Integrated Information Theory vs Global Workspace: What Was Settled

The contest between integrated information theory vs global workspace theory is the closest thing consciousness science has to a heavyweight title fight, and in 2025 the two camps finally stepped into the same ring under rules they had agreed on in advance. The result, published in Nature by the Cogitate Consortium (2025), did not crown a winner. It showed something narrower: both theories make predictions that can fail, and some did.

This article asks what exactly that experiment settled, what it left open, and why one of the two theories is so hard to test at all. It also runs a small computation, built for this piece, of IIT’s central quantity in toy circuits, to show what that number responds to and why nobody can compute it for a brain.

The question, stated precisely

Two families of theory dominate the scientific search for the neural basis of experience. Global neuronal workspace theory (GNWT) says a content becomes conscious when it is broadcast across a long-range network anchored in prefrontal and parietal cortex. Integrated information theory (IIT) says consciousness is identical to a specific kind of intrinsic causal structure, quantified as Φ (phi), and that in humans the relevant structure sits mostly in posterior cortex.

The unsettled question is not simply “which one is right?” Both could be wrong, and the more interesting puzzle is methodological: can a single experiment, or any set of experiments, discriminate between a theory about information access and a theory about intrinsic causal structure? If the answer is no, the field needs different tools. If yes, the 2025 study is a first data point, and the scorecard matters.

Public interest is easy to document. When more than a hundred researchers posted a preprint in September 2023 calling IIT pseudoscience, the dispute spilled into general outlets; The Conversation, for one, ran a piece on why a leading theory had been branded that way. For a gentler map of the territory, see this site’s overview of the main scientific theories of consciousness; the present piece narrows to the head-to-head test and the measurement problem behind it.

What each theory actually claims

In its modern form, GNWT holds that unconscious processing runs in parallel specialized modules, while conscious access involves a sudden, self-sustaining reverberation across a distributed network. Mashour and colleagues (2020) describe how “a non-linear network ignition associated with recurrent processing amplifies and sustains a neural representation,” making it available to many systems at once. Prefrontal cortex is central, and so is timing: ignition should appear when a content enters awareness and, per the preregistration, again when it changes.

IIT starts from the other end. Tononi first set it out in 2004, and the current formulation, IIT 4.0 (Albantakis et al. 2023), begins from axioms said to be true of every experience (it exists for itself, is specific, unified, definite, and structured) and translates them into requirements on a physical system’s cause-effect power. Consciousness, on this view, is whatever substrate maximizes integrated information, and the quality of an experience is the full structure of causal relations that substrate specifies. Tononi, Boly, Massimini and Koch (2016) argued that in the human brain this maximum likely lies in a posterior “hot zone” rather than in prefrontal cortex.

So the theories disagree about where, when, and why: prefrontal bursts at transitions for GNWT, sustained posterior activity carried by connectivity for IIT. Those differences made an adversarial test possible.

What the adversarial collaboration tested

The design was published before any data were collected. In the protocol paper (Melloni et al. 2023), proponents of both theories agreed on predictions, and six theory-neutral laboratories committed to running the experiments with three methods: functional MRI, magnetoencephalography and EEG, and intracranial recordings in epilepsy patients. Outcomes were framed as supporting, challenging, or refuting each theory, with Bayesian statistics used to separate absence of evidence from evidence of absence.

Participants (256 in total) watched clearly visible pictures shown for different durations. Three preregistered tests carried the weight:

• Where is the content? IIT predicted that the identity of what was seen would be best decoded from posterior cortex and that adding prefrontal signals would not help. GNWT predicted a necessary prefrontal contribution. Category could be read out from both regions; orientation came mainly from posterior areas; adding prefrontal cortex did not improve decoding.
• How is it maintained? IIT predicted activity in posterior cortex that lasts as long as the stimulus. GNWT predicted brief prefrontal ignition at onset and again at offset. Some posterior sites tracked duration. No prefrontal site showed the onset-and-offset pattern.
• How do areas talk? IIT predicted sustained gamma-band synchronization within posterior cortex. GNWT predicted brief long-range coupling between category-selective areas and prefrontal cortex. The preregistered phase-based measure supported neither; a later amplitude-based analysis gave a mixed picture.

The consortium summarized the damage to each side in two sentences from its abstract:

“For IIT, a lack of sustained synchronization within the posterior cortex contradicts the claim that network connectivity specifies consciousness.”

Cogitate Consortium, 2025, Nature

“GNWT is challenged by the general lack of ignition at stimulus offset and limited representation of certain conscious dimensions in the prefrontal cortex.”

Cogitate Consortium, 2025, Nature

What the study established, then, is limited but real. Prefrontal cortex is not required to decode at least some dimensions of a conscious visual content. The predicted prefrontal burst when a stimulus disappears did not show up. And the sustained posterior synchronization that IIT’s proponents had committed to also failed to appear. Neither theory was refuted outright; each lost something it had staked in public.

Three ways to read the scorecard

Here are the main readings, each in its strongest form.

Reading one: IIT came out ahead

Two of the three tests leaned toward the posterior-cortex story, and GNWT’s signature prediction, ignition at offset, simply was not there. On this view the experiment moved credence toward IIT, or at least toward “posterior” accounts in general. The weakness is that the posterior predictions are not unique to IIT. Recurrent processing theory and several other views also locate visual experience at the back of the brain, so evidence for posterior involvement is not evidence for Φ. The authors themselves noted that the challenges extend to other theories sharing the tested predictions.

Reading two: GNWT survives with auxiliary assumptions

Workspace theorists can argue that offset ignition is not needed if the experience fades gradually, or that the task did not require reporting, so prefrontal broadcast was muted. That is a legitimate move; theories are always tested alongside auxiliary assumptions. The cost is that it makes the theory harder to pin down, and critics will ask which observation GNWT would now accept as a defeat.

Reading three: the method, not the theories, is the main finding

A third reading treats the experiment as proof of concept for adversarial collaboration. Rivals agreed in advance on what would hurt, and both got hurt. The authors closed by calling for a quantitative framework for theory testing and invited readers to weigh the evidence themselves rather than declaring a verdict.

The testability fight around Φ

Running underneath the scorecard is a sharper dispute. In the September 2023 preprint, Fleming and colleagues argued that IIT should be treated as pseudoscience until it is empirically testable, and that the adversarial results did not test its core. A follow-up commentary in Nature Neuroscience by Klincewicz and colleagues (2025) made the case in its most direct form:

“we argue that the theory is indeed unscientific because its core claims are untestable even in principle.”

Klincewicz et al., 2025, Nature Neuroscience

The argument has several strands. The predictions that were tested (posterior activity, sustained synchronization) are derived from IIT plus assumptions about brain anatomy, not from Φ itself. The theory also implies that some simple, inactive lattices of logic gates would be conscious, a consequence the computer scientist Scott Aaronson pressed in 2014 and that Tononi accepted rather than disowned. And Doerig and colleagues (2019) offered the “unfolding argument”: any recurrent network can be replaced by a feedforward one with identical input-output behaviour, so if consciousness is inferred from behaviour or report, a causal-structure theory like IIT is either unfalsifiable or false.

Defenders reply on three fronts. First, every theory needs bridging assumptions to reach data. Second, IIT made risky predictions in the Cogitate study, and one of them failed, which is what falsifiable theories do. Third, the unfolding argument assumes that report is the only evidence for consciousness, an assumption IIT rejects by design. Whether that last move is principled or a shield against refutation is the crux, and it remains open.

One fact is not in dispute: Φ cannot be computed for anything brain-sized. Tegmark (2016) wrote that “the integration measure proposed by IIT is computationally infeasible to evaluate for large systems,” with costs that grow super-exponentially. That is where the small computation below comes in.

A small computation: what Φ tracks in toy circuits

For this article, a short Python script used PyPhi, the reference software released by Tononi’s group (Mayner et al. 2018), to compute Φ for six three-node networks of deterministic logic gates. PyPhi version 1.2 implements the IIT 3.0 formalism, not the newer 4.0 definitions, so the values below are IIT 3.0 values. The pipeline was checked by rebuilding one of PyPhi’s bundled example circuits from logic functions and recovering its published value of 2.3125.

Alongside Φ, the script computed a deliberately crude proxy: the mutual information between the whole system’s current and next state, minus the same quantity for each half, minimized over all ways of cutting the system in two. That proxy is labeled as such throughout. It is not IIT; it is the kind of “whole versus parts” measure people often assume Φ resembles.

import itertools, numpy as np, pyphi   # PyPhi 1.2 (IIT 3.0)

def tpm_from(fns, n):  # deterministic logic -> state-by-node TPM
    rows = []
    for s in itertools.product([0, 1], repeat=n):
        s = s[::-1]                        # PyPhi little-endian order
        rows.append([f(s) for f in fns])
    return np.array(rows)

# connectivity(tpm, n): flip each input bit, record which outputs change
def big_phi(fns, n=3):
    tpm = tpm_from(fns, n)
    net = pyphi.Network(tpm, cm=connectivity(tpm, n))  # cm read off the TPM
    out = []
    for st in itertools.product([0, 1], repeat=n):
        try:
            out.append(pyphi.compute.phi(pyphi.Subsystem(net, st, range(n))))
        except pyphi.exceptions.StateUnreachableError:
            out.append(None)               # state has no possible past
    return out

def proxy(fns, n=3):  # NOT IIT: whole-minus-parts lagged mutual information
    def mi(idx):
        joint = {}
        for s in itertools.product([0, 1], repeat=n):
            nxt = tuple(f(s) for f in fns)
            k = (tuple(s[i] for i in idx), tuple(nxt[i] for i in idx))
            joint[k] = joint.get(k, 0) + 1 / 2**n
        px, py = {}, {}
        for (a, b), p in joint.items():
            px[a] = px.get(a, 0) + p; py[b] = py.get(b, 0) + p
        return sum(p * np.log2(p / (px[a] * py[b])) for (a, b), p in joint.items())
    whole = mi(range(n))
    cuts = [whole - mi(p) - mi([i for i in range(n) if i not in p])
            for k in (1,) for p in itertools.combinations(range(n), k)]
    return whole, min(cuts)

nets = {"OR-AND-XOR loop": [lambda s: s[1] or s[2], lambda s: s[0] and s[2],
                            lambda s: s[0] ^ s[1]],
        "Feedforward chain": [lambda s: 0, lambda s: s[0], lambda s: s[1]],
        "Ring of copies":    [lambda s: s[2], lambda s: s[0], lambda s: s[1]]}
        # ...plus XOR ring, two-loop + isolated node, majority vote
for name, fns in nets.items():
    print(name, big_phi(fns), proxy(fns))

Results from one run (Φ in IIT 3.0 units, computed for every state the network can actually reach; proxy in bits):

A separate timing run gave a sense of scale. It drew two random fully connected logic networks at each size (fixed seed) and timed one Φ calculation for each. Three-node networks averaged about 0.4 seconds on the machine used. Four-node networks averaged about 10 seconds. Five-node networks averaged about 239 seconds, roughly four minutes. Two samples per size is a small sample, so treat these as orders of magnitude, not a fitted curve.

What the toy results show, and what they do not

First, Φ is about causal architecture, not about how much information flows. The feedforward chain passes two bits from one time step to the next, and the whole-versus-parts proxy scores it at a full bit of “integration.” Φ is exactly zero. IIT says a system with no feedback has no intrinsic perspective at all, however sophisticated its behaviour, and the computation confirms that the measure enforces this. It is also the property the unfolding argument targets: a feedforward network can mimic a recurrent one’s input-output behaviour while scoring zero.

Second, Φ rewards structure that does nothing interesting. A ring of three wires that just pass each bit along scores Φ = 1 in every state, higher than the majority-vote circuit, which at least computes something. That is a three-node echo of Aaronson’s complaint about large, simple lattices, and it shows why intuitions about “complex processing” are a poor guide to what the theory measures.

Third, Φ is state-dependent and depends on which states are reachable. The OR-AND-XOR loop ranges from 0.25 to 1.917 across states. Proxy measures computed over all inputs flatten that variation.

Fourth, the exclusion postulate matters. In the two-node loop with an isolated self-looping third node, Φ for the whole system is zero, while the loop on its own scores 1.0 and forms what IIT calls the major complex. Deciding what the conscious system is requires searching over all candidate subsets, which multiplies the cost again.

What the computation does not show matters as much. It says nothing about whether IIT is true. It cannot be scaled up to test the Cogitate predictions, which used empirical proxies (decoding, synchronization) rather than Φ. Three-node results do not necessarily extrapolate to large systems, and the IIT 4.0 definitions could give different numbers for the same circuits.

The timing numbers point to the central practical problem. A cost increase of roughly twenty-fold or more with each added binary unit (0.4, then 10, then 239 seconds) means direct calculation for anything resembling a cortical column is not on the horizon. Every empirical test of IIT therefore runs through indirect stand-ins, and every stand-in can be disputed.

The author’s reading, and what would settle it

Opinion, clearly labeled. My reading is that the Cogitate study settled a methodological question and a few empirical ones, but not the theoretical one. It showed that rival camps can agree on falsifiable neural predictions and that prefrontal cortex is not where every aspect of visual content lives, which weakens the strongest prefrontal versions of workspace theory. It did not test integrated information as such, because no one can measure it in a brain. The “pseudoscience” label strikes me as too strong for a theory whose proponents preregistered predictions and accepted a failure; “untested at its core” is fairer and harder to dispute. Anyone tempted to treat the result as a verdict on the hard problem of consciousness should resist: both theories address which brain processes correlate with experience, and neither result touches why there is experience at all.

Concrete tests that would move the debate:

• Offset ignition with report. Rerun the duration paradigm with an immediate report at stimulus offset. If prefrontal ignition still does not appear when the change must be reported, GNWT’s account of maintenance is in trouble; if it appears only with report, the earlier null reflects task demands.
• Causal manipulation of posterior connectivity. IIT predicts that experience depends on the connectivity of the posterior substrate even when activity levels are matched. Focal perturbations that alter connectivity while preserving firing rates, in patients or with stimulation, should change reports of experience if IIT is right.
• Validating proxies against exact Φ. Compute exact Φ for many small simulated networks (as in this article) and test whether the empirical stand-ins used in human studies, such as perturbational complexity, rank those networks in the same order. If they do not, IIT-inspired brain measures are not testing IIT.
• States with dissociated access and experience. Anesthesia, dreaming and some brain injuries separate reportable access from other markers of experience. GNWT and IIT predict different neural signatures there; this site’s piece on what anesthesia does to the brain covers why those states are informative.

A second adversarial round with a different paradigm would do more than another round of essays. Patients with surgically divided hemispheres, discussed in this site’s article on split-brain research, are an obvious candidate, since the two theories carve such cases differently.

Common questions

Did the 2025 study prove integrated information theory?

No. Some predictions associated with IIT were supported, mainly about posterior cortex, but its preregistered prediction of sustained posterior synchronization was not. The supported predictions are also shared by other theories, so they do not single out IIT.

Was global workspace theory falsified?

Not outright. Its prediction of a prefrontal burst at stimulus offset failed, and prefrontal cortex carried less content than predicted. Proponents can add assumptions about task and report, but the result removes some of the theory’s room to manoeuvre.

Why can’t scientists just measure Φ in the brain?

Computing Φ requires evaluating every way of cutting a system, across states and candidate subsystems, and the cost explodes as units are added. In the toy runs here, each added binary node raised the time per calculation by more than twenty-fold, reaching about four minutes for five nodes. Brain studies therefore rely on indirect measures.

Is IIT pseudoscience?

That is contested. Critics argue its core claims cannot be tested even in principle; defenders point out that it produced preregistered predictions, one of which failed. The label is a judgment about scientific status rather than an established finding.

What is an adversarial collaboration?

It is a study in which proponents of competing theories agree on an experiment and on which outcomes would count against each view before data are collected. Independent labs then run it, limiting after-the-fact reinterpretation.

Further reading

• The Feeling of Life Itself: Why Consciousness Is Widespread but Can't Be Computed by Christof Koch. A book-length popular account of integrated information theory by one of its leading proponents. See it on Amazon (affiliate link)
• Consciousness and the Brain: Deciphering How the Brain Codes Our Thoughts by Stanislas Dehaene. The standard popular statement of global neuronal workspace theory by its principal architect. See it on Amazon (affiliate link)

As an Amazon Associate I earn from qualifying purchases. A commission may be earned if you buy through this link, at no extra cost to you.

References

1. Albantakis, L., Barbosa, L., Findlay, G., Grasso, M., Haun, A. M., Marshall, W., et al. (2023). Integrated information theory (IIT) 4.0: Formulating the properties of phenomenal existence in physical terms. PLoS Computational Biology, 19(10), e1011465. https://doi.org/10.1371/journal.pcbi.1011465
2. Aaronson, S. (2014, May 21). Why I am not an integrated information theorist (or, the unconscious expander). Shtetl-Optimized (blog). https://www.scottaaronson.com/blog/?p=1799
3. Cogitate Consortium, Ferrante, O., Gorska-Klimowska, U., Melloni, L., et al. (2025). Adversarial testing of global neuronal workspace and integrated information theories of consciousness. Nature, 642, 133–142. https://doi.org/10.1038/s41586-025-08888-1
4. Doerig, A., Schurger, A., Hess, K., & Herzog, M. H. (2019). The unfolding argument: Why IIT and other causal structure theories cannot explain consciousness. Consciousness and Cognition, 72, 49–59. https://doi.org/10.1016/j.concog.2019.04.002
5. Fleming, S. M., Frith, C., Goodale, M., Lau, H., LeDoux, J. E., Lee, A. L. F., et al. (2023). The integrated information theory of consciousness as pseudoscience. PsyArXiv (preprint). https://doi.org/10.31234/osf.io/zsr78
6. Klincewicz, M., Cheng, T., Schmitz, M., Sebastián, M. Á., Snyder, J. S., et al. (2025). What makes a theory of consciousness unscientific? Nature Neuroscience. https://doi.org/10.1038/s41593-025-01881-x
7. Mashour, G. A., Roelfsema, P., Changeux, J.-P., & Dehaene, S. (2020). Conscious processing and the global neuronal workspace hypothesis. Neuron, 105(5), 776–798. https://doi.org/10.1016/j.neuron.2020.01.026
8. Mayner, W. G. P., Marshall, W., Albantakis, L., Findlay, G., Marchman, R., & Tononi, G. (2018). PyPhi: A toolbox for integrated information theory. PLoS Computational Biology, 14(7), e1006343. https://doi.org/10.1371/journal.pcbi.1006343
9. Melloni, L., Mudrik, L., Pitts, M., et al. (2023). An adversarial collaboration protocol for testing contrasting predictions of global neuronal workspace and integrated information theory. PLoS ONE, 18(2), e0268577. https://doi.org/10.1371/journal.pone.0268577
10. Tegmark, M. (2016). Improved measures of integrated information. PLoS Computational Biology, 12(11), e1005123. https://doi.org/10.1371/journal.pcbi.1005123
11. Tononi, G. (2004). An information integration theory of consciousness. BMC Neuroscience, 5, 42. https://doi.org/10.1186/1471-2202-5-42
12. Tononi, G., Boly, M., Massimini, M., & Koch, C. (2016). Integrated information theory: From consciousness to its physical substrate. Nature Reviews Neuroscience, 17(7), 450–461. https://doi.org/10.1038/nrn.2016.44