IIT Expert

Ledger

Φ cannot be computed, so IIT cannot be tested

To calculate Φ one needs to consider how every component of the system at every spatial and temporal scale — say from the quantum scale to brain regions, and from nanoseconds to days — constrains the past and future states of the system. Given the complexity of these calculations, it is impossible in principle to compute Φ, and hence impossible to identify conscious states — a point that proponents of IIT readily admit.

— IIT-Concerned, Klincewicz, Cheng, Schmitz, Sebastián & Snyder (2025), "What makes a theory of consciousness unscientific?", Nature Neuroscience 28:689–693, p. 689, "Scientific theories of consciousness should be testable"; reference number omitted. doi:10.1038/s41593-025-01881-x

For the brain, this means that, in principle, the search must be conducted from, at least, the level of its atoms to that of whole cortical areas (Balduzzi & Tononi, 2008; Tegmark, 2016). This compounded computational intractability alone prevents IIT from being tested on relevant brain data.

— Merker, Williford & Rudrauf (2022), "The integrated information theory of consciousness: A case of mistaken identity", Behavioral and Brain Sciences 45:e41, p. 3; note number omitted. doi:10.1017/S0140525X21000881

Restatement. Identifying a complex requires evaluating φs for every candidate system, every partition and every grain of units and updates, and unfolding its Φ-structure requires evaluating every candidate distinction and relation. The number of these grows combinatorially, so the exhaustive calculation cannot be carried out for a brain. The claim is that this makes the theory untestable, because the quantities that its core claims are about can never be measured on the systems those claims concern. The critics' "Φ" covers both φs and Φ.

Status: answered

Response. IIT agrees that the exhaustive calculation cannot be performed for a realistic system and denies that this prevents testing. IIT 4.0 states that "it is not possible in practice to exhaustively apply the postulates", that "a full analysis of complexes and their cause–effect structure can only be performed on idealized systems of a few units", and, in the same section, that the theory is "eminently testable" (iit:ref/albantakis-2023b, "Conclusions"). The paper the 2025 letter cites for its "in principle" admission says "in practice". The measures are defined, and for small systems they are computed exactly, up to about ten to twelve units (iit:ref/mayner-2018b, "Limitations"); for larger ones they can be estimated under assumptions and approximations, the number of relations can be determined analytically, and bounds on Φ are known (iit:ref/albantakis-2023b, "Conclusions"; iit:ref/zaeemzadeh-2024b). The reply compares the situation with quantum chemistry, where the Schrödinger equation is solved exactly only for textbook systems (iit:ref/tononi-2025a, Box 2, second reply). Predictions follow from what the postulates require of an architecture without a value of φs for the whole brain: a dense lattice should support more integrated information than a modular arrangement, and the reply names an outcome that would count against the theory, estimates of integrated information higher "for a model of molecular interactions within a neuron than for a model of synaptic interactions among neurons in a dense cortical lattice" (same reply).

Computation: iit:result/albantakis-2023-fig6d-lattice IIT has stated the limitation from the start, each time together with what remains possible. The 2003 paper calls exhaustive measurement of Φ in large systems "a practical limitation" (iit:ref/tononi-2003b, "Limitations and extensions"), and the 2004 paper adds that "approximations and informed guesses are certainly conceivable" (iit:ref/tononi-2004, "Implications of the hypothesis"). The paper Merker et al. cite says that computing φ for parts of the human brain "is not a feasible undertaking, though heuristics, estimates, and relative comparisons remain possible" (iit:ref/balduzzi-2008, "Discussion").

The letter's "every spatial and temporal scale" reflects a real requirement: the exclusion postulate selects the grain of units and updates that yields maximal φs, so in principle grains must be compared (iit:ref/marshall-2026). In practice heuristic criteria narrow the candidates: for the brain, to neurons or groups of neurons over milliseconds to a few seconds (iit:ref/hoel-2016, "Assessing the intrinsic spatiotemporal scale of the brain: limitations and heuristic strategies"; iit:ref/marshall-2026, "Discussion"). The reply to Merker et al. treats grain as a prediction: "if it turns out that the units of our PSC are not at the grain that allows for maximum Ф, IIT is wrong" (iit:ref/tononi-2022). The question for this entry is whether that makes the theory untestable or only its exhaustive application infeasible. A related but distinct claim, that Φ is not well defined for general physical systems (Barrett & Mediano 2019), belongs to a separate entry.

Sources