nucleation

NOTE-812: Cogitate Consortium — Adversarial testing of GNWT and IIT

12 min read

Properties2
StatusRead
PaperLIT-973 — Adversarial testing of global neuronal workspace and integrated information theories of consciousness

Graph View

Contribution

This is the first large preregistered adversarial collaboration on theories of consciousness. The proponents of integrated information theory (IIT) and of global neuronal workspace theory (GNWT) agreed in advance on divergent predictions, pass and fail criteria and interpretations. Theory-neutral teams then collected and analysed data in three modalities across several laboratories. The results challenge one central prediction of each theory: IIT’s sustained posterior synchronization, and GNWT’s offset ignition and prefrontal representation of all conscious content. They confirm others.

Key insight

A theory of consciousness meets data through its proposed biological implementation, not through its core. IIT’s core is integrated information and GNWT’s is the global workspace. What can be tested are the claims that the maximum of integrated information lies in posterior cortex, and that prefrontal cortex is the workspace hub. So a failed prediction challenges an implementation, and “theorists could respond to challenging data by modifying the proposed biological implementation while retaining the mathematical or computational core”. The design gets its force from clear signals. Every stimulus was clearly seen and fully attended, so a missing effect cannot be blamed on weak input.

Assumptions

  • Suprathreshold, clearly seen stimuli. Faces, objects, letters and false fonts, 20 identities each, in three orientations (front, left, right) and three durations (0.5, 1.0, 1.5 s). Seeing was confirmed in a separate experiment with a surprise memory test.
  • Experience tracks the stimulus. The design assumes the experience has the stimulus’s category, identity and orientation for as long as it is on. The Mona Lisa example (Fig. 1b) states it. Percept and stimulus never come apart.
  • Task relevance as the control for report. Participants reported only rare targets. Analyses focus on task-irrelevant stimuli, and orientation and duration were always task-irrelevant.
  • Theory-defined ROIs. Posterior “hot zone” for IIT and prefrontal cortex for GNWT, fixed in advance with the proponents.
  • Signals. High gamma (70–150 Hz) for iEEG as a proxy for spiking, source-reconstructed MEG, and BOLD.
  • Lenient falsification. “Evidence for any tested feature … was sufficient to rule out failure, rather than requiring consistency across all features.”

Key results

Samples: fMRI n = 120 (73 in the replication set), MEG n = 102 (65), iEEG n = 34 patients (29 to 31 per analysis). MEG and fMRI analyses were tuned on one third and run on the remaining two thirds. iEEG was not split.

Prediction 1, decoding of content.

  • Category (faces against objects) generalised across tasks in both regions. iEEG: more than 95% in posterior ROIs for about the stimulus duration, about 70% in prefrontal ROIs but limited to about 0.2–0.4 s. fMRI searchlight: about 75% in striate, extrastriate, ventral temporal and intraparietal cortex, about 60% in middle and inferior frontal cortex.
  • Orientation of faces, always task-irrelevant: decoded in posterior but not prefrontal ROIs.
    • iEEG: about 95% with pseudotrials; prefrontal null, BF01 = 5.11–8.65.
    • fMRI: about 45% posterior, chance 33%. 34–55% of prefrontal voxels support the null (BF01 3–71.5), and only 1–9% the alternative.
    • MEG: about 75% posterior, a weak 35% prefrontal, “with a possibility of signal leakage from posterior regions”.
  • Adding prefrontal ROIs to posterior ones did not improve decoding and sometimes reduced it. iEEG BF01 = 1.94 × 10⁴ (face–object), 1.91 × 10⁵ (letter–false font), 1,205 (orientation). MEG BF01 3.05–4.70.

Prediction 2, maintenance over time.

  • Activation (iEEG high gamma, LMMs on three windows).
    • 25 of 657 posterior electrodes tracked duration: 12 independent of category, mostly early visual, and 13 category-specific, mostly in ventral temporal cortex. Only 8 of 53 face-selective electrodes (15%) were sustained; most were transient at onset.
    • Prefrontal: 99 electrodes gave non-selective onset responses and 24 category-selective ones. “None of the 655 electrodes measured the temporal profile predicted by GNWT (that is, onset and offset)”, with BF01 > 3 for every prefrontal electrode. The same model found the onset–offset pattern in 10 striate or extrastriate electrodes, so the null is not a failure of the analysis.
    • An exploratory decoding of duration found one inferior frontal sulcus electrode with the GNWT pattern, at 0.15 s rather than 0.3–0.5 s.
  • Representation (cross-temporal RSA, 1.5-s trials, iEEG).
    • Posterior: a sustained face–object distinction, fitting the IIT model better than the GNWT model. Object identity sustained (also letters and false fonts), but not face identity. Face orientation appeared weakly at onset and decayed after about 0.5 s, “contrary to the predictions of IIT”.
    • Prefrontal: category transient at onset, none at offset, even in the task-relevant condition. No identity information at all. No orientation information.
    • MEG supported neither theory on category, identity or orientation.

Prediction 3, interareal connectivity.

  • PPC (preregistered).
    • iEEG: category-selective sites synchronised with V1/V2 early and briefly (under 0.75 s), at 2–25 Hz, “mostly explained by the stimulus-evoked response”. Gamma-band sustained synchrony was absent (BF01 = 1.15–4.9). There was no content-selective PPC with prefrontal cortex (BF01 = 2.62–5.32).
    • MEG: early, low-frequency, evoked; gamma BF01 > 3.
    • “Supported neither theory.”
  • DFC (exploratory, evoked response removed).
    • iEEG: brief object-selective coupling with V1/V2, mainly gamma. Face coupling with V1/V2 scarce (BF01 = 1.3). Gamma coupling of prefrontal cortex with both face- and object-selective sites in GNWT’s window.
    • V1/V2 had only 12 electrodes against 472 in prefrontal cortex.
  • fMRI gPPI. FFA showed face-selective coupling with V1/V2, inferior frontal gyrus and intraparietal sulcus, “consistent with predictions from both theories”. LOC showed none, and 62–94% of prefrontal voxels supported the null (BF01 3–7.75).

The consortium’s verdict (Discussion, Extended Data Fig. 7).

  • IIT. “The lack of sustained synchronization within posterior cortex represents the most direct challenge”, being “incompatible with IIT’s claim that the state of the neural network, including its activity and connectivity, specifies the degree and content of consciousness”. IIT passed the duration prediction, but orientation was not sustained.
  • GNWT. “The most substantial challenge … the lack of ignition at stimulus offset”, “especially surprising given the change in conscious experience at the onset of the blank fixation screen”. And no identity in prefrontal cortex, with orientation only in MEG, “raises the question of whether PFC is involved in broadcasting all conscious content … or only a subset”.
  • No verdict. “We invite readers to weigh the evidence themselves.”

Claims

idclaimstrengthsupport
C1Category of a clearly seen image is decodable in both posterior and prefrontal cortex, independent of taskstrongthree modalities, cross-task generalisation, Fig. 2
C2Face orientation and object identity are represented in posterior cortex and not detectably in prefrontal cortexstrongiEEG and fMRI with Bayesian support for the null; MEG weak above chance with possible leakage
C3Prefrontal cortex adds no decodable content beyond posterior cortexstrongiEEG and MEG model comparison with large BF01
C4No prefrontal site shows an onset-and-offset response to a sustained imagestrong655 electrodes, BF01 > 3 each, with a positive control elsewhere
C5Posterior cortex sustains a category representation for the stimulus duration, and a minority of its sites track durationmoderateiEEG RSA and LMMs; MEG did not replicate
C6Content-specific synchronization between category-selective cortex and V1/V2 is brief and low-frequency, not sustained gammamoderateiEEG and MEG PPC with Bayesian support; sparse V1/V2 iEEG coverage
C7Brief content-specific gamma coupling between prefrontal and category-selective cortex occurs in GNWT’s windowweakexploratory DFC only; preregistered PPC found none
C8Each theory passes some and fails some preregistered critical predictionsmoderatethe integration rules of Extended Data Table 1, applied by the consortium

Method

  1. Proponents and consortium jointly preregister three predictions with pass/fail criteria, weighted by centrality (Extended Data Table 1).
  2. Theory-neutral teams collect data in two or three laboratories per modality, with one shared paradigm.
  3. Optimisation on one third of MEG and fMRI data; the frozen pipelines run on the rest. iEEG electrodes pooled across patients into a “super participant” for multivariate analyses.
  4. Prediction 1: linear SVM decoding with cross-task generalisation, searchlight for fMRI, and posterior-against-combined model comparison.
  5. Prediction 2: LMMs comparing null, duration, time-window, IIT and GNWT models by BIC; cross-temporal RSA correlated with theory-predicted matrices.
  6. Prediction 3: PPC (preregistered), DFC by Gaussian-copula mutual information (exploratory), gPPI in fMRI.
  7. Bayes factors throughout to support nulls.

Concepts

  • Ignition. In GNWT, brief content-specific prefrontal activity, “approximately 0.3–0.5 s” after a change in conscious content, at onset and offset, with “activity-silent information maintenance in between”.
  • Posterior hot zone. IIT’s proposed site of the maximum of integrated information, a “temporo–parietal–occipital” region.
  • Critical prediction. One whose failure counts against a theory under the preregistered weighting; others are “non-critical”.
  • Task-irrelevant trial. A stimulus from the categories not being searched for in the block, used to isolate perception from task.
  • Lenient falsification. One passing feature is enough to avoid failure.

Connections

The design and paradigm come from the consortium’s protocol (Melloni et al. 2023, PLoS ONE) and from Melloni et al. (2021) on adversarial collaboration. The multi-duration paradigm is shared with Vishne, Gerber, Knight and Deouell (2023), LIT-644, which the paper cites as one of “two independent studies using comparable methods” whose prefrontal results agree with its own. The GNWT predictions follow Dehaene and Changeux; the IIT ones follow Tononi, Boly, Massimini and Koch (2016) and IIT 4.0 (Albantakis et al. 2023). Tononi, Koch and Dehaene are co-authors, and the paper declares their commercial interests.

Bearing on the record

  • THEORY-102 is replicated in part, and its title is too broad.
    • THEORY-102’s promote_when names “the COGITATE multi-duration results” as the obvious second recording. This paper is that: an independent intracranial cohort (28 patients for RSA) with 576 prefrontal electrodes, and it finds what THEORY-102 states for category. A sustained posterior category representation, and prefrontal category content confined to onset with nothing at offset.
    • It does not hold for every dimension. Face identity was not sustained in posterior cortex, and face orientation decayed after about 0.5 s though the image stayed on. THEORY-102’s title says “the content of a clearly seen, unchanging image is carried for as long as it stays on”. On this cohort that holds for category and object identity, not orientation or face identity.
    • Also, only 15% of face-selective electrodes were sustained; most were transient. That fits THEORY-102’s “amplitude decays”, but the paper did not test a stable population code against decaying amplitude in the way Vishne et al. did.
    • THEORY-102 should say “category and object identity”, and name orientation as not sustained in COGITATE. Its first promote_when kind is partly met: replicated for category, not for every dimension.
  • THEORY-096. Its promote_when says the account needs “the COGITATE results”. They count against two clauses as stated.
    • “Broadcasts that one content.” Prefrontal cortex held the category of a clearly seen image but not its identity, and orientation at most weakly. The paper asks whether prefrontal cortex broadcasts “all conscious content … or only a subset”.
    • Ignition at a change of content. No prefrontal site responded at offset, though the experience changed there.
    • The “sustains” tension recorded on THEORY-096 (2026-10-12) is confirmed. The proponents predicted activity-silent maintenance between onset and offset, and no sustained prefrontal content was found.
    • None of this tests ignition without differential post-perceptual cognition, the kind of result THEORY-096’s promote_when asks for first. The stimuli were all seen.
  • THEORY-098 (a recurrent sensory basis, prefrontal causal at most). The paper is consistent with it. Content in all three dimensions sits in posterior cortex, and prefrontal adds nothing decodable. But IIT’s prediction of sustained posterior gamma synchrony failed, and that prediction is about recurrence between high-level areas and V1/V2. So the data do not support sustained recurrent coupling as the carrier.
  • QUESTION-061.
    • The filing overstated it. LIT-973 was filed as “direct evidence on whether access is one all-or-none broadcast”. It is not direct. No dimension’s access was measured, and there was no report on identity or orientation.
    • What it does give. Posterior cortex carries a seen image’s orientation and object identity, while the hub carries only its category. That is evidence that the prefrontal hub does not receive every dimension of a conscious content.
    • What it cannot say. Whether an unrepresented dimension is “not accessed”, in the question’s sense, is not tested.
  • QUESTION-048.
    • The design is task-based no-report evidence. The authors concede that “some might argue our approach tracks stimulus processing rather than consciousness”. Percept and stimulus never came apart, so nothing here separates experience from processing, much less from access. That is THEORY-094’s limit.
    • The paper’s own reading of what a theory test can do supports THEORY-023’s evidential point. Tests reach biological implementations, not cores, and the field lacks “a formal framework … to quantitatively integrate evidence”.
  • Filed from this reading:
    • THEORY-245: prefrontal cortex carried category but not identity or (detectably) orientation, and no site responded at offset.
    • THEORY-296: no sustained content-specific posterior synchronization with early visual cortex.
  • No instruction for machine-learning practice; nothing for the anthology.

Limitations

The authors name these:

  • residual task engagement with category (not with orientation or duration);
  • no single units;
  • a phenomenal richness the design “still falls short of”;
  • sparse V1/V2 iEEG coverage, 12 electrodes;
  • MEG results sensitive to parameter choices, with possible leakage.

The reading adds these:

  • iEEG electrodes are pooled across patients, so the region-level multivariate results are a pseudo-population.
  • Patients with epilepsy, and three below the behavioural criteria, were kept.
  • The proponents’ replies in the Supplementary Information were not read.
  • The connectivity result for GNWT depends on which metric is used: preregistered PPC found nothing, and exploratory DFC found gamma coupling.

Open questions

  • Is orientation maintained by any signal over the image’s duration? The authors leave it open for IIT.
  • Does prefrontal cortex carry identity or orientation in other bands, or in single units? The authors point to a separate animal collaboration.
  • Would the same tests with percept and stimulus dissociated (rivalry, masking) move the results? That is what THEORY-102, THEORY-094 and QUESTION-061 all need.
  • How should results be integrated across modalities and predictions? The authors call for a quantitative framework and do not provide one.

Backlinks