When IIT and GNW Were Put to the Test: What the Cogitate Consortium Found
For most of the last two decades, two theories have dominated empirical consciousness research and, by extension, the methodology used to evaluate whether artificial systems might be conscious. Integrated Information Theory, developed by Giulio Tononi and colleagues at the University of Wisconsin-Madison, proposes that consciousness is identical to integrated information, a quantity denoted phi. Global Neuronal Workspace Theory, developed principally by Stanislas Dehaene (Collège de France and CEA) and Bernard Baars, holds that consciousness arises when information from specialized processing modules is broadcast globally across the brain and becomes available to all cognitive systems simultaneously.
Both theories have been applied extensively to AI consciousness research. IIT’s phi calculation has been used to ask whether artificial neural networks have sufficient integration to support consciousness. GNW architecture analysis has been used to evaluate whether transformer models, with their attention mechanisms providing something like global broadcast, have the computational structure that consciousness requires.
In 2025, both theories were subjected to the most rigorous empirical test either had faced. The Cogitate Consortium, a team of 256 researchers organized around a preregistered adversarial collaboration design, ran simultaneous fMRI, MEG, and intracranial EEG studies across multiple laboratories. The theoretical predictions of both IIT and GNW were specified in advance by the theories’ own proponents, Tononi’s group for IIT and Dehaene’s group for GNW, and the experimenters agreed before running the study on what results would count as evidence for and against each theory.
The study was published in Nature (doi: 10.1038/s41586-025-08888-1) in 2025. Neither theory emerged with its core predictions fully confirmed. For anyone working on AI consciousness, the implications are substantial.
What Adversarial Collaboration Means Here
The term “adversarial collaboration” was introduced by Daniel Kahneman and others as a method for resolving disputes between theories that have strong proponents in the scientific community. In a standard adversarial collaboration, researchers who hold competing theories agree to design an experiment together, specify in advance what results would support or challenge each theory, and accept the results as evidence regardless of the outcome.
The Cogitate project applied this method to one of the most contested disputes in consciousness science. The adversarial design has a specific epistemic virtue: it removes the ability of theorists to reinterpret results post hoc. If you specify in advance that “IIT predicts sustained posterior synchronization and GNW predicts a specific ignition signature at stimulus onset,” then a study that finds neither cannot simply be reinterpreted as consistent with both. The pre-registration locks the predictions before the data arrives.
This matters because consciousness theories have historically been frustratingly flexible in the face of negative evidence. Both IIT and GNW have accumulated large bodies of supporting evidence, but they have also accumulated critical responses that each camp tends to reinterpret as consistent with the underlying framework. The Cogitate design was explicitly constructed to prevent that reinterpretation.
What IIT Predicted and What Was Found
Integrated Information Theory identifies the posterior hot zone, a region spanning the posterior parietal and occipital cortices, as the primary neural correlate of consciousness. The theory predicts that when a subject is consciously aware of a stimulus, the posterior cortex should exhibit sustained, high-integration activity throughout the period of conscious perception. The integration is specifically across the posterior network: information from different sensory modalities and representations should be bound together in a way that corresponds to unified experience.
The IIT prediction tested in Cogitate was operationalized as sustained posterior neural activity that persists for the duration of conscious perception, with specific signatures of integration across posterior cortical regions.
The study found that this prediction was challenged. While some posterior activation was present in conditions associated with conscious perception, the sustained, integrated posterior synchronization that IIT specifically predicted was not reliably present across the range of stimuli and conditions tested. The signal that IIT identified as the hallmark of consciousness did not appear with the consistency and magnitude the theory requires.
Tononi and colleagues did not concede that IIT is refuted. They argued that the stimuli used in Cogitate did not adequately engage the posterior hot zone under conditions that IIT specifies as necessary for high-phi integration. The dispute about whether the test adequately operationalized IIT’s predictions continued into 2026. But the adversarial collaboration design means this cannot simply be called a null result from a study that was designed by IIT’s opponents. The test was designed by IIT’s proponents.
What GNW Predicted and What Was Found
Global Neuronal Workspace Theory predicts that conscious perception involves a specific dynamic: when a stimulus becomes conscious, there is a sudden, widespread “ignition” event in which prefrontal and parietal regions broadcast the stimulus representation globally. This ignition is the moment of global availability. The theory further predicts that this ignition should be absent for stimuli that are processed subliminally: the stimulus is handled locally, in modality-specific regions, without reaching the prefrontal broadcast stage.
The GNW predictions tested in Cogitate focused on the ignition signature: its presence for conscious stimuli, its absence for non-conscious stimuli, and the role of prefrontal cortex in maintaining the globally broadcast representation.
The study found that GNW’s predictions were also partially challenged. The ignition pattern was not reliably found at stimulus offset, a point where GNW specifically predicts that the broadcast should cease in a detectable way. The representation of certain conscious dimensions in the prefrontal cortex was weaker than the theory requires. The global broadcast signature that GNW identifies as the computational basis of consciousness did not appear in the form the theory specifies.
Dehaene and colleagues similarly disputed the interpretation, arguing that some experimental conditions in Cogitate were better suited to testing IIT’s predictions than GNW’s. The back-and-forth following the paper’s publication has been substantive: several response papers appeared in 2025 and 2026, with theorists on both sides arguing about the operationalization choices.
The Measurement Problem in AI Consciousness
The Cogitate results have an implication for AI consciousness research that has not yet been fully absorbed by the field.
The two most commonly used frameworks for evaluating whether artificial systems might be conscious are derived from IIT and GNW. When researchers apply phi calculations to artificial neural networks, they are asking whether the system has sufficient integrated information to satisfy IIT’s consciousness criterion. When researchers evaluate transformer architectures for GWT-compatible properties, asking whether the attention mechanism constitutes a global workspace and whether information is broadcast across the system, they are applying GNW’s framework.
If both IIT and GNW have their core predictions challenged in their home domain, the biological brain, this raises a direct question about the AI applications. If the phi calculation is not reliably identifying the neural correlate of consciousness in biological systems, the fact that an AI system has high phi is less informative than it appeared. If the GWT architecture analysis is not reliably distinguishing conscious from non-conscious processing in biological systems, applying it to AI architectures becomes methodologically uncertain.
The scores versus profiles debate about AI consciousness measurement already identified that reducing consciousness to a single score versus a multidimensional profile involves fundamental trade-offs. Cogitate adds a layer to that debate: it raises the question of whether the theories that generate those measurements are themselves on solid empirical footing.
This does not mean IIT and GNW are worthless for AI consciousness research. Both theories captured genuine phenomena. IIT’s phi does measure something about information integration. GWT’s global broadcast analysis does capture something about how information becomes widely available in a cognitive system. The question is whether those somethings are identical to consciousness, and Cogitate’s results make that identification more uncertain than it previously appeared.
What Neither Side Claims
It is important to be precise about what the Cogitate results do not establish.
They do not establish that IIT is false. A failure to find the predicted posterior synchronization in the conditions tested is not proof that consciousness is not integrated information. It may mean the stimuli were wrong, the measurement resolution was insufficient, or the operationalization of IIT’s predictions was incomplete.
They do not establish that GNW is false. A failure to find the predicted ignition signature in some conditions is not proof that conscious access does not involve global broadcast. It may mean the experimental design was better suited to some aspects of GNW than others.
What the results do establish is that neither theory’s most central predictions held up under the conditions of the first preregistered adversarial test. That is different from refutation, but it is also different from confirmation. The theories need revision, further specification, or constraints on the conditions under which their predictions are expected to hold.
What the Field Has Done Since
The 2026 response to Cogitate has produced a range of reactions. Some researchers have argued that the results motivate moving beyond the IIT-GNW framing altogether and developing theories that are agnostic about where in the brain consciousness arises. Others have argued for hybrid frameworks that combine IIT’s integration criterion with GWT’s broadcast criterion. A third group has argued that the adversarial design itself was flawed and that a better-designed test would have produced more decisive results.
The Brock University and Institute of Noetic Sciences research on applying IIT equations to artificial systems is proceeding alongside this debate. That work treats the phi calculation as a useful tool even in the context of the Cogitate findings, on the grounds that a challenged theory is not a useless theory. The 14 indicator checklist developed by Patrick Butlin, Robert Long, and colleagues draws on multiple theoretical frameworks including IIT and GNW, and the Cogitate results add a layer of uncertainty to the indicators derived specifically from those frameworks.
The testing of consciousness theories in artificial systems has always faced the problem that no single theory commands consensus. Cogitate has made that lack of consensus empirically salient in a new way: it is now possible to point to a specific study, designed by the theories’ own proponents, that found their core predictions challenged. That is a different kind of uncertainty than theoretical disagreement.
The Methodological Gap This Opens for AI
The most productive reading of Cogitate for AI consciousness research is not “IIT and GNW are wrong, so we cannot evaluate AI consciousness.” It is “the measurement tools most commonly used for AI consciousness evaluation rest on theories whose empirical foundations are more uncertain than we assumed. We need either better-validated theories or methods that are less theory-dependent.”
This is the problem that Palminteri and Wu’s behavioral inference principle, discussed separately in a companion piece, is attempting to address: it proposes a methodology for attributing consciousness that does not depend on identifying the correct theory of consciousness in advance, but instead asks whether consciousness attribution is useful for predicting and explaining an entity’s behavior.
The Cogitate result and the behavioral inference proposal, taken together, suggest that consciousness science in 2026 is in a phase where the dominant theoretical frameworks are being empirically tested and revised, and where the AI consciousness field will need to either wait for that revision to settle or develop methodologies that can operate under theoretical uncertainty.
Neither option is fully satisfying. But the Cogitate result makes honest engagement with both options unavoidable.