Supplement of ARC-COGITATE
Adversarial Collaboration

This page presents IIT's section of the Supplementary Discussion of this adversarial collaboration:

Cogitate Consortium., Ferrante, O., Gorska-Klimowska, U. et al. Adversarial testing of global neuronal workspace and integrated information theories of consciousness. Nature (2025). https://doi.org/10.1038/s41586-025-08888-1.

The results corroborate IIT’s overall claim that posterior cortical areas are sufficient for consciousness, and neither the involvement of PFC nor global broadcasting are necessary. They support preregistered prediction #1, that decoding conscious contents is maximal from posterior regions but often unsuccessful from PFC, and prediction #2, that these regions are sustainedly activated while seeing a stimulus that persists in time. They do not support prediction #3 concerning sustained synchrony, although this negative finding is quite possibly the result of sparse electrode coverage (see 12.1 Further observations). Below we illustrate how these predictions were motivated by IIT. 

Posterior regions are often considered mere ‘information processors’; their activation, it is claimed, may be necessary but not sufficient for experiencing specific contents. For example, they may show activations during deep sleep or anesthesia and for unreported stimuli under contrastive, near-threshold paradigms [19]. This seems to warrant the need for additional ingredients, such as ‘global broadcasting’ [19] or ‘higher-order monitoring’ by PFC [20]. 

For IIT, however, posterior regions are sufficient for consciousness as long as they satisfy the requirements for maximal integrated information. Why this prediction? Unlike other approaches, IIT infers the essential physical requirements for the substrate of consciousness from the essential properties of experience [21, 22]. This leads to the claim that the quality and quantity of an experience are accounted for by the ‘cause–effect structure’ specified by a substrate with maximal integrated information, called the ‘main complex’ [21, 22]. We conjectured that posterior cortical regions should provide an excellent substrate for the main complex owing to their dense local connections arranged topographically into a hierarchical, divergent–convergent 3D lattice [22], leading to prediction #1. Nevertheless, by IIT, posterior regions can only support consciousness if their physiology ensures high integrated information—which indeed breaks down [23] due to bistability when consciousness is lost in deep sleep and anesthesia [24–26]. 

Much of PFC, in contrast, seems to be organized not as a grid but as a patchwork of segregated columns [27], unfavorable for high integrated information. Even so, any PFC region organized in a grid-like way with dense interconnections with posterior regions may well be part of the main complex. As previously emphasized [28], “…we bear no preconceived enmity to the prefrontal cortex. Indeed, searching for the NCC of specific aspects of experience…in certain anterior regions is an important task ahead.” For example, parts of IFG might contribute to, say, an abstract/evaluative/actionable experiential aspect of faces, which could be consistent with some pNCC analysis results. However, IIT predicts that we would still experience faces (sans aspects contributed by PFC regions) if PFC were selectively inactivated. 

For IIT, all quality is structure: all properties of an experience are accounted for by properties of the cause–effect structure specified by the main complex. Every conscious content (face, object, letter, blank screen) is thus a (sub)structure of integrated information (irreducible cause-effects and their overlaps [21]); it is neither a message [29] that is encoded and broadcasted globally [19, 30, 31], nor a distributed activity pattern, nor a neural process. Indeed, IIT’s research program aims to account for specific conscious contents—why space feels extended, time feels flowing, and phenomenal objects feel like binding general concepts (invariants) with particular features—all exclusively in terms of their corresponding cause-effect structures [21, 32, 33]. As highlighted in the Introduction, when we see Mona Lisa, we see that it is a face, with her particular features, at a particular location on the canvas, and we see her for as long as we look at her. This is why we predicted (prediction #2) that the NCC in posterior cortex would last for the duration of the percept, notwithstanding the widespread evidence for neural adaptation and onset/offset neural responses (probably due to transient excitation/inhibition imbalance), and (prediction #3) that synchrony would occur (reflecting causal binding) between units in higher and lower areas, supporting respectively invariant concepts and particular features. 

To conclude, moving beyond the contrastive paradigm between seen and unseen stimuli and beginning to account for how experience feels is one key reason why the experiments reported in this adversarial collaboration mark an important development. Another is that they inaugurate a powerful new way of making progress on a problem often considered beyond the reach of science. The group that carried out this endeavor did so in a way that was explicit, open, and truly collaborative—in short, in a way that is paradigmatically scientific.

The present results failed to confirm IIT’s prediction about sustained, content-specific synchrony in the gamma range between category-specific cortical regions and V1/V2. By IIT, when we see, say, a face with its contours, features, and location in space, there should be ‘causal relations’ among all the units contributing those contents. Relations require an overlap between the intrinsic causes and effects of those units [21, 32], and those are likely to result in increased synchrony among active units, hence the prediction. Given the lack of evidence for sustained synchrony in the gamma band, the prediction may be wrong with respect to the frequency range. Synchrony may occur instead in lower frequency ranges that may be more sensitive to broader cortical interactions (see e.g. Casimo et al. [34]). Indeed, main results using iEEG (Figure 4b) show increased content-specific synchrony (PPC) between category-specific cortical regions and V1/V2 (but not with PFC) in the 2-25 Hz frequency range between 0 and ~750 ms post-stimulus onset.  

The failure to confirm sustained synchrony in the gamma range with iEEG in the present study may also stem from technical limitations. As pointed out in the main text, there were only 12 iEEG electrodes in V1/V2 (against 472 in PFC), with only a minority showing sustained activity for longer trials – which was required to include them in the synchrony analysis. Of note, the analysis of iEEG data did not strictly follow pre-registered analyses methods for selecting ROIs, which required restricting measures of synchrony with anatomical V1/V2 to ‘category-selective' channels (showing a stronger response to either faces compared to objects or vice versa) exclusively showing sustained activation compared to baseline in all three pre-specified time windows (0.3-0.5, 0.5-0.8 and 1.3-1.5 sec). Instead, the final analysis included ‘category-selective' channels showing an increase in activity, compared to baseline, in any time window. Such a lenient selection of ROIs decreased the relevance of the synchrony analysis to test IIT by including many areas which were less likely to be NCC. It may also partially explain findings of transient rather than sustained synchrony patterns. Results of an ongoing adversarial collaboration replicating the present paradigm using large-scale iEEG recordings along with single-units in monkeys (https://www.templetonworldcharity.org/accelerating-research-consciousness-our-structuredadversarial-collaboration-projects) may help resolve these issues and further test this prediction of IIT. 

In the present experiment, MEG ROIs for synchrony analyses were chosen using spatial filters defined on broad-band ERFs, rather than based on local gamma-band activation. Because broad-band ERFs likely comprise signals from both activated and deactivated brain areas (e.g. Vidal et al. [35]; Mukamel et al. [36]), measuring changes in synchrony averaged across such ROIs did not formally test IIT’s pre-registered prediction, which only concerns activated areas. 

While several findings revealed sustained tracking of duration in iEEG HGP both within V1/V2 and inferior temporal cortex, results of other analyses (especially for orientation RSA and synchrony) failed to identify such a sustained pattern (instead decaying beyond 500 ms). One possible explanation for such discrepancy is that the two latter analyses strongly depend on content-specificity over time. This poses a challenge in the case of V1-V2 units, which contribute low-level stimulus features and were sparsely sampled. Note that micro-saccades appear to be more numerous beyond 500 ms after stimulus onset (Supplementary Figure 7), implying that the perceived location of faces or objects in the visual field may have shifted slightly. If so, IIT predicts that a different set of units in V1/V2 (sensitive to the different retinotopic location) would contribute to experienced low-level features, and thus to representational similarity and synchronization with category-specific units in higher areas. Again, animal experiments with denser iEEG sampling of V1/V2 may be necessary to resolve this issue. 

The finding of a few fMRI-activated voxels in the PFC in the pNCC analysis may have various explanations. One possibility, mentioned in the main text, is that such areas may genuinely contribute some content of consciousness (such as some abstract/evaluative/actionable aspect of faces [11, 28, 37]). However, the pNCC contrast against baseline is admittedly not very specific (much less than duration-tracking using iEEG). Additionally, none of the activated PFC areas also showed significant fMRI decoding against baseline, further questioning their relevance for consciousness. The weak PFC fMRI activation patterns may also reflect inputs to PFC from posterior cortex or non-specific onset responses (found to be widespread within PFC areas using iEEG, in contrast with absent PFC duration-tracking). Further experiments employing slowly morphing stimuli may shed light on what determines such onset (or offset) responses. 

IIT’s prediction #1 of less consistent decoding of conscious contents in PFC compared to posterior cortex was verified not only through 1) a failure of decoding orientation in both iEEG and fMRI datasets, but also by 2) a failure of cross-task decoding for letter vs false fonts in PFC using fMRI (Extended Data Figure 1a); 3) a lack of duration tracking for decoding in PFC using iEEG (Extended Data Figure 1d-e); and 4) a lack of cross-task iEEG decoding for faces vs. objects using pseudotrials (Extended Data Figure 1g). In contrast, positive results were consistently found in posterior cortical areas. IIT’s prediction of maximum decodability within posterior regions, with no significant additional information about content added by PFC regions, was confirmed using two different multivariate model comparison methods. These results all fit with IIT’s prediction that posterior cortical areas are the primary constituents of a complex having maximal integrated information. 

Of note, both the sparsity of iEEG electrodes showing duration-tracking (prediction #2) and the widespread deactivation patterns found in PFC (pNCC analysis) strongly suggest that changes in contents of consciousness may be supported by localized activity changes, rather than by a global ‘broadcasting’ and ignition across the brain. These findings are in line with IIT’s prediction that local changes in the state of the substrate of consciousness are sufficient to determine changes in its contents.