Interoceptive inference, emotion, and the embodied self (Seth 2013)
Seth’s landmark opinion piece extends the predictive-coding / Bayesian-brain framework — previously applied mostly to exteroception (vision, audition) and action — inward to interoception. The core claim, termed interoceptive-inference, is that emotions are not read out from bottom-up visceral signals but are actively inferred: the brain runs hierarchical generative models predicting the causes of interoceptive afferents, and subjective feeling states are the integrated content of those predictive representations.
Key move: generalizing appraisal theories
Seth situates his model as the predictive-coding successor to the lineage that runs from James–Lange through cognitive-appraisal theories (Schachter & Singer). Where appraisal theories say emotion = physiological change + cognitive interpretation, Seth reframes interpretation itself as top-down Bayesian inference on the causes of interoceptive signals. This dissolves the feed-forward assumption that had, per Seth, been imported uncritically from classical theories of perception (see feedforward-vs-predictive-interoception).
The anterior insular cortex as comparator
The AIC is proposed as the anatomical locus: structurally placed to detect and cause changes in physiological condition and to integrate interoceptive with exteroceptive signals; functionally, the site where predictions meet prediction errors. Supportive fMRI evidence: AIC encodes anticipated pain and pain prediction error, predicted risk and risk prediction error, and scales with subjectively reported itch and with feeling prediction errors in economic games. A posterior-to-anterior gradient maps primary interoceptive representations (posterior insula) onto subjective re-representations (AIC). The AIC’s density of von Economo neurons is noted as a possible substrate for rapid long-range integration.
Active inference and autonomic control
A distinctive contribution over earlier bottom-up models (Damasio, Craig, Critchley): interoceptive predictions can be discharged as reference points for autonomic reflexes (heart rate, respiration, smooth muscle) — the interoceptive parallel to how proprioceptive predictions enslave motor reflexes in predictive-coding accounts of motor control. This requires transiently low precision on interoceptive prediction errors (decreased attention to them), or the errors would revise predictions rather than drive action.
Body ownership evidence
Suzuki et al. (2013) and Aspell et al. (2013): synchronous cardio-visual feedback (a virtual hand/body changing colour in time with the heartbeat) increases ownership, measured by questionnaire and proprioceptive drift. Lower interoceptive-sensitivity predicts greater susceptibility to the rubber-hand-illusion (Tsakiris). These are read as multisensory predictive integration minimizing self-related precision-weighted prediction error — see experience-of-body-ownership.
Contradiction / distinction flagged
Seth explicitly distinguishes his view from Damasio, Craig, and Critchley: those frameworks ground selfhood in physiological representation but treat the interoceptive hierarchy as largely bottom-up. None identify emotion with top-down inference of interoceptive causes. This is the live theoretical fault line captured in feedforward-vs-predictive-interoception.