Interoceptive inference
A view that bodily feelings are not signals read straight off the body but predictions the brain makes about what the body should be reporting, corrected as real signals arrive. Feeling becomes a best guess that is checked and updated rather than a direct readout. It is set against the older picture in which sensations travel up from the body and are simply registered as experience.
The theoretical position, developed in Anil Seth's 2013 Trends in Cognitive Sciences paper and extended in Barrett and Simmons 2015 and Klaas Stephan's Zurich computational programme, that subjective feeling states are not direct readouts of body state but actively-inferred predictions generated by the brain about what its interoceptive afferents should be saying, with prediction errors driving learning and updating. The frame adapts Karl Friston's free-energy principle from the perceptual-inference domain to the body-state- monitoring domain.
Etymology§
The term joins interoception with inference in the technical Bayesian-brain sense developed by Karl Friston from the early 2000s on. Inference here is active prediction, not deduction — the brain is constantly generating expectations about its sensory inputs (including interoceptive ones) and updating those expectations against incoming signals. The phrase first appears as a unified construct in Seth 2013, though the underlying ideas were in the predictive- processing air for several years before.
Interoceptive inference inverts the directionality of Craig's classical model. Where Craig has homeostatic afferents flowing up through the spinothalamic tract to be re-represented in the right anterior insula as conscious feeling, the inference frame has the anterior insula generating top-down predictions about what those afferents should report, with prediction errors (mismatches between predicted and actual signals) driving learning. Feeling, in the inference frame, is the brain's best-guess about what the body is doing, constantly checked against incoming signal.
The inference frame is one of two main theoretical alternatives to Craig's direct-readout model — the other being Lisa Feldman Barrett's theory of constructed emotion, with which interoceptive inference is closely aligned (Barrett and Simmons's EPIC model is the architectural sibling of Seth's theoretical statement). Both are predictive-processing frames; both place interoception inside a wider Bayesian-brain framework; both have generated computational-modelling follow-up, particularly from Klaas Stephan's group at Zurich (Petzschner, Iglesias, Stephan and colleagues). The 2018 Khalsa Roadmap treats the inference frame as a legitimate alternative- to-Craig rather than declaring a winner.
The frame is theoretically elegant but empirically difficult: distinguishing top-down prediction from bottom-up readout requires careful experimental design, and many existing interoception findings are compatible with both readings. Anil Seth's group at Sussex and Stephan's at Zurich have produced specific computational and behavioural tests; the empirical literature continues to accumulate.