Homeostasis (and feelings as homeostatic sensations)
The organizing principle of Craig’s account of interoception: the interoceptive system is the afferent complement of the efferent autonomic nervous system — the long-missing sensory limb of homeostasis, carried by the lamina-i-spinothalamocortical-pathway. See craig-2002-interoception, ad-craig.
Feelings are homeostatic sensations
Craig’s central claim: every feeling from the body (temperature, pain, itch, hunger, thirst, air hunger, muscular ache, sensual touch) has both a sensory and an inseparable affective/motivational aspect, and that affect is keyed to homeostatic need. The same cool stimulus is pleasant when you are overheated and aversive when “chilled to the bone” — because the feeling signals a homeostatic role and motivates behavioural correction (thermoregulation → survival). Affect is not decoration on sensation; it is the homeostatic signal.
Why this reframes pain and emotion
- Pain becomes “both a sensation and a motivation” — a homeostatic imbalance signal, not merely tissue-damage detection. Central pain is reinterpreted as thermoregulatory distress.
- Emotion becomes grounded in homeostasis: the cortical image of body state (insular-cortex) is the substrate of subjective feeling, connecting to James–Lange and the somatic-marker-hypothesis, and to embodied-selfhood (“the material me”).
Homeostasis is already more than a thermostat (Petzschner et al. 2021)
The word carries a static connotation this page should not let stand. Petzschner et al. (2021) define homeostasis as the active processes keeping physiological states in survivable ranges, and note that the classical reactive framing has been superseded: homeostatic control “can also be predictive, by selecting actions that shift the state of an organism away from the setpoint in anticipation of a predicted future deviation.”
That gives the wiki a three-way distinction where it usually has two:
| set-point | when it acts | |
|---|---|---|
| reflexive homeostasis | fixed | after deviation (the baroreflex; comparator + hard-wired response) |
| predictive homeostasis | fixed | before deviation — deliberately leaving the set-point in anticipation (raising body temperature ahead of expected cold) |
| allostasis | moved | the set-point itself is adjusted to anticipated demand |
Evidence that human and animal homeostatic control is flexible in these ways is described as substantial: it adjusts to context, to expected future events, and — via placebo effects — to abstract beliefs. A regulated variable that tracks a belief is the strongest single argument on offer that homeostasis is inferential rather than mechanical.
Both of the anticipatory rows presuppose a model of how bodily states will unfold, which is the arc almost nobody formalizes — see forecasting and sensory-control-loop. And homeostatic-reinforcement-learning shows the whole column can in principle be delivered by reward machinery rather than inference, without any percept of the body.
How much can the periphery do alone? (Quigley et al. 2021)
A small disagreement the wiki should hold open, because two sources in the same special issue draw the line differently.
Quigley et al. (2021): homeostasis “can operate independent of predictive or anticipatory regulation by the brain, but can also be modulated by the brain when necessary.” Its job is local — fine-tuning peripheral functions, and correcting error “when predictive regulation fails to precisely match local needs.” Together with allostasis it lets the brain, “the body’s most metabolically expensive organ,” run near peak efficiency.
Bonaz et al. (2021): homeostatic reflex arcs are “enslaved” by descending allostatic policies from agranular visceromotor cortex, with set-points continuously written from above.
On the first reading the periphery has a default mode and the brain intervenes; on the second the periphery’s set-points are never its own. The mechanistic content may be nearly identical — both allow descending modulation, both allow local reflex — but the emphasis determines what counts as a failure. If homeostasis is autonomous-with-override, dysfunction can be peripheral. If it is enslaved, dysfunction is a policy error. Bonaz et al. need the second reading to license interoceptive dysfunction “at any level of the neuraxis”; Quigley et al. need the first to make homeostasis the error-correction backstop for imperfect anticipation.
Neither source engages the other. Recorded as a difference of framing, not a contradiction — and a reminder that “predictive homeostasis” in the table above sits precisely on the seam.
A third source lands on the seam: is the hypothalamus predictive? (Berntson & Khalsa 2021)
The disagreement above is about how much the periphery does alone. Berntson & Khalsa (2021) put the same question to the structure that actually issues the commands, and their answer is unusually well-scoped.
Hypothalamic control of homeostasis is classically viewed as reactive. However, there is growing appreciation that the hypothalamus can exert both reactive and predictive forms of control [Burdakov 2019, on orexin/hypocretin neurons], although this concept is mostly based on non-neuroendocrine functions.
And then the conditional that makes it worth recording: identifying a predictive hypothalamic controller of homeostatic endocrine function “would require rethinking of the impact of this brain region on allostatic processes.”
Three things follow.
It splits the predictive claim by output. The wiki’s predictive-regulation material — EPIC, allostasis, interoceptive active inference — treats anticipatory control as a general property of the visceromotor hierarchy. Berntson & Khalsa report it as established for some hypothalamic functions and not yet for the neuroendocrine ones, which are the slowest and the most obviously “set-point” shaped. The hpa-axis is exactly the case in question.
It is a real empirical stake in an argument that has mostly been framing. The Quigley/Bonaz difference above turns on emphasis with near-identical mechanics. This is a finding that could go either way and would change which of them is easier to hold: a predictive neuroendocrine controller supports the enslaved-reflex reading, its absence supports autonomous-with-override.
It is offered as an open question, not a result. Recorded in the middle row of the table above — “predictive homeostasis” — as the place the evidence is currently thinnest, and cross-listed at forecasting, since a predictive controller presupposes a forecast of how the state will evolve, which is the arc nobody has written down.
Relation to allostasis and predictive accounts
Craig’s homeostasis is largely a reactive/regulatory frame (feelings report deviations to be corrected). Compare allostasis (Sterling & Eyer; stability through anticipatory change), which Farb et al. and predictive accounts foreground — where the body is regulated toward predicted set-points, not just current deviations. Craig 2009 itself gestures toward prediction (the AIC represents “predictions of future feelings”), a bridge between his homeostatic hierarchy and the predictive-coding reframing in interoceptive-inference.