Language of Stress and
Integrated Information Theory (IIT)
Critical examination of IIT’s claim that integration (Φ) determines consciousness. LoS shows integration is neither necessary nor sufficient. What matters: persistent self-model, variable rigidity, genuine stakes, unified value assessment.
Pace, J. C. (2026). The Language of Stress and Integrated Information Theory: Where we Converge and Diverge. FigShare. DOI: 10.6084/m9.figshare.31286344
Overview
Integrated information theory is the closest thing in contemporary consciousness research to a framework built the way this one is. Both start from properties experience must have and reason backward to what a system would need in order to have it. Both reject functionalism: behaving like a conscious system is not sufficient for either account. Both make an identity claim rather than a correlation claim, holding that experience is a certain kind of structure rather than being reliably produced by one.
That methodological agreement is where the similarity ends. Integrated information theory identifies experience with integrated cause-effect structure, which means certain simple, highly connected arrangements qualify while systems with rich behaviour and no integration do not. This framework identifies candidates for experience by whether anything is at stake for the system: whether it holds a self-model it defends, at intensities varying with what depends on them, across a value space in which competing demands can be weighed. The step from candidacy to experience is the identity inference, which the framework argues rather than demonstrates. The two criteria are different in kind, not only in content: one is a measure of integration, the other a question of what the system has riding on its own coherence.
The document takes integrated information theory in its current formulation rather than an older or simplified one, and is explicit that arguments aimed at the earlier version can miss. It also declines the easy objections. Thermostats and photodiode grids are not counterexamples, because the theory already assigns them zero or near-zero integrated information, and reaching for those illustrations misdescribes what it claims. The serious version of the objection comes from Scott Aaronson, who showed that certain regular constructions can be made to have arbitrarily large integrated information while being, by ordinary judgement, not the sort of thing that has experiences, and whose objection does not depend on misreading anything.
On current artificial systems the two frameworks point the same way by entirely different routes. Integrated information theory gives a verdict, and is the more restrictive of the two; this framework says only that their architecture gives no ground for candidacy. The document notes that convergence without treating it as support for either.
What separates them, ultimately, is what each takes as primitive. One begins from information and integration; the other from value and what is staked on it. Whether that difference can be made to bear on any observation is the harder question, and the document says where it thinks the answer lies.
Version 2.1 tests the identity by felt relief when a demand is resolved, not by report, which narrows a falsifier. It also takes the unfolding argument separately for each claim: the argument does not reach the architectural claim, and it does reach the identity.