Of these, the record holds Damasio & Carvalho (LIT-388); Botvinick & Cohen, Lenggenhager et al., Blanke & Metzinger and Apps & Tsakiris are being filed alongside it. The reference list gives Suzuki et al.’s title with “interoceptive and exteroceptive” reversed; Crossref has “exteroceptive and interoceptive”.

Bearing on the record

• The interoception cluster and body ownership. Before this filing the record held interoception as the ground of feeling (LIT-388, NOTE-330; LIT-628, NOTE-507), of emotion (LIT-542, NOTE-434) and of brain function (LIT-996, NOTE-761). It held body ownership only through Metzinger (LIT-450, NOTE-372). This is the work that joins them: it makes ownership an inference over interoceptive and exteroceptive signals, and it cites evidence that the heartbeat enters it (C6, C7).
• THEORY-045 (felt affect). The account leaves open whether what is felt is a prediction of the body’s state or a map of it. Seth takes the prediction side explicitly, and places Damasio’s account, including LIT-388 (his ref. 14), on the bottom-up side (C1). He gives no case where prediction and map come apart, which is what THEORY-045’s promotion test asks for. He does cite the other kind of result it names: false cardiac feedback that moves emotional ratings and salience (C4). That is reported at second hand, with emotive stimuli, so it falls inside an emotion episode, not outside one as the test requires. He also cites, as a hedge, Damasio et al. 2013 on feelings surviving insula damage, the case NOTE-330 reads for Damasio’s side. A source for THEORY-045, not a test of it. It bears too on QUESTION-059, through C2: on this account the felt state is tied to a body that the predictions act on.
• NOTE-434 (Barrett). Barrett’s 2017 synthesis cites this article for its claim C1. The two agree on prediction. They differ on what does the work. For Seth it is the inferred cause of interoceptive change; for Barrett it is a concept that categorises. Seth’s comparator is the anterior insula, while Barrett makes the agranular limbic cortex the source of predictions. The comparison is mine.
• LIT-135, NOTE-169 (Seth 2024). This is the earlier statement of the interoceptive predictive processing on which Seth’s biological naturalism rests (NOTE-169, C4). Here it is a claim about emotion and self, not about consciousness in general, and nothing in it says that only living systems could run such inference.
• LIT-450, NOTE-372, THEORY-064 (the self-model). Metzinger’s C6, that body ownership can be manipulated by multisensory conflict, gets a mechanism here: the self-model is revised to cut precision-weighted error. It also gets an interoceptive channel the article does not have. Seth’s account of the rubber hand illusion is a self-model being revised, which is consistent with THEORY-064. But it is silent on transparency and on whether anything beyond the model is a self, so it neither supports nor tests THEORY-064’s distinctive claims. Box 3’s depersonalization proposal (C9) is the kind of case THEORY-064’s promotion test names. It is offered there as a precision failure, not as a self-model becoming opaque, and it is speculative.
• LIT-1181, NOTE-960 (Friston et al. 2010). This is the article’s ref. 21, the source of its active inference. Seth extends that paper’s motor account, action fulfilling proprioceptive predictions, to autonomic reflexes (C2). The extension is by analogy and is not derived.
• LIT-996, THEORY-291 (Theriault et al.). Figure 2’s salience network, anchored on the anterior insula and cingulate and driving brainstem visceromotor targets, is an early form of the visceromotor–interoceptive system THEORY-291 describes. Seth gives it as a schematic, not as connectivity data.
• LIT-1009, NOTE-827 (Haggard on agency). Box 4 asks how interoceptive models relate to “similar models of agency”. Haggard’s review reports the anterior insula as the only area consistently associated with positive self-agency, the region Seth makes the interoceptive comparator. Neither paper draws the link; it is mine.
• THEORY warranted. No THEORY in the record states that body ownership draws on interoceptive as well as exteroceptive signals. This reading supports proposing one, and the sibling filings on the rubber hand and full-body illusions would be its other sources. THEORY-767, filed on 2026-10-20, now states it.
• No instruction for machine-learning practice; nothing for the anthology.

Limitations

• An opinion article. Every empirical result is cited and not shown, except Figure 3, which reproduces the author’s own study. No effect sizes are given in the text.
• The author says “there is not yet any direct confirmatory evidence” (p. 567). Most of the support depends on assuming the anterior insula is the comparator.
• The body-ownership account explains the standard rubber hand results again in predictive-coding terms. The article calls them “compatible with PC” and gives no result that predictive coding predicts and a plain multisensory account does not.
• The interoceptive-sensitivity finding rests on heartbeat detection, whose validity the article does not discuss, and on one study cited for the rubber hand illusion.
• The autonomic results (cooling, histamine, skin conductance) are said to be “compatible with” active inference, not predicted by it.
• Two comment letters in Trends in Cognitive Sciences (2014) are listed by PubMed and were not read.

Open questions

• What result would show that felt emotion follows the predicted rather than the actual interoceptive state? The false-feedback studies come closest, but they change the signal, not the prediction.
• Is the effect of cardio-visual synchrony on ownership interoceptive, or just one more temporal correlation that vision can lock to? A control using a non-bodily periodic signal at heart rate would separate the two.
• How do interoceptive predictive models relate to “similar models of agency”? Box 4 asks this and leaves it open.
• Does low interoceptive sensitivity mean low precision on interoceptive error, as Seth reads it, or simply a worse heartbeat-counting performance?