Why do organisms do anything at all, and why do they feel anything at all? Where does the agency to act come from, and why does anything matter to a system? The Homeostasis Theory of Cognition and Consciousness (HTCC) proposes that homeostasis is the core principle behind why and how we observe, think, feel and act.
The current version of the theory (v0.2) starts with feelings, i.e. why some states feel good or bad, and builds cognition on top of that. Feelings are also what scaling up today’s AI models doesn’t give us (see the capability–sentience gap).
Why is pain bad? #
Illusionism is the view that the “feel” of conscious experience is a representation created by the brain’s self model, with no further fact behind it. A well known objection to illusionism (Kammerer, 2019) asks: if there is no phenomenal fact, why is pain bad?
HTCC’s answer is that the badness of pain was never a phenomenal fact to begin with. An organism has to keep itself within viable bounds (e.g. body temperature and blood sugar) to stay alive, and that sets up its own norm. Some states are actually bad for it. This “bad-for-it” is real, and it exists before any feeling is involved. We call it the constitutive viability relation. The felt quality of pain (the quale) is how the self model represents that relation.
So HTCC is illusionist about the feel, but realist about the stake. The badness is real. The feel is the self model’s representation of it.
The realist homeostatic theories of feeling (Damasio, Panksepp, Solms, Metzinger) also identify consciousness with homeostatic affect. HTCC agrees with that identification, but draws the opposite metaphysical conclusion.
Prior work and our contributions #
The idea that viability is where “mattering” comes from is not new. It comes from enactivism (Weber & Varela, 2002; Di Paolo, 2005), and we build on it. The regulation mechanisms we use (perceptual control theory, homeostatic reinforcement learning, active inference) are also existing work that we build on.
HTCC contributes 3 things:
- An answer to the normative challenge to illusionism. HTCC explains why experiences matter, and not only how they are structured, by locating the norm in the viability relation. The organism owns the stake through operational closure, and the self model is a transparent model of it.
- A falsifiable prediction. If cognitive and physiological drives compete in one common decision space, there should be a stable, state-modulated marginal rate of substitution (MRS) between them. Think of how long someone will keep a hand in ice water for different amounts of money, and how that trade-off shifts with their internal state. This can be measured with a cold-pressor experiment. Homeostatic value is also multidimensional, not a single “common currency”.
- A two-ingredient test for machine sentience (below).
What HTCC predicts about current AI #
Global Workspace theory explains how information becomes available across the brain, and IIT explains how it is integrated. HTCC adds the “stakes layer”, i.e. what makes any of that information matter. So HTCC complements these theories.
According to HTCC, felt valence requires both:
- an intrinsic, self-preserving stake, i.e. a real viability boundary the system can move towards or away from, and
- a self model that is its own implementation, i.e. a self that is re-instantiated by the very dynamics it reports on. This needs recurrent state that dominates the system’s processing.
An AI system can have a functional self that it can report on (see the Global Workspace evidence in the gap) and still have no self-preserving stake. Today’s models have stakes set from outside, e.g. the rewards they were trained on and the instructions in their prompt, but none of them is about keeping the model itself going. So today’s models can be capable, maybe even functionally self-aware, without being sentient.
HTCC v0.2 therefore disqualifies today’s models. Their stakes are set from outside, so they have no self-preserving stake. And they carry no internal state from one pass to the next. Their only recurrence goes through generated text, e.g. a chain of thought, and nobody has shown that this is enough for a self model that is its own implementation.
Status #
Manuscript v0.2, “Why Pain Is Bad: An Illusionist Homeostatic Theory of Valence and Consciousness”, is in preparation. It addresses two philosophical gaps: the normative bridge and owner-individuation. The main open item is real MRS data from the cold-pressor experiment.
Publications #
- Early exploration poster presentation: Kelkar, Amol. (2018). Homeostasis Theory of Consciousness (Solving the Hard Problem using physiological correlates of consciousness). 10.13140/RG.2.2.35033.83047.
- (2019-24) Evolutionary perspective on how homeostasis leads to cognition and consciousness
- (2018-24) Manuscript v0.1
- (2026) Manuscript v0.2, valence-first version (in preparation)
Keywords: homeostasis, valence, cognition, consciousness, qualia, illusionism, active inference, machine sentience