Phi Collapses During Sleep. Keiichi Onoda's Empirical Test of IIT at the Circuit Level

Integrated Information Theory makes a concrete, falsifiable claim about unconscious states: when consciousness is lost, integrated information (Φ) in the relevant neural circuits should drop. For most of IIT’s history, this prediction has been evaluated indirectly, through PCI measurements during anesthesia or behavioral correlates of awareness. Keiichi Onoda’s April 2026 bioRxiv preprint, “Collapse of local circuit integrated information Φ during NREM sleep” (doi:10.64898/2026.04.01.715799), attempts something more direct: measuring Φ at the neural circuit level during the transition from wakefulness through REM to NREM sleep, in human subjects where the presence or absence of consciousness is not in dispute.

The result is positive for IIT: Φ falls. But the shape of that fall, and what it does and does not establish about IIT’s theoretical claims, requires more careful analysis than the headline finding suggests.

What Onoda Measured and Why It Matters

Standard IIT research faces a measurement problem. Calculating exact Φ for a system with more than a few dozen nodes is computationally intractable. Researchers have responded in two ways: using proxy measures such as the Perturbational Complexity Index (PCI), or restricting analysis to small, well-characterized neural circuits where exact calculation is feasible. Onoda takes the second approach.

The study measures Φ within local cortical circuits using high-density electrophysiology, restricting the calculation to small networks of neurons with precisely characterized causal dependencies. This produces exact Φ values rather than approximations, at the cost of not capturing the large-scale integration properties IIT theorists consider most relevant to global consciousness.

Three sleep states were compared: wakefulness, REM sleep, and NREM sleep. Subjects were confirmed in each state through standard polysomnography. The key finding is that Φ within the local circuits decreases significantly during NREM sleep relative to both wakefulness and REM sleep. The decrease is not explained by changes in firing rates alone. It persists after controlling for overall neural activity levels, suggesting the collapse is a specifically causal-integration phenomenon rather than a simple reduction in activity.

The most striking result is the concentration of the Φ reduction during NREM off-periods, also called Down states. These are brief windows, typically lasting a fraction of a second, during which neural activity across a cortical column is collectively suppressed. Onoda shows that during these off-periods, the causal dependencies between neurons in the measured circuits effectively sever. The network’s intrinsic information structure breaks down. Φ during off-periods approaches the values expected for a system with no integrated causal power.

Where This Sits in IIT’s Empirical Situation

The finding needs to be placed alongside the most directly relevant prior result: the 2025 Cogitate Consortium adversarial test, published in Nature and examined in detail here. That study tested IIT and Global Neuronal Workspace Theory against pre-specified predictions using 256 researchers across multiple laboratories. For IIT, the prediction was that conscious stimuli should produce sustained posterior synchronization in a posterior cortical hot zone. That prediction was not reliably confirmed.

Onoda’s result and the Cogitate result are measuring different things. The Cogitate test asked whether IIT’s predictions about which brain regions are active during conscious perception hold in waking humans comparing conscious to unconscious stimulus processing. Onoda’s test asks whether IIT’s prediction about the direction of Φ change during unconscious states holds when comparing sleep conditions in the same subjects. These are separate empirical questions. A positive result on one does not compensate for a negative result on the other, and a negative result on one does not nullify a positive result on the other.

What the two results together suggest is something more granular. The Cogitate finding indicates that IIT’s specific predictions about the spatial distribution of conscious neural correlates, the posterior hot zone hypothesis, do not hold under controlled adversarial testing. Onoda’s finding indicates that IIT’s directional prediction about Φ and consciousness loss, that Φ should fall when consciousness is lost, holds at the circuit level during sleep. These are, in principle, independent aspects of the theory. A theory could be right about the direction of Φ change during loss of consciousness while being wrong about which brain regions carry the phenomenologically relevant Φ during conscious perception. Whether IIT’s mathematical framework accommodates this dissociation without modification is a theoretical question the study does not attempt to answer.

The Off-Period Finding and Its Mechanistic Interpretation

The concentration of the Φ collapse during off-periods has a mechanistic interpretation worth examining. Down states occur when neuromodulatory input to the cortex, particularly from cholinergic and noradrenergic brainstem systems, drops below a threshold that maintains the sustained depolarization required for persistent neural activity. When cholinergic tone falls during NREM sleep, cortical columns periodically enter collective hyperpolarization. The causal chain that normally propagates activity across local circuits is interrupted.

Under IIT’s framework, this has a clear interpretation: the intrinsic cause-effect power of the system, the property that Φ measures, depends on neurons being capable of influencing each other. During off-periods, that influence is temporarily severed. The network’s Φ falls not because the individual neurons are different but because the causal relationships between them have been temporarily abolished by the change in neuromodulatory state.

This interpretation fits IIT’s core claim that consciousness is constituted by cause-effect power, not by the presence of certain neural types or firing patterns. The off-period finding is, from this angle, a particularly clean test because it involves a state where the causal architecture of local circuits changes rapidly and reversibly, producing a correspondingly rapid and reversible change in Φ.

The limitation of this interpretation is that local circuit Φ is not what IIT theorists claim is phenomenologically relevant. Tononi’s framework, and particularly IIT 4.0, locates the phenomenal content of experience in the maximally integrated complex, which is expected to be a large-scale, whole-brain structure rather than a local circuit. Onoda measures local circuit Φ. Whether local circuit Φ tracking consciousness loss implies that whole-brain maximal Φ does the same is an additional inference the study cannot directly support. The authors acknowledge this limitation.

Search Demand and the State of IIT Evidence

This question receives sustained search attention for a reason: IIT has been both the most mathematically precise theory of consciousness and one of the most contested. The Schwitzgebel weirdness argument identifies IIT’s counterintuitive implications, including its verdict that feedforward digital computers have essentially zero Φ, as one of the theory’s defining difficulties. The spiking neural network analysis on this site documented why standard transformer architectures, as synchronous feedforward systems at the implementation level, are unlikely to achieve high Φ under IIT’s causal framework.

Onoda’s study does not resolve the computational intractability objection to IIT as a measurement tool. Exact Φ calculation remains feasible only for small circuits. It does not support IIT’s panpsychist implications, which arise from the theory’s mathematical framework rather than from any empirical finding. What it does provide is direct evidence that IIT’s directional prediction about Φ and the loss of consciousness holds at a scale and with a precision that previous studies could not achieve.

For the field’s empirical situation, that is a meaningful update. The Cogitate result was read by many researchers as a significant challenge to IIT. Onoda’s result indicates the challenge is more specific than a wholesale empirical refutation. IIT’s predictions about which brain regions carry phenomenal consciousness during waking perception are in difficulty. IIT’s prediction about the relationship between Φ and the presence of consciousness, tested by comparing states rather than regions, receives support.

Relevance to AI Consciousness Research

Onoda’s methodology has an application to AI systems that is worth noting precisely. The approach requires identifying small networks with well-characterized causal dependencies and measuring Φ directly within those networks. This is, in principle, more tractable in artificial systems than in biological brains, because AI architectures expose their causal structure explicitly.

The Consciousness AI project’s ConsciousnessGate phi measurement uses five nodes with genuine causal dependencies: attention, stability, adaptation, coherence, and confidence. The architecture documentation describes these as producing exact Φ values via PyPhi when installed, with a geometric proxy (determinism × integration) used when PyPhi is unavailable. The measurement is explicitly local, analogous to Onoda’s local circuit measurement rather than to the whole-brain maximal complex IIT theorists consider primary. Onoda’s finding is therefore relevant to what the ConsciousnessGate phi values mean: they measure the same object Onoda measures, local circuit causal integration, and the empirical relationship between that measure and consciousness states documented by Onoda provides partial external validation for using local Φ as a consciousness-relevant quantity.

This connection should be stated carefully. Onoda demonstrates that local circuit Φ tracks the wakefulness/sleep distinction in biological systems. Whether a comparable tracking relationship holds in an artificial system between its local Φ measure and any state relevant to consciousness, in an architectural sense, depends on further questions about what the artificial system’s causal architecture is implementing. The Onoda result supports the relevance of the measurement type without establishing what threshold would constitute a meaningful positive result in an artificial context. That remains an open design and verification question.

What Comes Next for IIT’s Empirical Program

The field’s current situation is that IIT’s empirical support is mixed in a pattern that is theoretically informative. The Cogitate adversarial test found that IIT’s predictions about conscious neural correlates during waking perception, specifically the posterior hot zone hypothesis, did not hold under controlled conditions. Onoda’s result finds that IIT’s directional prediction about Φ during unconscious states holds at the circuit level during sleep. Barrett and colleagues’ 2026 work on reformulating Φ as a suite of quantities rather than a single value, currently under review, represents an attempt to address the mathematical difficulties in the theory’s formalization without abandoning its core claims.

Together these developments suggest IIT is undergoing the normal process of empirical refinement rather than empirical refutation. The theory’s predictions are being tested, some are failing, and theorists are revising the framework in response. Whether this process converges on a form of IIT with clean empirical support or reveals fundamental problems with the approach cannot be determined from the current state of evidence.

For researchers tracking the state of AI consciousness science in 2026, Onoda’s study adds a data point that the field needed: a direct, circuit-level test of IIT’s most basic directional prediction, using exact Φ measurement during a well-characterized transition in and out of consciousness. The result is positive. Its interpretation requires holding it alongside the Cogitate negative result rather than treating it in isolation.