IIT Expert

Ledger

The unfolding argument

Here, using theorems from the theory of computation, we show that causal structure theories are either false or outside the realm of science.

— Doerig, Schurger, Hess & Herzog (2019), "The unfolding argument: Why IIT and other causal structure theories cannot explain consciousness", Consciousness and Cognition 72:49–59, abstract. doi:10.1016/j.concog.2019.04.002

Restatement. Any recurrent network can be replaced ("unfolded") by a feedforward network with the same input–output function. IIT assigns the two different values — the feedforward twin necessarily has φs = 0 — while their behaviour is identical, so no behavioural experiment can distinguish them. If evidence about consciousness must be behavioural (the argument's premise), IIT is either contradicted by that evidence or untestable by it.

Status: disagreement

Response. IIT rejects the premise, not the construction. The premise — that input–output behaviour is the only admissible evidence about consciousness — is functionalism taken as an epistemic axiom, and IIT does not grant it: the theory starts from the existence of experience (the axioms) and identifies an experience with the cause–effect structure specified by a substrate, so two substrates with the same input–output function but different causal structure differ in Φ. That the unfolded feedforward twin has φs = 0 and is not conscious is a prediction of the theory, not a refutation of it, and it is testable against non-behavioural evidence (perturbation and lesion studies among others). Both parties understand each other and differ on the premise of what counts as evidence, hence the status. The published reply is iit:ref/tsuchiya-2020; the dispute is also catalogued in iit:ref/albantakis-nd.

Computation: iit:result/feedforward-3node The unfolding construction itself is not in dispute, and it did not originate with Doerig et al.: IIT published it five years earlier. IIT 3.0 has a section titled "Conscious complexes and unconscious 'zombie' systems can be functionally equivalent" (iit:ref/oizumi-2014, Fig. 21). It cites the universal approximation theorems for feedforward networks (Cybenko 1989; Hornik et al. 1989), unfolds a recurrent system over time into a feedforward network with the same input–output behaviour, and shows that the recurrent system forms a complex with Φ > 0 in every state while the feedforward one forms no complex at all. Albantakis et al. (iit:ref/albantakis-2014, Discussion) make the same point the same year: "any task could, in principle, be solved by a modular brain with Φ = 0 given an arbitrary number of elements and time-steps". Albantakis's commentary "Unfolding the Substitution Argument" (in the corpus document consciousnessrealist) notes both precedents: "Guess who introduced this issue?" IIT used the construction to argue that functional equivalence does not entail phenomenal equivalence; the unfolding argument invokes the same proof against the theory (iit:ref/tononi-2025b, "The 'unfolding' argument"; iit:ref/albantakis-2021c makes the same point with the Krohn–Rhodes theorem). What the theory denies is that behavioural equivalence settles the question of experience — on IIT's account the physical substrate's cause–effect structure, not the function it computes, is what an experience is. The linked computation shows the feedforward half of the dissociation concretely: a three-unit feedforward chain has φs = 0 because it is not strongly connected, so a partition cutting in the (absent) feedback direction makes no difference.

Sources