Neither Theory Survived: What the Cogitate Consortium's Adversarial Test Found

In April 2025, the Cogitate Consortium published what remains the most rigorous direct empirical test of the two most influential scientific theories of consciousness. The study enrolled 256 human participants across multiple laboratories and put them through fMRI, magnetoencephalography (MEG), and intracranial EEG. The result, published in Nature (Volume 642, Issue 8066, pages 133, 142; doi.org/10.1038/s41586-025-08888-1), was unambiguous. Neither Integrated Information Theory nor Global Neuronal Workspace Theory emerged with its core empirical predictions intact.

TL;DR. The Cogitate Consortium was the first preregistered adversarial collaboration in consciousness science. The proponents of IIT and GNW agreed in advance what findings would falsify their own theories. IIT was challenged by the absence of sustained posterior neural synchronization during conscious experience. GNW was challenged by absent ignition at stimulus offset and weaker-than-predicted prefrontal representation of some conscious dimensions. Neither theory is refuted, but neither can be applied to artificial systems in 2026 without accounting for the limits this study exposed.

What Adversarial Collaboration Means in Practice

Most empirical tests of consciousness theories follow a predictable pattern. A research group aligned with one theoretical framework designs an experiment, collects data, and interprets the results through that framework. When findings are inconvenient, the interpretation shifts: the paradigm was a special case, the operationalization was imperfect, the stimuli were not optimal. With enough interpretive flexibility, almost no result counts as a genuine falsification.

The adversarial collaboration model closes this exit. Before the Cogitate Consortium ran a single participant, the proponents of each theory sat across from each other and agreed on two things: what their theory specifically predicted, and what would count as evidence against it. Stanislas Dehaene (Collège de France) led the Global Neuronal Workspace arm. Giulio Tononi (University of Wisconsin) led the Integrated Information Theory arm. A theory-neutral coordination team oversaw the entire process. The predictions were pre-registered. The analysis plan was locked before any data was unblinded.

This design changes the epistemic status of the findings. The 2025 Nature results cannot be explained away by either side. Two of the most prominent living theorists of consciousness watched their own predictions fail to hold on data they had already agreed constituted a fair test. Nothing quite like that had happened before in consciousness science.

The Experimental Setup

The study enrolled 256 participants. The imaging battery covered three complementary modalities, chosen because each captures something the others cannot. fMRI provided spatial precision at millimeter resolution, revealing where in the brain activity was concentrated. MEG captured the temporal dynamics of neural activity at millisecond resolution, tracking the sequence of events as stimuli were processed. Intracranial EEG, recordings from electrodes placed directly in the brain, used with patients undergoing clinical monitoring for epilepsy surgery, provided the highest-resolution direct access to neural activity in specific regions, free from the distortion introduced by the skull and scalp.

Participants performed tasks designed to generate sharp contrasts between conscious and unconscious processing. Paradigms included visual masking (where a target stimulus is briefly flashed and then masked, producing conditions under which the same stimulus is sometimes consciously perceived and sometimes not), binocular rivalry (where different images are presented to each eye, forcing alternation between conscious percepts), and threshold detection. The core logic across all conditions was consistent: when a participant reports consciously perceiving a stimulus versus when they report no perception, what is the neural signature of that contrast, and does it match what IIT and GNW respectively predict it should look like?

What Integrated Information Theory Predicted, and What Was Found

Integrated Information Theory holds that consciousness is identical to integrated information, the information generated by a system above and beyond what its parts generate independently. This quantity, phi, captures the degree to which the system constitutes a unified causal whole irreducible to its components. Tononi’s framework provides a specific computational structure for what high-phi systems look like and where in the brain they operate.

For the biological case, IIT makes a testable anatomical prediction. Conscious experience should correlate with sustained synchronous neural activity in the posterior cortical hot zone, a network spanning occipital, temporal, and parietal cortex. This posterior region, in IIT’s formalism, is where integrated information processing is richest. During the period when conscious experience is present, IIT predicts sustained synchronization should be detectable in this network and should track the content and intensity of conscious experience across conditions.

The Cogitate data did not support this prediction. The sustained posterior synchronization that IIT’s framework requires was absent. Posterior cortical regions responded to conscious stimuli, as prior work had suggested they do, but the specific pattern of sustained synchronization that IIT’s causal geometry predicts was not found at the scale or consistency the theory demands.

This has consequences that extend beyond the biological case. The 2026 Brock University and Institute of Noetic Sciences study, which applied IIT’s phi mathematics to computational systems to quantify artificial cause-effect power, builds on a theoretical foundation that was challenged at its core empirical level by the Cogitate findings. Any researcher citing phi as evidence of consciousness in an artificial system in 2026 is working with a metric whose biological grounding is now contested in the most demanding test of IIT yet conducted.

What Global Neuronal Workspace Theory Predicted, and What Was Found

Global Neuronal Workspace Theory, developed by Bernard Baars (Neurosciences Institute) and extended by Dehaene and Jean-Pierre Changeux (Institut Pasteur), proposes a different mechanism for conscious access. The brain, in this view, processes most information locally, in encapsulated specialized modules. Consciousness is the event that breaks this modularity: a sudden, widespread ignition that broadcasts locally processed information to a global workspace accessible by many downstream cognitive systems simultaneously. Attention, working memory, verbal report, and volitional action all depend on this global broadcast.

GNW’s predictions for the Cogitate study were correspondingly specific. Conscious experience should produce clear ignition events at stimulus onset, when the content first becomes globally available, and at stimulus offset, when the stimulus ends and its conscious representation either persists or decays. Prefrontal cortex, as the central node of the global workspace, should show robust representation of conscious content across the range of stimulus dimensions the study probed.

The findings partially supported GNW but not uniformly. Ignition at stimulus onset was observed across paradigms. Ignition at stimulus offset was largely absent, a result that sits in direct tension with GNW’s prediction that the offset of a conscious stimulus should produce its own broadcast event. Prefrontal representation of some conscious dimensions was weaker than the theory predicted, with variability across stimulus types and participants that GNW’s framework had not anticipated.

For AI consciousness research, this matters in the same structural way as the IIT findings. The Theater of Mind architecture (arXiv:2604.08206), the first explicit GNW implementation in a large language model, uses global broadcast logic as its computational foundation. If global broadcast in biological systems does not produce the full pattern GNW predicts, and the Cogitate data suggest it does not, the question of what a machine implementation of that mechanism is actually capturing becomes analytically harder.

Predictions Against Results

Neither Theory Is Refuted. Both Are Weaker.

The Cogitate results do not end IIT or GNW as research programmes. Both theories have explanatory scope well beyond the specific predictions the Consortium pre-registered, and both have proponents who argue, with some justification, that the operationalizations tested were imperfect representations of the theories’ deepest commitments. This response is available because IIT and GNW are mathematically rich enough to generate many distinct empirical trajectories, and one contested trajectory does not falsify the whole.

What the results do is change the epistemic status of both frameworks. Neither can be described, after 2025, as empirically validated in the way that would justify treating phi or global broadcast as established correlates of consciousness. They are theoretical frameworks that generate testable predictions, and in the most demanding test yet conducted, the predictions did not hold across the board. That is a different kind of empirical position from the one both theories occupied before the Consortium published.

The methodological shift visible in 2026 consciousness research reflects this. The move from binary consciousness scores toward multidimensional profiles is in part a post-Cogitate response: if the field cannot be confident any single theoretical framework is correct, then consciousness assessments that bet everything on one framework are epistemically fragile. Butlin et al.’s indicator approach, PRISM’s methodological agnosticism, and Birch’s centrist manifesto all work from the premise that the theoretical landscape is genuinely contested, which is precisely what the Cogitate results confirm.

What This Means for AI Consciousness Research

The implications for artificial consciousness research are specific rather than diffuse, and they apply to both the positive and negative sides of the debate.

On the affirmative side, scholars who use IIT’s phi to argue that a given AI architecture has measurable consciousness are working with a metric whose biological grounding is empirically contested. Phi may still capture something real about the causal structure of information processing in artificial systems. But connecting phi to consciousness requires a theoretical step that the Cogitate data have made more contested, not less.

On the skeptical side, scholars who use GNW to argue that AI systems lack consciousness because they lack genuine global broadcast are relying on a mechanism that, even in biological systems, does not produce the full predicted pattern. The absence of biological-style global broadcast in an LLM may still be a legitimate reason to doubt machine consciousness. But it is a weaker argument than it was before the Consortium reported.

The broader field is now in a position that the current scientific debate on AI consciousness captures in its ongoing form: the most influential theoretical frameworks are empirically challenged simultaneously, and the question of what it would take to assess consciousness in an artificial system is open in a deeper way than it appeared before 2025. Welfare advocates, alignment scholars, and governance bodies who build their assessments on IIT or GNW foundations face a field that has explicitly tested and complicated both.

That is not a reason to abandon theoretical grounding for AI consciousness research. It is a reason to hold those theoretical grounds with the epistemic humility that the Cogitate Consortium’s own proponents, having run the test, were compelled to exercise.

The 2025 Nature paper is not the end of either framework. It is the clearest signal yet that the field needs methodological approaches capable of testing mechanisms that IIT and GNW, in their current forms, do not fully anticipate, and that any assessment of consciousness in artificial systems should begin by accounting for how much has remained unsettled in the biological case.