How do you feel — now? (Craig 2009)
The sequel to Craig (2002), extending the interoceptive hierarchy from the body all the way to awareness itself. See ad-craig. Where 2002 established the posterior insula as primary interoceptive cortex, 2009 argues the anterior insular cortex (AIC) — the top of the re-representation hierarchy — is the substrate of human awareness.
The argument from convergence
Craig catalogs imaging studies in which the AIC activates across wildly different tasks — interoception, pain, agency, self-recognition, music, emotion, risk, time perception, attention, decision-making, error awareness, the “feeling of knowing.” His inference: the one thing all these share is that they engage awareness. No other region is common to all. Therefore the AIC “engenders human awareness” and merits discussion as a neural correlate of consciousness. (The obvious counter: this is reverse inference — a caveat noted in the page’s limitations.)
The integration gradient
A posterior→mid→anterior progression: primary interoception (dorsal posterior insula) → mid-insula integrates emotionally salient environmental, hedonic (nucleus accumbens), and multisensory input → AIC produces the ultimate meta-representation. This maps onto the general front-of-cortex gradient of increasing complexity and the human expansion of the anterior insula. See insular-cortex.
The global emotional moment
Craig’s signature model: salience across all conditions is integrated into a unified global-emotional-moment — an image of “the material me” at one instant (“now”) near the AIC/frontal-operculum junction. A timebase-indexed sequence of these moments produces continuous, “cinemascopic” awareness across a finite present (a “meta-memory”), with storage buffers enabling comparison of past/present/future feelings — instantiating a reflexive “observer.” The model also predicts subjective time dilation (salient moments fill up faster) and frames music as an emergent property of awareness (rhythmic progression of emotional moments).
VENs, salience, and asymmetry
- von-economo-neurons (VENs): spindle cells in AIC + ACC, concentrated in hominoids (and elephants/whales), proposed as fast AIC↔ACC interconnections; their loss in frontotemporal dementia tracks loss of self/emotional awareness.
- salience-network (Seeley 2007): AIC + ACC + amygdala + hypothalamus; right AIC switches between executive-control and default-mode networks.
- Forebrain emotional asymmetry (Box 3): right AIC = sympathetic/arousing/withdrawal/survival emotions; left AIC = parasympathetic/affiliative/approach emotions (see the asymmetry section on insular-cortex).
How much of the interoceptive leg is cardiac? (Haruki & Ogawa 2023)
Craig’s argument for the AIC is an argument from convergence: it activates for pain, agency, self-recognition, music, risk, time, error awareness and interoception, and awareness is the only thing they share. That structure has a hidden dependency — each domain in the list must be contributing an independent instance.
The interoceptive one may be narrower than the list implies. Haruki & Ogawa (2023) ran cardiac and gastric attention against each other in the same subjects and found the right dorsal anterior insula was the only region preferring the heart; the accuracy-to-right-AIC correlation this literature rests on does not appear for breathing (Wang et al. 2019) or skin conductance (Baltazar et al. 2021); and the meta-analysis usually cited in support (Schulz 2016) is explicitly of heart-focused interoception.
So “interoception” enters Craig’s convergence list as one channel, not as a modality. That does not touch the other domains, and Craig never claimed cardiac specificity — but it means the interoceptive item is doing less independent work in the argument than its name suggests, and the anatomy at the top of insular-cortex is anchored more narrowly than the wiki had been treating it. The rival reading Haruki & Ogawa offer is that the right AI is where arousal is constituted, and the heart is simply the organ that reports arousal, which would make this region’s ubiquity a fact about arousal rather than about awareness.
Recorded as a qualification, not a refutation: the study is a single N = 31 attention paradigm with no ground truth, and it explicitly endorses Craig’s posterior→anterior gradient elsewhere in its own discussion.
The predictive footnote
Notably, Craig 2009 says the AIC represents predictions of future feelings and that anxiety/functional somatic disorders involve “distorted interoceptive predictions.” So Craig’s model contains a comparator and a predictive gesture — but the architecture remains an ascending integration/re-representation hierarchy with a comparator at the top, not a descending generative model. That distinction is exactly what Seth (2013) formalizes differently — see feedforward-vs-predictive-interoception.