Insular cortex

The cortical home of interoception and the anatomical spine running through nearly every study in this wiki. Craig’s two papers (2002, 2009) established its organization; Seth, Critchley, and Damasio all build on it. See ad-craig.

The posterior→mid→anterior gradient

• Dorsal posterior insula — the primary interoceptive cortex: the topographic terminus of the lamina-i-spinothalamocortical-pathway (via VMpo/VMb). Objective bodily state (e.g., graded cooling) maps here linearly. Distinct from the parietal somatosensory cortices (S1/S2) — a neural separation of interoception from exteroception.
• Mid-insula — integrates the primary interoceptive image with emotionally salient environmental, multisensory, and hedonic input (amygdala, nucleus accumbens; see salience-network). Associated with body ownership/agency (e.g., rubber-hand-illusion activation).
• Anterior insular cortex (AIC) — the re-representation: subjective feeling (“how you feel”) rather than objective state. In humans this is lateralized (see asymmetry below) and is proposed as the seat of embodied-selfhood and, per Craig 2009, of awareness itself (the global-emotional-moment).

The gradient derived from rest, not cytoarchitecture (Cauda et al. 2011)

The posterior→mid→anterior gradient above is stated from anatomy — the lamina I terminus, the granular/agranular seam, the cytoarchitectural tripartition. Cauda et al. (2011) recover nearly the same split from a wholly independent source: the resting-state correlation structure of spontaneous BOLD, no task and no reference to cell types. Ten insular seed ROIs in 16 adults, and three clustering methods (k-means, hierarchical, fuzzy c-means) converge on a ventral-anterior / dorsal-posterior / transitional tripartition: the ventral-anterior insula couples to rACC, middle/inferior frontal and temporoparietal cortex (a limbic/salience pattern), the dorsal-posterior insula to premotor/SMA/sensorimotor cortex and mid-posterior cingulate (sensorimotor integration), with the middle gyri transitional. The two networks are anticorrelated (anterior vs posterior time courses r = -0.47 in a sample subject), which the wiki’s predictive sources later partition by computational role (anterior = descending prediction, posterior = ascending error — see the EPIC section below). So the functional bipartition that the rest of this page assumes is not merely inferred from Nissl stains; it falls straight out of how the resting insula fluctuates. The paper’s lateralization — right anterior ROIs preferentially wired to brainstem, pons and right thalamus — is also the resting-state anticipation of the Haruki deflation (right AIC as an arousal node) and of Craig’s right-AIC-as-arousal asymmetry, reached from connectivity rather than autonomic argument.

The anterior insula is itself two subdivisions (Deen et al. 2011; Chang et al. 2013)

Cauda above recovers a bipartite insula from rest. Deen, Pitskel & Pelphrey (2011) resolved the anterior end into two first — voxelwise k-means clustering of resting connectivity split each insula into a ventral anterior, a dorsal anterior-to-middle, and a posterior subregion, with the ventral-anterior cluster coupling to pregenual ACC and limbic cortex, the dorsal-anterior to dorsal ACC and the cognitive-control network, and the posterior to somatomotor cortex — and validated the split on an independent disgust task (both anterior clusters responded to disgusting images, posterior did not). It also drew the cingulate end: pregenual ACC↔vAI, dorsal ACC↔dAI, middle cingulate↔PI, so the insula+ACC pairing this page treats as one unit is really three parallel loops (a ventral emotional, a dorsal cognitive, a posterior sensorimotor). Chang, Yarkoni, Khaw & Sanfey (2013) reproduce that tripartition in the right insula and — the distinctive part — attach a function to each subdivision by large-scale reverse inference rather than by activation. Data-driven k-means parcellation of the right insula (validity index selecting k=3) first splits anterior from posterior, then splits the anterior insula again into a dorsoanterior node (coupled to ACC and DLPFC) and a ventroanterior node (coupled to amygdala, ventral striatum, VTA, OFC). The tripartition replicates from resting-state connectivity and from meta-analytic coactivation across ~4400 studies — two independent data classes returning the same three networks.

This matters for a page that mostly treats “the AIC” as a single object. The anterior insula’s dorsal part is an executive node and its ventral part an affective-chemosensory one — the same seam the wiki’s finer results keep landing on: Haruki & Ogawa’s right dorsal anterior insula preferring cardiac/arousal attention, and the EPIC model’s agranular ventral anterior insula as the visceromotor/prediction pole (below). Reverse-inference decoding gives each subdivision a specific function: ventroanterior = emotion, chemosensation (olfaction/gustation), autonomic; dorsoanterior = executive control (inhibition, error, conflict, switching); posterior = pain, sensorimotor, language.

A third method reaches the same anterior split. Vercelli et al. (2015) — Cauda’s own Turin group returning to its bipartition — drive the resolution to 12 fuzzy-c-mean nodes per insula and, once the shared anterior/posterior variance is regressed out, find the individual nodes carrying the “echoes” (Leech et al. 2012) of whole-brain networks nested inside the insula: default-mode, sensorimotor and dorsal-attentional. At the anterior end this validates Touroutoglou et al.’s (2012) dorsal (dFP) / ventral (vFP) frontoparietal split — dorsal for top-down/bottom-up salience integration, ventral for emotional salience and “the integration of bodily feelings” — the same dorsal-executive/ventral-affective seam Deen and Chang reach, now from a fourth angle (high-dimensional fuzzy clustering). So the anterior-insula-is-two-subdivisions claim now holds across connectivity-clustering (Deen), reverse inference (Chang) and high-dimensional fuzzy parcellation (Vercelli). The “echoes” result is also a domain-generality data point (below): insular nodes wearing other networks’ signatures fit the general-integrator reading better than a dedicated body-sense one.

The conceptual payload is a distinction this whole page depends on and rarely states: consistency is not specificity. By forward inference the dorsoanterior network is the most consistently activated of the three across nearly all tasks — the pattern that led some to call it a general “goal-directed cognition” region. By reverse inference it is nonetheless specifically executive; its ubiquity reflects that executive functions (attention, goal-monitoring, arousal) are prerequisites for most tasks, not that the region lacks a function. That is the same logic the Lindquist section below uses to reject insula-as-disgust-module (selective ≠ specific), and it is the caution behind every “the insula activated, therefore…” inference on this page. See meta-analytic-reverse-inference.

The parcellation is experience-dependent (Gong et al. 2015)

The sections above treat the anterior/posterior functional split as a stable property of the structure — recovered identically from cytoarchitecture, from rest, and from meta-analytic coactivation. Gong et al. (2015) add the one thing that evidence cannot show: the split can be moved by experience. Taking Cauda’s ten ROIs and A-/P-network scheme as a template, they compared 27 action-video-game champions against 30 amateurs and found, in experts, higher functional integration between the anterior and posterior insula (left-lateralized, correlating r=0.68 with weekly playing time), increased grey-matter volume near the anterior/posterior transitional seam (correlating with both the integration index and playing time), and a rewiring in which the posterior sensorimotor insula becomes coupled to a frontal attentional node (MFG). The amateurs, meanwhile, reproduce Cauda’s bipartite split — a small independent replication of the parcellation.

Two cautions before this is read as “the insula is plastic.” The design is cross-sectional, so it cannot separate training-sculpts-insula from insula-selects-champions; the playing-time correlations are the only lever toward an experience account and are themselves correlational. And it is emphatically not an interoception study — the insula is treated as a generic attention/sensorimotor integrator, no interoceptive measure taken, its function imported wholesale from Cauda. That second point is itself a data point for the domain-generality question: an account of “what this cortex does” that never mentions the body works fine here, echoing the Farb and Haruki deflations that the anterior insula “is not an area of pure body awareness.” Sits beside the Farb et al. (2007) result below — the wiki’s other evidence that the insula’s functional partners, not just its activation, are trainable.

Limbic sensory cortex

Craig casts the insula as limbic sensory cortex (feeling) paired with the anterior-cingulate-cortex as limbic motor cortex (motivation/agency): insula + ACC = an emotion. The two are co-activated in nearly all emotion studies and interconnected by von-economo-neurons.

Human specificity

The AIC re-representation extends only to the middle of the insula in macaques but to the anterior limit in humans — Craig argues the human AIC “has no equivalent in the monkey.” Its expansion (and its VEN content) is proposed to underlie uniquely human self-awareness.

Forebrain emotional asymmetry (Craig 2009, Box 3)

Craig proposes the two AICs are functionally asymmetric, tied to autonomic opponency:

• Right AIC ↔ sympathetic / energy-expenditure / arousal / withdrawal (aversive) / individual-survival emotions (e.g., pain, anger, anxiety).
• Left AIC ↔ parasympathetic / energy-nourishment / relaxation / approach (appetitive) / affiliative emotions (e.g., maternal & romantic love, joy, smiling, pleasant music).

He grounds this in the asymmetric autonomic innervation of the heart and brain energy-optimization. It is a distinctive but contested claim (a hypothesis awaiting split-brain and direct-comparison tests), recorded here rather than given its own page.

The predictive fault line

Craig’s insular hierarchy is ascending re-representation with a comparator at the top (AIC). Seth (2013) adopts the same anatomy but reinterprets its computation as top-down predictive inference (predictions descend, prediction errors ascend). That is the core of feedforward-vs-predictive-interoception. Seth & Friston (2016) add a structural argument: the AIC (as one of the visceromotor-areas) is agranular, lacking the granular layer IV that would mark it as a target of ascending prediction-error input — anatomy consistent with it being a source of descending prediction, not merely a comparator perched atop an ascending stack.

The insula split by granularity: the EPIC model (via Khalsa et al. 2018)

The Khalsa et al. (2018) roadmap adopts Barrett & Simmons’s Embodied Predictive Interoception Coding (EPIC) model (its Fig 4), which cuts the insula along the granular/agranular seam and assigns the two ends opposite computational roles:

• the granular mid and posterior insula is primary interoceptive sensory cortex — it returns ascending prediction error to the visceromotor regions;
• the agranular ventral anterior insula is a visceromotor region — it issues descending allostatic predictions (to hypothalamus, brainstem, spinal cord) and simultaneously predicts the sensory insula.

Under usual conditions the agranular regions are “relatively insensitive” to the ascending error, which the roadmap notes is why interoceptive predictions stay stable despite bodily fluctuation. The clinical hypothesis (Paulus & Stein) is that mental illness is that insensitivity breaking down: interoceptive input (posteriors) decouples from interoceptive predictions (priors), and the resulting elevated error “may present in the brain as ‘noisy afferent interoceptive inputs’.” This maps the wiki’s posterior→anterior gradient onto a sensory→visceromotor prediction loop and gives it a failure mode — see theory-of-constructed-emotion, visceromotor-areas, computational-psychiatry. The roadmap also cites Hassanpour et al. (2018) that the insula “dynamically maps changes in cardiorespiratory interoception,” empirical support for the mid-insula as an active interoceptive mapper rather than a static relay.

Not a “disgust module”: the meta-analytic reframe

Lindquist et al.’s (2012) neuroimaging meta-analysis directly tested — and rejected — the locationist claim that the anterior insula is specifically the brain seat of disgust (Calder, Wicker, Jabbi). The insula is functionally selective for disgust (more consistently active for it than other categories) but not specific to it: it also activates during body-movement awareness, gastric distension, and orgasm, and during anger and sadness experience. Their positive reframing explicitly cites Craig (2002) and Craig (2009) as support: the insula’s real role is representing core affective feelings and interoceptive awareness generally, not any one discrete emotion. This confirms Craig’s own claims (he never proposed disgust-specificity) while refuting the more locationist popular reading of them — see locationist-vs-constructionist-brain-emotion. It does, however, sharpen a genuine tension with Craig’s forebrain emotional asymmetry hypothesis below, which is itself a category-cluster-to-hemisphere locationist claim not yet tested against Lindquist et al.’s functional-specificity criteria.

Damasio’s insula, and why the annexation is one-directional

Craig casts his interoceptive anatomy as the substrate Damasio’s hypothesis requires: the right AIC supplies the “mental image of one’s physical state” that somatic markers presuppose. Reading Bechara & Damasio (2005) from the other side shows the traffic runs one way only.

In the somatic marker framework the insula appears exclusively as “insular/SII, SI cortices” — an undifferentiated block, always named as a triple, doing two jobs: storing the patterns that somatic states leave behind, and holding the conscious feeling of a somatic state (“what it feels like”), as against the non-conscious patterns in brainstem nuclei like the PBN. Posterior VM (BA 25) connects to it directly; anterior VM indirectly (Ongur & Price 2000), which is what makes posterior coupling fast and strong. See ventromedial-prefrontal-cortex.

What is absent is everything this page is built on: no posterior→mid→anterior gradient, no lamina I pathway, no VMpo, no re-representation, no distinction between the insula and the parietal somatosensory cortices it is lumped with — and no citation of Craig anywhere in the paper.

So Craig’s framework is compatible with the somatic marker hypothesis and considerably more specific than it. It is not drawn from it, and Damasio shows no sign of having taken it up. Two consequences worth recording:

• The claim that somatic markers are an interoceptive theory is Craig’s contribution, not Damasio’s. In the primary source, the insula is where somatic patterns live; whether the subject perceives them is never at issue, and the striatal biasing level is explicitly covert. See does-somatic-feedback-guide-decisions.
• The as-if body loop — somatic states activated directly in insula/SII/SI without peripheral enactment — is a claim about this cortex running a body model offline. That is the same job interoceptive-inference and the simulation-map assign it, in vocabularies thirty years apart. But Bechara & Damasio never specify whether the stored pattern is activated as a prediction, a memory read-out, or an efference copy, so the somatic marker hypothesis takes no position on feedforward-vs-predictive-interoception. It predates the machinery that would give it one.

The insula decoupled from the self-model (Farb et al. 2007)

Before the 2010 sadness study, Farb et al. (2007) had already made the right insula the pivot of a self-reference dissociation, and the finding is about its connectivity rather than its activation. Attending to momentary experience (experiential focus) rather than to an enduring self-concept (narrative focus) recruited a right-lateralized viscerosomatic network — insula, SII, inferior parietal lobule — but only reliably in mindfulness-trained participants. The mechanistic result: in untrained novices the right insula was strongly coupled to the vmPFC (R = 0.61), so interoceptive signals were by default read through the medial-prefrontal self-model; training abolished that coupling (R = 0.06) and re-routed the insula to dlPFC (R = 0.28 → 0.78). This is the wiki’s earliest evidence that the insula’s functional partners — not just its activation level — are plastic, and that present-moment body awareness may depend on releasing the interoceptive signal from the narrative self. It also seats the 2007-era reading of the right AIC as the viscerosomatic-awareness hub that this page’s later sections (Haruki & Ogawa, Farb 2013) go on to narrow.

The insula as a trainable, protective response to emotion (Farb et al. 2010)

Most of this page’s evidence is anatomical or phylogenetic; Farb et al. (2010) is where the insula shows up as a state that intervention can move and clinical outcome can track. Watching a sad film, untrained controls deactivated the right insula (alongside subgenual ACC and somatosensory cortex) while activating a midline self-referential + language network — the felt body dropping out as verbal elaboration comes online. MBSR completers, at equal felt sadness, kept the right insula online (the strongest training-related increase, MT>control), and right insula recruitment correlated with lower depression (BDI r=-.465), anticorrelated with language-area (Wernicke’s) recruitment, which itself tracked higher depression. So the same cortex Craig places at the top of the interoceptive hierarchy behaves here as an experiential counter-pole to elaborative processing, and more of it under a dysphoric challenge goes with less dysphoria.

Two connections worth flagging. First, the effect is right-lateralized — consistent with Craig’s forebrain emotional asymmetry (right AIC ↔ withdrawal/arousal/individual-survival emotions like sadness), though the paper does not frame it that way and the design cannot test the asymmetry claim. Second, the finding partially closes a gap the wiki keeps flagging in the other Farb papers (farb-2011-relapse-prediction, farb-2022-relapse-biomarkers), whose protective “sensory” pole is exteroceptive (visual, somatosensory) rather than interoceptive; here it is the insula itself — at the cost of a cross-sectional rather than prospective design.

The depressed insula’s abnormality is at rest, not in the stimulus response (Wiebking et al. 2010)

Farb 2010 above reads insula engagement under emotional challenge as protective. Wiebking et al. (2010), from Northoff’s “material me” programme, look instead at the insula’s baseline in depression, and their central result is a methodological warning this page should carry into every insula-in-depression claim. Comparing 17 medicated MDD patients with 17 controls across heartbeat-counting (interoception), tone-counting (exteroception) and rest, they found: the interoceptive signal did not differ between groups; the exteroceptive signal appeared to differ (reduced deactivation, even positive BOLD, in anterior insula); but that exteroceptive difference vanished under baseline correction and reappeared as a difference in the preceding rest period — the depressed insula under-deactivates at rest (reduced negative BOLD in left anterior/middle and right middle insula), which inflates whatever stimulus response is measured against it.

So “reduced insula reactivity in depression” — the summary this page carries from Quadt et al. and converges on in Nord et al. — may in part be a rest-baseline artefact wherever the source design could not separate rest from stimulus. The other durable result is a decoupling: in controls, right-anterior-insula rest activity correlated with body-perception (BPQ) scores (r≈0.5–0.56); in patients that correlation was absent, while reduced left-anterior-insula rest deactivation instead tracked depression severity (BDI r=0.57). The felt body and its interoceptive cortex come apart in depression. Held at the interoceptive-imaging discount plus three of its own — all patients medicated, “rest” only 4–10 s fixation epochs (not a resting-state design), and the “material me” a frame rather than a measured self. See embodied-selfhood, interoceptive-psychopathology.

The insula turned outward: empathy, and its training (Lutz et al. 2008)

The Farb result is the insula representing one’s own state under challenge. Lutz et al. (2008) is the insula representing someone else’s — and being trained to do it more. On the perception-action model of empathy the paper adopts, feeling with another recruits the affective (not sensory) components of one’s own experience: anterior insula and ACC (Singer et al. 2004). So the same cortex Craig places atop the interoceptive hierarchy is where another person’s distress gets re-represented as a felt state.

Generating a compassion state up-regulated that circuitry. Long-term meditators, more than novices, increased their insula/SII response to emotional human vocalizations — a follow-up isolated a right-insula cluster driven by experts responding more strongly to a crying woman than a laughing baby during meditation vs. rest. The insula signal also tracked the practitioner’s verbally-reported quality of the meditation (stronger in “good” than “poor” blocks), tying it to the felt success of the state rather than the stimulus alone. Critically, the effect survived regressing out pupil-indexed autonomic arousal (ANCOVA F(1,27)=20.2, p<.0005), so the insula is doing something beyond registering generalized arousal — consistent with a specifically empathic representation.

Read alongside Farb 2010, the two contemplative studies make a matched pair: mindfulness keeps the insula online to self-directed sadness; compassion up-regulates it to other-directed distress. In both, a contemplative intervention increases interoceptive-cortex engagement under emotional challenge — the “more insula is better” assumption in its most sympathetic form (more insula, more empathy).

The insula as the experiential organ of illness (Quadt et al. 2018)

Most of this page’s disease-adjacent evidence is about emotion and self. Quadt, Critchley & Garfinkel (2018) add the insula’s role in the felt experience of being physically ill — the sickness-behaviour literature. After an inflammatory challenge (typhoid vaccine, endotoxin), the insula carries the subjective side of the response: right anterior insula metabolism tracks loss of interest in social interaction, AI–middle-cingulate connectivity predicts subjective malaise and discomfort, and mid-insula reactivity is implicated in fatigue. The same cortex that re-represents cardiac and homeostatic state also renders inflammation as felt malaise — inflammation entering consciousness through the interoceptive hierarchy. Across the review’s disease sections the insula recurs as the shared node: reduced dorsal-mid-insula reactivity in depression, reduced AI activation during heartbeat/stomach attention in anorexia, altered insula reactivity and connectivity in autism. See interoceptive-psychopathology.

What lesions to it take away (Bonaz et al. 2021)

Nearly all the evidence above is correlational — activation, connectivity, volume. Bonaz et al. (2021) assemble the lesion literature, which is the closest this wiki comes to causal evidence about the insula, and it dissociates things the imaging cannot.

Viscerosensory and somatosensory heartbeat experience come apart. Patients with bilateral ischemic damage to insula and ACC may retain residual somatosensory but not viscerosensory experience of their own heartbeats, tested under pharmacologically induced tachycardia (García-Cordero et al. 2016; Khalsa et al. 2009’s isoproterenol method). “Feeling your heart” is at least two capacities, and this cortex carries only one of them. That is a constraint on every heartbeat-task interpretation on the wiki — see is-the-heartbeat-counting-task-valid.

Interoception, emotion recognition and theory of mind fall together. After fronto-insular-temporal ischemic stroke, deficits in heartbeat detection are linked to deficits in emotion recognition and mentalizing — read by the authors as confirming that network’s role in integrating interoceptive information into emotional and social representation. Compare empathy and the Lutz result above, which show the same network engaged; the lesion data show it required.

Focal insular lesions produce specific motivational and affective losses: disgust insensitivity, acquired alexithymia, psychopathy, and loss of drug-craving, alongside case reports of central pain and altered body awareness. The craving finding is the sharpest — damage this cortex and a specific motivational feeling disappears while the rest of motivational life continues, which is unusually direct evidence that a felt bodily state is constitutive of a drive rather than accompanying it. See craving, where-are-feelings-constituted.

Two more from the same review. In Alzheimer’s disease, insular and cingulate degeneration accompanies a four-way cardiac-interoceptive deficit — accuracy, heartbeat-evoked potential modulation, interoceptive learning, and metacognitive awareness — one of the few places all four dimensions of the taxonomy are measured in one sample. And in behavioural-variant frontotemporal dementia, frontal-predominant degeneration hits anterior insula and its interactions with cingulate, medial frontal, frontopolar, orbitofrontal and basal ganglia, compromising both perceptual sensitivity and metacognitive insight. Across neurodegenerative disorders, loss of self-awareness tracks damage to cingulate, prefrontal and fronto-temporal cortices with strong connectivity to the insula — which is Craig’s awareness claim meeting degenerative evidence, and about as close to a test of it as the wiki holds.

Note the disgust item cuts against the Lindquist reframe above only apparently: selectivity without specificity is compatible with lesions producing a disgust deficit among several others, which is exactly the reported pattern. See disgust, where the primary cases are now held.

The lesion literature read first-hand, and it does not confirm the model (Jones et al. 2010)

The section above is a clause in a clinical review. Jones, Ward & Critchley (2010) is the dedicated survey, eleven years earlier, written specifically to appraise models of insula function built from neuroimaging — this page’s models — against the neuropsychology. Four things it changes here.

1. The gradient survives, in a form. Anterior-insula lesions disrupt affective and language functions (amusia — loss of the emotional content of music; aphasia; apraxia of speech; especially left). Posterior-insula lesions disrupt the representation of interoceptive information — pain, temperature, tactile perception. That is roughly the posterior-primary / anterior-integrative split at the top of this page, arriving from a causal design rather than from cytoarchitecture.

2. But awareness of the body localizes posteriorly, not anteriorly. This is the direct collision. On Craig’s account awareness is the AIC’s business; Jones et al. record that “some previous reviews imply” anosognosia for hemiplegia and hemianaesthesia are disorders of emotional awareness reflecting anterior damage — and report that the evidence puts them at the posterior insula (Karnath et al. 2005, lesion overlay; Cereda et al. 2002, a focal right posterior lesion sufficient for somatoparaphrenia). See anosognosia, where the three available readings are set out; the authors’ own is that these are failures of a comparator losing its afferent term rather than failures of feeling.

3. The insula is not required for pain awareness. Two patients with large left-MCA strokes involving insula rated noxious stimuli on the affected side as significantly more intense than controls, with no insula activation in either hemisphere, while their unpleasantness ratings were normal — dissociating pain’s two dimensions (Starr et al. 2009). The authors’ conclusion: “subjective awareness of noxious stimuli involves multiple, distinct patterns of brain activity where insular cortex is not a prerequisite.” Set against posterior-insula lesions elsewhere raising pain thresholds or producing pain asymbolia, the lesion record on pain is inconsistent in direction. See chronic-pain.

4. And it may be dispensable. Duffau et al.’s (2006) 42 insular low-grade glioma resections produced substantial immediate deficits — 21 transient hemiparesis, 10 articulatory disorders, 7 with abulia — and at three months, no visceral, cardiovascular, sensorimotor, gustatory, auditory-vestibular or language disorders in any patient. Whatever this cortex does, other tissue largely does it within a quarter. That is the single most deflationary fact on this page and it comes from the review most committed to the structure’s importance.

The methodological warning that governs all of it. Isolated insular infarcts are uncommon; the insula’s supply is the middle cerebral artery, whose territory includes internal capsule and basal ganglia. Hillis et al. (2004) used diffusion/perfusion imaging to show that apraxia of speech, localized to left anterior insula in Nature (Dronkers 1996) and cited for eight years, actually correlates with damage to Broca’s area — the insula being “most commonly damaged because it is most vulnerable to disruption of the middle cerebral artery.” A canonical insular localization produced by vascular anatomy rather than function. See lesion-symptom-mapping.

The review’s verdict, from a co-author who builds these models: “models of insula function fall short of adequately explaining in detail the range of neuropsychological effects reviewed above,” and specifically, how interoceptive signals are perceived, how they become feeling, and how predicted/actual mismatch is detected “is less obvious in the presence of marked insula damage.” Craig’s global emotional moment is presented as the leading unitary attempt and is not endorsed.

Worth recording for feedforward-vs-predictive-interoception: three years before Seth (2013), Critchley is already describing this cortex as a comparator detecting incongruence between predicted and actual emotional states, with anterior insula storing “an error representation when the states are mismatched” (Preuschoff et al. 2008; Singer, Critchley & Preuschoff 2009), demonstrated via false feedback (Gray et al. 2007). Prediction-error language without the generative model — nearer Craig’s ascending comparator than Seth’s descending one.

What insula lesions take from evaluation: both valence and arousal (Berntson et al. 2010)

The Jones section deflates the neuroimaging models; Berntson, Norman, Bechara, Tranel & Cacioppo (2010) adds a lesion result that is positive for this cortex’s evaluative role. Seven patients with unilateral strokes destroying >50% of the insula rated IAPS pictures on valence and arousal (affective-picture-viewing). Against control-lesion patients they showed attenuated arousal for both pleasant and unpleasant extremes (neutral spared) and attenuated valence — smaller increments in positivity as pleasant grew more pleasant, in negativity as unpleasant grew more unpleasant. The loss spans both dimensions and both hedonic signs; their neutral ratings and their loneliness/depression self-ratings were normal, so it is a failure to register affective content, not a rating bias.

This is causal support for the core-affect reading at the top of this page: the insula “broadly involved in the recognition, processing, and assignment of evaluative valence,” and contributing to arousal — arrived at by subtraction rather than activation. It also refines the Haruki deflation: where imaging suggests the right anterior insula tracks arousal specifically, removing the insula costs valence too, so the structure’s evaluative contribution is not arousal-only. The same paper’s amygdala group loses only negative arousal (valence intact) — a double dissociation of the two core-affect dimensions across the two structures; see amygdala. Held at the lesion discount: n=7, large MCA-territory strokes, unilateral and L/R-mixed, so the design cannot test the anterior/posterior gradient or Craig’s laterality.

Interoceptive-awareness differences track right-insula activation, to aversive material (Pollatos et al. 2007)

The Berntson result above is lesion evidence that the insula is required for evaluating IAPS pictures. Pollatos, Gramann & Schandry (2007) add the individual-differences complement, and it is now held first-hand rather than as the bare citation this page carries in the Haruki section below. Sixteen high vs sixteen low heartbeat-perception subjects viewed the same picture set under 62-channel EEG, and the P300 was localized by distributed source reconstruction (stCDR) to five generators. High perceivers showed greater activation in right insula (BA 13), anterior cingulate, somatosensory and prefrontal cortex — but not in the visual generator, the one cluster with no place in the feeling circuit. Crucially, the right-insula enhancement appeared only for unpleasant pictures (F(1,30) = 8.61; pleasant and neutral ns), while the somatosensory enhancement was emotion-general. So within one dataset the interoceptive-cortex difference is affect-gated and specifically negative, congruent with the right-insula-and-aversive-state reading this page carries from Craig and Critchley, and with Berntson’s broadly-evaluative lesion result.

Two discounts. This is EEG-inverse anatomy over a template brain — an ill-posed, low-resolution reconstruction, so “the right insula (BA 13)” is a current-estimate centroid, not an fMRI-grade localization. And, as everywhere in this Munich programme, no bodily response to the pictures was measured, so a sympathetic-tone account (“good perceivers’ hearts are louder, their cortices more responsive”) predicts the whole result set, source clusters included, with no perception in it. See pollatos-2007-neural-systems-feelings and heartbeat-detection-task. This is also the paper cited (with Critchley et al. 2004 and Caseras et al. 2013) as the heartbeat-accuracy correlation the Haruki & Ogawa deflation below narrows to the cardiac channel.

The right anterior insula as an early site of degeneration (Farb et al. 2012)

The lesion sections above ask what removing the insula takes away. Farb et al. (2012) show the early signature of its degeneration in frontotemporal dementia, and it is convergent across three independent local metrics: in both bvFTD and semantic dementia the right anterior insula showed reduced low-frequency power (fALFF), reduced regional homogeneity (REHO), and loss of its normal resting-state connectivity with prefrontal cortex. Notably the whole-brain ICA network analysis did not flag this region — only the local-signal metrics caught it — which the authors offer as evidence that the right anterior insula is an early, sensitive marker of dementia pathophysiology that network-decomposition methods can miss.

This fits the wiki’s picture of the anterior insula as the apex of the salience hierarchy: it is where the failure shows up first and where its loss is read as decoupling affective/visceral information from behaviour. The paper’s framing — the anterior insula “integrates emotional and visceral information into representations of present moment context that guide socially appropriate behavior” (citing Farb 2007, Seeley 2007, Craig 2009) — makes FTD a natural disconnection experiment: sever this integration and the behavioural cost is apathy (with prefrontal isolation) and, where the limbic channel is lost, an inability to produce disinhibition. See salience-network for the network-level reading.

The posterior insula as a respiratory-rate sensor, and as a training target (Farb 2013, ×2)

The gradient at the top of this page is usually stated in terms of cardiac and thermal afference. Weng et al. (2021) describe the posterior insula as “sensitive to the respiratory rate and putatively considered primary interoceptive cortex” — the same primary-cortex role, indexed to breathing. See respiratory-interoception.

Both primary sources are now held first-hand, and reading them corrects a conflation the secondhand note had inherited — there are two Farb 2013 papers, and the Weng-sourced summary had merged an attention effect and a training effect that belong to different ones.

What the [[farb-2013-attentional-modulation|Cerebral Cortex attention paper]] shows (a within-subject IA-vs-EA contrast, MBSR a covariate):

• During interoceptive attention to the breath, respiratory rate is tracked by the posterior insula (IA r=0.63); during exteroceptive attention the same respiratory signal is tracked by somatosensory cortex instead (posterior-insula coupling drops to r=0.23). Attention selects which cortex represents the signal — a within-subject gain effect, not a between-groups training increase.
• Interoceptive attention enhanced insula connectivity with the posterior ventromedial thalamus — the VMpo relay — relative to exteroceptive attention.
• The insula shows a graded posterior→anterior attentional tuning: posterior gyri tuned to interoception, anterior gyri to exteroception, so the anterior insula is “not an area of pure body awareness” (see the Haruki & Ogawa section below, which reaches the same deflation).

What the [[farb-2013-training-interoceptive-cortex|SCAN training paper]] shows (a between-groups MT-vs-untrained contrast — the “mindfulness training raised posterior insula” claim the wiki inherited, corrected on two points now it is first-hand):

• The group-level training effect is in the anterior, not posterior, insula: MT graduates responded more to interoceptive attention in the dorsal anterior insula (accessory and anterior short gyri), read as integrating body sensation with external context.
• The posterior insula moved with practice dose, not group membership: within the MT group, % of prescribed daily practice completed predicted greater posterior-insula IA activation and interoceptive bias (r=0.61). So “training raised posterior insula” is really “practice raised posterior insula.”
• The training paper’s connectivity story is about DMPFC (deactivated during IA in MT, negatively coupled to posterior insula) and the putamen, not the VMpo thalamus — the thalamic-relay finding above is the attention paper’s, and attributing it to training was the conflation.

The thalamic result is the more consequential for this page. Most of the wiki’s contemplative evidence concerns anterior insula and its cortical partners — re-representation, appraisal, the global emotional moment. An attentional effect on connectivity with the thalamic relay is a claim about the afferent pipeline itself, low in the hierarchy, which is where Craig’s account puts the primary interoceptive map. Whether that is best read as improved signal transmission or as top-down precision-weighting of a low-level channel is exactly the feedforward-vs-predictive-interoception question, and one connectivity result does not settle it.

Weng et al. also give the anterior insula a network job — hub shared between the interoception and salience networks, and the switch that meditation is proposed to exercise. See that page.

Recorded directly: the insula’s band changes with attentional direction (García-Cordero et al. 2017)

The Farb 2013 attention effect above is fMRI — it shows which cortex represents a signal as attention shifts, at the temporal resolution of the BOLD response. García-Cordero et al. (2017) put depth electrodes in the posterior insula (among amygdala, somatosensory cortex, and inferior frontal gyrus) in two epileptic patients and recorded the same interoceptive-vs-exteroceptive attention contrast at millisecond resolution. In all four regions, basal interoception showed greater broadband high-frequency power (35-110 Hz) and exteroception greater low-frequency power (1-35 Hz). High-frequency broadband activity indexes internal, effortful, memory- and consciousness-related processing; lower bands, exteroceptive sensory processing.

So the posterior insula does not merely light up under interoceptive attention — the spectral character of its activity flips with the direction of attention, recorded from inside the human insula rather than inferred from scalp or BOLD. This is convergent, causal-adjacent evidence for the attention-selects-representation picture the Farb section builds, and it is the wiki’s only intracranial recording from human insula distinguishing an interoceptive from an exteroceptive task. The scalp complement — a frontal HEP more negative for interoception than exteroception at 200-500 ms — is on garcia-cordero-2017-attention-in-and-out. The usual caution holds doubly here: n = 2 epileptic patients, leaned on for cross-patient consistency rather than for sample size.

The insula is itself body-mapped — the insula homunculus (via Zeharia et al. 2019)

The posterior→anterior gradient this page is built on is organizational (primary map → re-representation). Orthogonal to it, the interoceptive cortex is also somatotopically organized by body region: the same group behind the new precuneus homunculus had earlier reported an insula homunculus (Zeharia et al. 2012) — a topographic body layout within the insula. Zeharia, Hofstetter et al. (2019) revisit this in their discussion, proposing the insula and precuneus homunculi as a putative “node linking body and mind,” and (in the new data) find that the anterior, leg-related voxels of the precuneus body map are functionally connected to the insula. This is the one place somatotopy and interoception genuinely meet in the wiki: the interoceptive cortex is not a featureless “feeling” map but is itself laid out by body part. Recorded as the authors’ framing — the 2019 connectivity links precuneus-leg voxels to the insula, but the study measures no interoceptive function, and the “body-mind node” is explicitly a conjecture for future work.

The wiki now holds the same claim demonstrated directly, and for a lamina I modality. Brooks et al. (2005) — with Craig co-authoring — mapped painful heat to face, hand and foot at high-resolution 3T and found an orderly body map in the dorsal posterior insula (face anterior/lateral, foot medial in the circular sulcus, hand lateral), the only one of five operculo-insular regions to carry a stimulus-site-dependent map — SII carried none. The coordinates match human depth-electrode stimulation and recording. So the primary-interoceptive-cortex role assigned to the dpIns at the top of this page comes with an internal body layout, exactly as the VMpo→dpIns labelled-line account predicts — a measured somatotopy of a homeostatic afferent, not (as with the 2012/2019 insula-homunculus citations) an inferred one. Note this cuts a little against the “posterior insula did nothing in an attention paradigm” reading elsewhere on this page (Haruki): with an actual graded noxious stimulus the dpIns is highly responsive and finely organized — the difference being that Brooks perturbed the body while attention paradigms do not.

The mouse insula predicts (Berntson & Khalsa 2021)

Almost everything above is human, and the predictive material is human imaging or cytoarchitecture. Berntson & Khalsa (2021) supply the wiki’s first animal evidence for anticipatory interoceptive representation, and it is in this cortex.

• Aversive state processing is localized to posterior insular cortex in mice, where it can flexibly shift adaptive feeding behaviour (Gehrlach et al. 2019) — a region with extensive top-down and bottom-up connectivity to other cortical and subcortical areas (Gehrlach et al. 2020).
• Hunger/thirst-dependent activity patterns in mouse insula shift in anticipation of expected satiety of food or water signals (Livneh et al. 2020), which the review reads as showing “that mice can generate and modify interoceptive predictions in response to ongoing contextual changes in the external environment.”

Two consequences. For feedforward-vs-predictive-interoception, this is the first evidence on that page that is neither human imaging, cytoarchitecture, nor theory: an anticipatory interoceptive representation in an animal that can be lesioned, silenced and recorded. For can-we-know-animal-feelings, it tightens the screw — the same review states there are no animal models of interoceptive awareness, so we now have a species that demonstrably updates predictions about its internal state and no way to ask whether anything is felt.

Note the anatomical wrinkle: the mouse result is posterior insula for aversive processing, where the human gradient assigns posterior insula the primary-mapping role and puts affect and re-representation anteriorly. Whether that is a species difference, a granularity difference, or the two literatures using “posterior” differently is not settled by anything the wiki holds — and Craig’s own claim that the human AIC “has no equivalent in the monkey” makes cross-species insular inference hazardous in both directions.

How much of this transfers across species, itemized (Chen et al. 2021)

The wrinkle above is the local case of a problem Chen et al. (2021) state as a list. Their verdict is blunt: insular “properties and cell types appear to differ substantially across species, limiting the applicability of some findings from rodents to human interoceptive health.”

What differs, specifically:

• Von Economo and fork neurons sit in AIC and ACC of macaques, great apes and humans, plus elephants and whales — a distribution that includes macaques, correcting the claim this wiki inherited from Craig. See von-economo-neurons, where the correction is recorded.
• The salience network splits. Its ventral division, for arousal-based affective experience, is largely homologous between monkeys and humans; the dorsal division for attentional control is “much more developed in humans.” The salience-network page has been treating one network as one thing.
• Amygdalar organization differs markedly between rodents and monkeys, and the insula receives major amygdalar input — so a rodent insula is receiving from a differently-organized structure.

What is consistent across species is worth stating with equal precision, because it is the part the wiki can lean on: the posterior-to-anterior topography of the insular interoceptive map holds “in mammals ranging from rodents to humans.” The gradient transfers; the cell types and the network context do not.

They add an open question the wiki had not recorded: given the different cytoarchitectonic regions of insular cortex, is there a single representation of the body in the insula, or multiple overlapping maps? That sits directly beside the insula homunculus material above and is not answered by it.

The right anterior insula may be cardiac rather than interoceptive (Haruki & Ogawa 2023)

Almost everything on this page above treats the right AIC as the seat of interoceptive awareness as such. Haruki & Ogawa (2023) put a second organ next to the heart in the same scanner session and the region went with the heart.

Thirty-one participants attended, in ten-second blocks, to their heartbeat, to their stomach, or to a fading visual target (interoceptive-attention-task). Contrasted against the visual control, cardiac and gastric attention produced the same map — insula, frontal and parietal operculum, middle cingulate, SMA. Contrasted against each other, the right dorsal anterior insula extending into frontal operculum was the only region preferring cardiac (48 voxels, t = 4.65), with the mean-signal analysis agreeing (right ASG p = 0.021, right MSG p = 0.016).

Three reasons to take the deflation seriously rather than as one small cluster:

1. The supporting evidence for the general claim is itself almost all cardiac. The correlations that established this region — Critchley et al. (2004), Pollatos et al. (2007), Caseras et al. (2013) — are heartbeat-accuracy correlations, and Schulz’s (2016) meta-analysis is explicitly of heart-focused interoception. The generalization to “interoceptive awareness” was never tested against another organ.
2. Where it has been tested on other channels, it fails. The accuracy-to-right-AIC relationship does not appear for awareness of breathing (Wang et al. 2019) or of skin conductance (Baltazar et al. 2021).
3. There is a functional explanation ready. Cardiac awareness is how arousal is noticed; the right insula is where arousal is made. Resection diminishes physiological and emotional arousal (Terasawa et al. 2021; Holtmann et al. 2022), and false accelerated cardiac feedback both raises perceived arousal and activates right AI (Gray et al. 2007; Kleint et al. 2015 — see false-feedback-paradigm). On that reading the right AI is not the general seat of feeling the body; it is the seat of feeling aroused, and the heart is the organ that reports arousal.

This does not refute Craig, who argued from convergence across many non-interoceptive domains (pain, agency, time, error awareness) and never claimed cardiac specificity. It narrows the interoceptive leg of his argument to one channel. See is-interoception-domain-general, where this is the neural counterpart of Banellis et al.’s behavioural decorrelation.

And a complication the same study reports without pursuing. Comparing interoceptive attention (both organs averaged) against the exteroceptive control, the right ASG was more activated by the visual task (F(1,30) = 10.28, p = 0.003); the interoception-preferring subdivisions were left MSG and PSG. So the dorsal anterior insula prefers the heart within interoception while preferring a demanding visual task to interoception altogether — which fits the salience-network and attentional-switching readings of this cortex better than it fits a dedicated interoceptive one.

Organ identity is coded in the middle insula

The same study ran MVPA over twelve anatomical insula subdivisions (Faillenot et al. 2017 atlas). Only the left posterior short gyrus — the dorsal middle insula — classified cardiac from gastric attention above chance (56.32%, d = 0.61, FDR p = 0.024), with left/right ASG and left MSG marginal (ps = 0.089). Ventral and posterior subdivisions did not classify at all.

So a region that shows no mean-activation preference for either channel nonetheless carries which organ is being attended to in its multivoxel pattern. Two consequences for this page:

• It gives the first direct evidence on the question Chen et al. (2021) raised and left open above — one body map in the insula, or several overlapping ones? The mid-insula behaves like a place where organ identity is still distinguishable, which the single-map picture does not predict.
• It puts viscerotopic specificity at the integrative stage of the posterior→mid→anterior gradient rather than the primary one, and asks that stage to preserve information the usual reading of the gradient has it discarding on the way to a unified anterior feeling.

Hold the effect size in view: 56% against 50%. The pattern is there and it is small.

The posterior insula, meanwhile, did nothing — neither classifying the conditions nor activating above rest. The authors read this as support for the primary-cortex role assigned at the top of this page: posterior insula codes ongoing change in the bodily signal, and an attention paradigm perturbs nothing, so there is no change to code. Coherent, and an argument from a null in a study that recorded no physiology and so cannot confirm that nothing changed. See feedforward-vs-predictive-interoception.

The insula holds both sides of the update equation (Harrison et al. 2021)

The EPIC section above splits this cortex by granularity and assigns the halves opposite computational roles: granular mid/posterior insula returns ascending prediction error, agranular anterior insula issues descending prediction. Seth, Barrett & Simmons and Stephan et al. all propose some version of that anterior/posterior split.

Harrison et al. (2021) is the first study in this wiki able to test it, because it is the first to have separate, trial-by-trial estimates of prediction and prediction error to regress activity against (BLT, 7T, reduced FOV over insula and midbrain). The test failed:

• Anterior insula deactivated with prediction certainty — activity falling as predictions grew more confident, alongside dlPFC, ACC and MFG.
• Anterior insula activated with prediction error magnitude — alongside ACC, MFG and the PAG.
• The authors state it plainly: prediction and prediction-error activity “did not appear to be dissociated between anterior and posterior insula cortices, as has been previously hypothesized.”

So one region carried both terms, with opposite signs. The authors’ reading is favourable to the framework rather than against it — greater certainty reduces belief updating and greater error increases it, so a region whose job is representing and updating a model of the body should track both. That is a coherent gloss, and it is also the reading that makes the anatomical prediction unfalsifiable by this design.

The one crumb for the split is a valence effect: unexpectedly receiving a resistance activated left posterior insula more than unexpectedly escaping one, which the authors suggest may mean homeostatically threatening inputs are enhanced in primary interoceptive cortex. A small effect, in the direction the theory wants, on the one contrast where it appeared.

Two limits before this is treated as a refutation. The functional slab was a reduced field of view, and layer-resolved imaging would be required to identify granular versus agranular contributions — the authors say the resolution is not there. And this is respiratory resistance, a single channel and a single feature; the EPIC anatomy is stated about interoception generally.

And anxiety changed the prediction side only. The one group effect in the whole imaging analysis was a valence × anxiety-group interaction in bilateral anterior insula for prediction certainty: low-anxiety participants deactivated aIns more when confidently predicting an upcoming resistance, moderate-anxiety participants the reverse. Nothing in prediction error. Given that the wiki’s clinical story (computational-psychiatry) locates psychopathology in elevated prediction error, that asymmetry is the study’s most consequential result for this page’s disease sections.

The mid-insula as a transdiagnostic locus of disruption (Nord et al. 2021)

The Haruki MVPA above put organ-identity coding in the left dorsal middle insula in healthy participants. Nord, Lawson & Dalgleish (2021) arrive at almost the same coordinate from the opposite direction — a preregistered ALE meta-analysis of 33 neuroimaging experiments (626 patients, 610 controls) across depression, bipolar disorder, anxiety, anorexia and schizophrenia, asking where disrupted interoceptive activation converges. It converges on a single cluster: the left dorsal mid-insula (peak MNI −36, −2, 14). Two independent methods — a within-subject organ-decoding study and a transdiagnostic disruption meta-analysis — both single out the left mid-insula as the interoceptively load-bearing subdivision, and both put the action at the integrative stage of this page’s gradient rather than the anterior seat of feeling.

Two conjunction results sharpen where on the gradient it sits. Against a large core-affect database the mid-insula cluster showed no overlap; core affect instead peaked in a huge left anterior insula cluster. So an independent meta-analysis reproduces the anterior=affective / mid=integrative split this page is built on — and it fits the Haruki deflation that the anterior insula (especially right) may be about arousal/affect rather than the body in general. And against the neural targets of antidepressants and psychotherapy the cluster again showed no overlap, which the authors turn into a claim that the mid-insula is an untouched treatment target.

Nord et al. read the dysgranular precentral insular gyrus — hybrid connectivity to both anterior and posterior insula — as the anatomical candidate for encoding interoceptive prediction errors, i.e. the comparator between descending prediction (agranular anterior) and ascending signal (granular posterior). This is a different placement from Harrison et al. above, who found the anterior insula carrying both prediction and error terms with no anterior/posterior dissociation. Nord (anatomy + convergence, no model) and Harrison (trial-by-trial model, no meta-analysis) are complementary and mutually under-determined bets about where interoceptive prediction error lives in this cortex — mid- versus anterior. See feedforward-vs-predictive-interoception, interoceptive-psychopathology. Held with the meta-analytic brakes: coordinate-based (no effect sizes), a surprising left laterality possibly driven by verbal probes, and pooling of interoceptive channels the paper concedes may not be integrated.

The anterior insula gates exteroceptive awareness by cardiac synchrony (Salomon et al. 2016)

Almost every result on this page has the insula representing the body — cardiac, gastric, thermal, nociceptive afference, or the felt self. Salomon et al. (2016) put the anterior insula between the body and the outside world: a visual target flashed in time with the viewer’s own heartbeat takes longer to break continuous flash suppression into awareness (and is discriminated less accurately under crowding) than the same target flashed at a slightly different frequency. At 7T, the bilateral anterior insula shows reduced BOLD to cardiac-synchronous visual stimuli — for both consciously seen targets and targets rendered fully invisible by CFS — while ACC, rSTG and occipital control regions do not, and the effect survives RETROICOR correction (so it is neuronal, not pulsation).

Three things this adds to the page:

1. It extends the AIC’s multimodal-integration role past self-awareness. The wiki reads the anterior insula as the site where interoceptive and exteroceptive signals converge (Craig, Seth, the EPIC model); here that convergence is caught shaping what reaches consciousness of the external world, not just what the body feels like. The authors frame it as predictive suppression of the heartbeat’s self-generated sensory consequences — though they concede a non-predictive temporal-disparity account fits equally.
2. It is a cardiac-frequency effect, cleanly dissociated from cardiac phase. Phase-shifting the target by half a cycle left the suppression unchanged, so this is not a diastole pulsed-inhibition effect — a different axis of cardio-visual coupling, and one that resolves the “vision shows mostly null cardiac-phase effects” observation by looking at frequency instead.
3. It is bilateral. Against the Haruki and Craig right-AIC-as-arousal readings, the cardio-visual suppression appears in both anterior insulae with no lateralization (F(1,7)=0.008, n.s. left vs right), so whatever cross-domain-synchrony job this is, it is not right-lateralized the way the arousal story predicts.

Held at the usual discounts: small behavioural effects (~0.2 s time-to-emergence; d’=0.38–0.65), small fMRI n (7 and 8 after exclusions), and a mechanism that is argued rather than tested. See feedforward-vs-predictive-interoception, is-interoception-domain-general.

Shaped by early experience

Craig’s account is anatomical and phylogenetic; it says little about how the structure gets built in a given person. Oldroyd et al. (2019) assemble the evidence that early social experience leaves marks on exactly this cortex, and hang a developmental thesis on it. Their starting observation is that the insula shows protracted post-natal development, leaving a long window for environmental input.

The cited findings: children classified as anxiously or avoidantly attached show markedly lower insular volume and smaller surface area than controls (Kühn & Gallinat 2013; Sheffield et al. 2013; Lim et al. 2014), and adults with an avoidant attachment style show decreased insular activation to stimuli relative to securely attached individuals (DeWall et al. 2011). Childhood maltreatment also alters insular network centrality (Teicher et al. 2014).

Treat this as suggestive rather than settled: the structural studies are drawn from maltreatment, PTSD and psychosis samples rather than from normative variation in attachment, so the inference from “insecure attachment” to “smaller insula” is looser than the summary implies, and none of it is longitudinal. But it is the strongest available anatomical support for social-origins-of-interoception — and it is what makes the social-vs-biological-origins-of-interoception debate a question about this structure rather than an abstract dispute.