KnowledgeInSight
Phenomenology for Consciousness Studies
0% of Course 4 complete

Module 2 · Lesson 3

Hard Problem: Scientific Theories

This lesson surveys four leading scientific theories of consciousness:

What you will be able to do

  • Describe the aims and commitments of global workspace, integrated information, higher-order, and predictive processing theories

0% of this lesson · 9 items · 1h 3m total · 48m without the optional journal

Contents of this lesson9 items
  1. ReadingTesting Theories of Consciousness: Stakes in the Clinic and in AI3 min
  2. ReadingGlobal Workspace and Integrated Information Theories of Consciousness4 min
  3. ReadingHigher-Order and Predictive Processing Theories of Consciousness4 min
  4. ReadingAdversarial Collaboration and the Cogitate Test of GNWT and IIT4 min
  5. ReadingThe IIT Pseudoscience Controversy and Four Theories Compared4 min
  6. Guided ReadingGuided Walkthrough: Reading the Cogitate Adversarial Collaboration7 min
  7. Guided ConversationCompare the Theories12 min
  8. Journal · optionalJournal Entry on Theories of Consciousness15 min
  9. Knowledge CheckScientific Theories of Consciousness10 min

Reading 3 min

Testing Theories of Consciousness: Stakes in the Clinic and in AI

Consciousness science has no shortage of theories. Reviews of the field list dozens, and several have large research programs behind them. What it has lacked is decisive tests. Rival theories often explain the same data in their own terms, and their proponents rarely agree on what result would show them wrong.

That would be an academic problem if nothing depended on it. A good deal does.

Take patients with severe brain injuries who don't respond to commands. A large multicenter study in 2024 used brain imaging and electrophysiology to ask such patients to imagine movements. About a quarter of those who showed no behavioral response produced brain activity that followed the commands (Bodien et al. 2024). The finding is called cognitive motor dissociation. It means that bedside behavior can miss awareness. Deciding what to look for instead requires some view of what consciousness is and where it arises in the brain.

Anesthesia raises a related question every day in operating rooms: when is a patient unconscious, and how would we know? So does artificial intelligence. A 2023 report by a group of researchers derived "indicator properties" of consciousness from leading scientific theories and assessed current AI systems against them. It concluded that no current system was a strong candidate for consciousness, but that there were no obvious technical barriers to building systems that satisfy the indicators (Butlin et al. 2023). That conclusion is only as good as the theories behind it. If the theories disagree about which properties matter, the same system can pass one test and fail another.

One response to the testing problem has drawn wide attention. In an adversarial collaboration, proponents of rival theories agree in advance on an experiment and on which results would support or challenge each theory, and they register those predictions publicly before data are collected. In 2025, Nature published the results of one such collaboration, and an editorial argued that the time for this approach had come (Nature 2025).

For a researcher, the skill at stake is reading a theory for its commitments. You need to know what a theory predicts, which predictions are central and which are auxiliary, and what would count against it. Without that, you can't tell a test of a theory from a test of one researcher's version of it, or a clinical or ethical claim that rests on a theory from one that doesn't. This content reflects the field as of September 2026.

References

  • Bodien, Yelena G., Judith Allanson, Paolo Cardone, Arthur Bonhomme, Jerina Carmona, Camille Chatelle, Srivas Chennu, et al. 2024. "Cognitive Motor Dissociation in Disorders of Consciousness." New England Journal of Medicine 391 (7): 598–608.
  • Butlin, Patrick, Robert Long, Eric Elmoznino, Yoshua Bengio, Jonathan Birch, Axel Constant, George Deane, et al. 2023. "Consciousness in Artificial Intelligence: Insights from the Science of Consciousness." arXiv:2308.08708.
  • Nature. 2025. "Make Science More Collegial: Why the Time for 'Adversarial Collaboration' Has Come." Editorial. Nature 641 (8062): 281–82.

Report an issue with this item

Reading 4 min

Global Workspace and Integrated Information Theories of Consciousness

Introduction

Two theories dominate current neuroscience of consciousness, and they disagree about nearly everything that matters: what consciousness is, where in the brain it arises, and what a theory should start from. That disagreement is what makes them testable against each other.

This reading explains global workspace theory, in its original and neuronal forms, and integrated information theory. It also notes one related view, recurrent processing theory. This content reflects the field as of September 2026.

Global Workspace Theory

Bernard Baars proposed the global workspace theory in 1988 (Baars 1988). The mind, on his account, contains many specialized processors working in parallel and outside awareness. They handle vision, language, memory, and motor control. Consciousness depends on a global workspace: in Baars's theory, a limited-capacity system whose contents are broadcast to many specialized processors at once, so that they become available throughout the cognitive system.

Baars used a theater image. The workspace is a lit stage. Attention is a spotlight. The specialized processors are the audience in the dark. What's on stage is conscious because everyone in the audience receives it.

Stanislas Dehaene and his colleagues turned the idea into a neural theory (Dehaene 2014, ch. 5). The global neuronal workspace is Dehaene's neural version of global workspace theory, in which conscious access is the brain-wide broadcast of information through long-range networks centered on prefrontal and parietal cortex.

The theory makes a distinctive claim about timing. When a stimulus becomes conscious, activity doesn't just grow. It undergoes ignition: in global neuronal workspace theory, a sudden, self-sustaining surge of activity across the workspace network that marks the moment information becomes conscious. Before ignition, a stimulus can be processed without being consciously seen. After it, the stimulus is available for report, memory, and reasoning.

Notice what the theory targets: conscious access, the availability of information for use throughout the system. Its critics ask whether that is the same as experience.

Integrated Information Theory

Integrated information theory (IIT), developed by Giulio Tononi, starts at the other end (Tononi 2004). It doesn't begin from brain data or from cognitive function. It begins from experience itself.

In its current version, IIT 4.0, the theory first states what it takes to be essential properties of every experience (Albantakis et al. 2023). These are its axioms: for example, that experience exists for itself, is specific, is unified, and is definite. It then asks what a physical system must be like to account for those properties. The answers are its postulates, stated in terms of a system's cause–effect power: how its parts constrain its own past and future states.

The theory's central claim is an identity. An experience is identical to the cause–effect structure of a physical substrate that has maximal integrated information: in IIT, the degree to which a system's causal structure is unified and irreducible to that of its parts. The quantity of that integration is written Φ (phi): in IIT, the quantity of integrated information, which the theory uses to measure how much consciousness a system has.

Two commitments follow that matter for testing. First, what counts is causal structure, not what a system does. Two systems that perform the same functions could differ in consciousness. Second, IIT locates the main substrate of human experience in posterior cortex, in parietal, temporal, and occipital regions, rather than in prefrontal cortex.

Global neuronal workspaceIntegrated information theory
Starts fromCognitive function and brain dataProperties of experience
Consciousness isGlobal availability of informationIntegrated causal structure
Key regionPrefrontal and parietal networksPosterior cortex
Key signatureIgnition and broadcastSustained integration within posterior cortex

A third view sits between them. Recurrent processing theory holds that experience arises when sensory areas engage in local feedback loops, without global broadcast (Lamme 2006).

Conclusion

Global workspace theory treats consciousness as global availability: contents broadcast from a limited-capacity workspace. Its neuronal version ties this to prefrontal and parietal networks and to ignition. Integrated information theory starts from properties of experience and identifies consciousness with integrated causal structure, measured by Φ and located mainly in posterior cortex.

Key Terms

  • Global workspace: In Baars's theory, a limited-capacity system whose contents are broadcast to many specialized processors at once, so that they become available throughout the cognitive system.
  • Global neuronal workspace: Dehaene's neural version of global workspace theory, in which conscious access is the brain-wide broadcast of information through long-range networks centered on prefrontal and parietal cortex.
  • Ignition: In global neuronal workspace theory, a sudden, self-sustaining surge of activity across the workspace network that marks the moment information becomes conscious.
  • Integrated information: In IIT, the degree to which a system's causal structure is unified and irreducible to that of its parts.
  • Φ: In IIT, the quantity of integrated information, which the theory uses to measure how much consciousness a system has.

References

  • Albantakis, Larissa, Leonardo Barbosa, Graham Findlay, Matteo Grasso, Andrew M. Haun, William Marshall, William G. P. Mayner, et al. 2023. "Integrated Information Theory (IIT) 4.0: Formulating the Properties of Phenomenal Existence in Physical Terms." PLOS Computational Biology 19 (10): e1011465.
  • Baars, Bernard J. 1988. A Cognitive Theory of Consciousness. Cambridge: Cambridge University Press.
  • Dehaene, Stanislas. 2014. Consciousness and the Brain: Deciphering How the Brain Codes Our Thoughts. New York: Viking.
  • Lamme, Victor A. F. 2006. "Towards a True Neural Stance on Consciousness." Trends in Cognitive Sciences 10 (11): 494–501.
  • Tononi, Giulio. 2004. "An Information Integration Theory of Consciousness." BMC Neuroscience 5:42.

Report an issue with this item

Reading 4 min

Higher-Order and Predictive Processing Theories of Consciousness

Introduction

Not every theory of consciousness locates it in a broadcast or a causal structure. Higher-order theories locate it in a relation between mental states. Predictive processing, a general account of how brains work, has been developed into an account of experience as the brain's best guess.

This reading explains higher-order theories, in Rosenthal's version, and predictive processing, in Clark's framework and Seth's account of perception and self. This content reflects the field as of September 2026.

Higher-Order Theories

Higher-order theories start from an intuitive thought (Gennaro 2026). A conscious state is one you're aware of being in. An unconscious perception, such as seeing an obstacle you avoid without noticing it, is a state you're in but not aware of being in. What makes the difference?

The answer is a higher-order representation: a mental state that represents another of the subject's own mental states; on higher-order theories, a state is conscious when it is suitably represented by one. The first-order state represents the world: the red tomato. The higher-order state represents the first-order state: that I am seeing red.

David Rosenthal's version is the best known (Rosenthal 1997). On his account, the higher-order state is a higher-order thought: in Rosenthal's theory, an assertoric thought, not arrived at by conscious inference, to the effect that one is in a certain mental state. The higher-order thought needn't be conscious itself. It becomes conscious only if a further thought represents it, as in introspection.

The shift is from the state's content to the state's being represented. On a higher-order theory, nothing about the first-order state alone makes it conscious.

Higher-order theories face a well-known objection (Gennaro 2026). What if a higher-order thought represents a state you aren't actually in? The theory seems to imply that you'd have an experience with no first-order state to be the experience. Higher-order theorists have offered several replies, and the issue remains open.

Predictive Processing

Predictive processing starts from a different problem: how does a brain, sealed in a skull, make sense of the signals that reach it? Andy Clark's answer is that the brain is a prediction machine (Clark 2013, 2016).

Predictive processing is the view that the brain continually predicts its sensory input and updates its predictions using the differences between prediction and input, so that perception is the brain's best guess about the causes of its signals. The brain maintains a layered model of the world. Higher layers send predictions down. Lower layers send back prediction error: the difference between the input the brain predicts and the input it receives, which is used to revise predictions. Perception is the prediction that best explains away the error. Action, on Clark's account, can reduce error too, by changing the input to fit the prediction (Clark 2016).

Predictive processing is a general framework for perception, action, and cognition. It isn't, by itself, a theory of consciousness. Anil Seth has developed it into one (Seth 2021).

Seth: Controlled Hallucination and the Beast Machine

Seth describes perception as a controlled hallucination: Seth's phrase for perception as the brain's predictions about the causes of its sensory signals, held in check by those signals. Hallucination, on this view, is uncontrolled perception: prediction that has come loose from the input. Ordinary perception is the same process, kept on track.

Seth extends the account to the self. The beast machine is Seth's name for his account of the self as grounded in the brain's predictive regulation of the body's internal states to keep the organism alive. The deepest layer of selfhood, on this view, is the feeling of being a living body. It comes from predictions about heartbeat, breathing, and other signals from inside the body, which the brain must keep within bounds.

Seth doesn't claim to solve the hard problem directly. He proposes to explain, predict, and control specific properties of experience, and to see whether the hard problem looks different once that's done (Seth 2021).

Conclusion

Higher-order theories hold that a state is conscious when a suitable higher-order representation, in Rosenthal's version a higher-order thought, represents it. Predictive processing treats perception as the brain's best guess, revised by prediction error. Seth develops it into an account of perception as controlled hallucination and of the self as a beast machine.

Key Terms

  • Higher-order representation: A mental state that represents another of the subject's own mental states; on higher-order theories, a state is conscious when it is suitably represented by one.
  • Higher-order thought: In Rosenthal's theory, an assertoric thought, not arrived at by conscious inference, to the effect that one is in a certain mental state.
  • Predictive processing: The view that the brain continually predicts its sensory input and updates its predictions using the differences between prediction and input, so that perception is the brain's best guess about the causes of its signals.
  • Prediction error: The difference between the input the brain predicts and the input it receives, which is used to revise predictions.
  • Controlled hallucination: Seth's phrase for perception as the brain's predictions about the causes of its sensory signals, held in check by those signals.
  • Beast machine: Seth's name for his account of the self as grounded in the brain's predictive regulation of the body's internal states to keep the organism alive.

References

Report an issue with this item

Reading 4 min

Adversarial Collaboration and the Cogitate Test of GNWT and IIT

Introduction

For decades, theories of consciousness were tested mostly by their own proponents. Each lab ran experiments that suited its theory, and results rarely changed anyone's mind. In 2025, Nature published the results of a different kind of study, in which the two leading theories were tested against each other by design.

This reading explains the adversarial collaboration model and the Cogitate consortium's test of global neuronal workspace theory (GNWT) and integrated information theory (IIT). Citations give the article, which is open access. This content reflects the field as of September 2026.

The Adversarial Collaboration Model

An adversarial collaboration is a research design in which proponents of rival theories agree in advance on experiments and on which results would support or challenge each theory. The approach was developed in psychology as an alternative to disputes carried on through competing papers (Nature 2025).

Three features make it work.

  1. Agreed predictions. The rival theorists must commit to predictions that differ. A result that both theories predict can't decide between them.
  2. Preregistration: publicly recording hypotheses, methods, and analyses before collecting data, so that predictions can't be adjusted to fit the results.
  3. Separation of roles. Theory-neutral teams collect and analyze the data. The theorists design the test but don't run it.

The editorial accompanying the published results argued that the approach works only if everyone involved accepts that they might be wrong (Nature 2025).

The Cogitate Study

The Cogitate consortium tested GNWT against IIT (Cogitate Consortium et al. 2025). Leaders of both theories agreed on the design and on three sets of diverging predictions, which were preregistered.

The experiments tested 256 participants with three methods: functional MRI, magnetoencephalography (MEG), and intracranial recordings from patients with implanted electrodes. Each method was run in more than one laboratory. Participants saw clearly visible images of faces, objects, letters, and meaningless symbols, shown at different orientations and for different durations. Some image categories were relevant to a task and others weren't. That manipulation aimed to separate the neural basis of being conscious of an image from the neural basis of doing something with it.

The predictions concerned three questions.

  1. Where is conscious content represented? IIT predicted it would be found mainly in posterior cortex. GNWT predicted it would be found in prefrontal cortex. The main test was decoding: using a statistical classifier to read out which stimulus feature, such as a category or orientation, is represented in a pattern of brain activity.
  2. How is content maintained over time? IIT predicted sustained activity in posterior cortex lasting as long as the image was seen. GNWT predicted brief ignition in prefrontal cortex when the image appeared and again when it disappeared.
  3. How do regions communicate? IIT predicted sustained synchronization within posterior cortex. GNWT predicted brief, long-range synchronization between prefrontal cortex and category-selective areas.

What the Results Showed

The authors report that the results aligned with some predictions of each theory while "substantially challenging key tenets of both" (Cogitate Consortium et al. 2025). For IIT, the key challenge was the lack of sustained synchronization within posterior cortex. For GNWT, it was the general lack of ignition when images disappeared and the limited representation of some conscious features, such as orientation, in prefrontal cortex.

The paper doesn't declare a winner. Its discussion gives separate interpretations from the IIT and GNWT proponents and from the neutral authors. That format concedes a point about testing. A theory's proponents can respond to a failed prediction by revising an auxiliary assumption: a supporting hypothesis, such as a claim about which brain signals track a theory's quantities, that links a theory's core claims to observable results. Whether a failed prediction counts against the core of a theory, or only against an auxiliary, is itself open to dispute.

Conclusion

In an adversarial collaboration, rival theorists agree on diverging predictions in advance, preregister them, and leave data collection to neutral teams. The Cogitate study tested GNWT and IIT on where content is represented, how it is maintained, and how regions communicate. Its results challenged predictions of both theories, and its proponents disagree about what follows.

Key Terms

  • Adversarial collaboration: A research design in which proponents of rival theories agree in advance on experiments and on which results would support or challenge each theory.
  • Preregistration: Publicly recording hypotheses, methods, and analyses before collecting data, so that predictions can't be adjusted to fit the results.
  • Decoding: Using a statistical classifier to read out which stimulus feature, such as a category or orientation, is represented in a pattern of brain activity.
  • Auxiliary assumption: A supporting hypothesis, such as a claim about which brain signals track a theory's quantities, that links a theory's core claims to observable results.

References

  • Cogitate Consortium, Oscar Ferrante, Urszula Gorska-Klimowska, Simon Henin, Rony Hirschhorn, Aya Khalaf, Alex Lepauvre, et al. 2025. "Adversarial Testing of Global Neuronal Workspace and Integrated Information Theories of Consciousness." Nature 642 (8066): 133–42.
  • Nature. 2025. "Make Science More Collegial: Why the Time for 'Adversarial Collaboration' Has Come." Editorial. Nature 641 (8062): 281–82.

Report an issue with this item

Reading 4 min

The IIT Pseudoscience Controversy and Four Theories Compared

Introduction

In 2023, a public dispute broke out over whether one of the leading theories of consciousness counts as science at all. The dispute is still unresolved, and it has become a debate about the standards any theory of consciousness should meet.

This reading sets out the controversy over integrated information theory (IIT), presented as a debate about standards of evidence, then compares the four major theories side by side. This content reflects the field as of September 2026.

The Controversy

The sequence of events runs as follows.

  1. The 2023 open letter. After preliminary results of the adversarial test of IIT and global neuronal workspace theory were publicized in 2023, a group of more than a hundred researchers posted an open letter calling IIT pseudoscience (IIT-Concerned et al. 2023). They argued that the tested predictions weren't logically tied to the theory's core ideas, and that media coverage had presented the theory as empirically supported. They also pointed to implications they found implausible, such as that an inactive grid of simple logic gates could be conscious.
  2. Responses. Other researchers, including IIT's proponents, objected. They argued that the label was unwarranted and that it damaged scientific debate.
  3. The 2025 commentary. The letter's authors developed their case in Nature Neuroscience (IIT-Concerned Consortium et al. 2025). They argued that IIT is unscientific because its core claims are untestable even in principle.
  4. The editorial. Publishing the adversarial test's results, Nature endorsed adversarial collaboration and criticized the letter's language. Such language, it said, has no place in a process meant to build working relationships between rival groups (Nature 2025).
  5. The reply. Several of the letter's authors answered in a Nature Correspondence (Snyder et al. 2025). They defended the label as carefully reasoned and argued that all theories must be subject to critical evaluation.

A Debate about Standards of Evidence

The disputants don't mainly disagree about data. They disagree about what a scientific theory of consciousness must offer. Three questions divide them.

  • What must be testable? The critics hold that a theory's core claims, not only its auxiliary predictions, must be empirically testable. IIT's defenders hold that the theory generates testable predictions and should be judged by them, as other theories are.
  • How should implications count? The critics treat counterintuitive implications as evidence against the theory. Defenders reply that intuitions about which systems are conscious aren't evidence either way.
  • What does the label do? The critics hold that calling a theory pseudoscience can be a justified judgment. The editorial held that it undermines the collegial process that adversarial testing needs.

Four Theories Compared

The table compares the four theories (Morales 2024; Cogitate Consortium et al. 2025).

Global workspaceIntegrated informationHigher-orderPredictive processing
What it explainsConscious access: global availability of informationExperience, starting from its essential propertiesWhat makes a state conscious rather than unconsciousPerception and self as the brain's best guess
Core mechanismBroadcast through prefrontal–parietal networks, with ignitionIntegrated causal structure in posterior cortexA higher-order representation of a first-order statePrediction and prediction error in a layered model
Key predictionIgnition and prefrontal involvement in conscious contentSustained posterior integration during experienceConsciousness tracks higher-order representation, not first-order contentExpectations shape what is perceived
Main criticismExplains access, perhaps not experienceCore claims may be untestable; counterintuitive implicationsStates can be misrepresented by higher-order statesA framework for perception, not yet a specific theory of experience

Core Terms at a Glance

TermDefinition
Global workspaceIn Baars's theory, a limited-capacity system whose contents are broadcast to many specialized processors at once, so that they become available throughout the cognitive system.
IgnitionIn global neuronal workspace theory, a sudden, self-sustaining surge of activity across the workspace network that marks the moment information becomes conscious.
Integrated informationIn IIT, the degree to which a system's causal structure is unified and irreducible to that of its parts.
ΦIn IIT, the quantity of integrated information, which the theory uses to measure how much consciousness a system has.
Higher-order representationA mental state that represents another of the subject's own mental states; on higher-order theories, a state is conscious when it is suitably represented by one.
Predictive processingThe view that the brain continually predicts its sensory input and updates its predictions using the differences between prediction and input, so that perception is the brain's best guess about the causes of its signals.
Adversarial collaborationA research design in which proponents of rival theories agree in advance on experiments and on which results would support or challenge each theory.

Conclusion

Critics called IIT pseudoscience in 2023 and argued in 2025 that its core claims are untestable. Its defenders and Nature's editorial rejected the label, and the critics replied that every theory must face critical evaluation. The dispute turns on standards: what must be testable, how implications count, and what labels do.

Key Terms

  • Global workspace: In Baars's theory, a limited-capacity system whose contents are broadcast to many specialized processors at once, so that they become available throughout the cognitive system.
  • Ignition: In global neuronal workspace theory, a sudden, self-sustaining surge of activity across the workspace network that marks the moment information becomes conscious.
  • Integrated information: In IIT, the degree to which a system's causal structure is unified and irreducible to that of its parts.
  • Φ: In IIT, the quantity of integrated information, which the theory uses to measure how much consciousness a system has.
  • Higher-order representation: A mental state that represents another of the subject's own mental states; on higher-order theories, a state is conscious when it is suitably represented by one.
  • Predictive processing: The view that the brain continually predicts its sensory input and updates its predictions using the differences between prediction and input, so that perception is the brain's best guess about the causes of its signals.
  • Adversarial collaboration: A research design in which proponents of rival theories agree in advance on experiments and on which results would support or challenge each theory.

References

  • Cogitate Consortium, Oscar Ferrante, Urszula Gorska-Klimowska, Simon Henin, Rony Hirschhorn, Aya Khalaf, Alex Lepauvre, et al. 2025. "Adversarial Testing of Global Neuronal Workspace and Integrated Information Theories of Consciousness." Nature 642 (8066): 133–42.
  • IIT-Concerned Consortium, Michał Klincewicz, Tony Cheng, Michael Schmitz, Miguel Ángel Sebastián, Joel S. Snyder, Derek H. Arnold, et al. 2025. "What Makes a Theory of Consciousness Unscientific?" Nature Neuroscience 28 (4): 689–93.
  • IIT-Concerned, Stephen M. Fleming, Chris D. Frith, Mel Goodale, Hakwan Lau, Joseph E. LeDoux, Alan L. F. Lee, et al. 2023. "The Integrated Information Theory of Consciousness as Pseudoscience." PsyArXiv, September 16, 2023.
  • Morales, Jorge. 2024. "The Neuroscience of Consciousness." In The Stanford Encyclopedia of Philosophy, edited by Edward N. Zalta and Uri Nodelman. https\://plato.stanford.edu/entries/consciousness-neuroscience/.
  • Nature. 2025. "Make Science More Collegial: Why the Time for 'Adversarial Collaboration' Has Come." Editorial. Nature 641 (8062): 281–82.
  • Snyder, Joel S., Tony Cheng, Michał Klincewicz, Michael Schmitz, and Miguel Ángel Sebastián. 2025. "Adversarial Collaborations: All Theories Must Be Subject to Critical Evaluation." Correspondence. Nature 642 (8069): 867.
    • Free: None (paywalled).
    • Publisher: Nature

Report an issue with this item

Guided Reading 7 min

Guided Walkthrough: Reading the Cogitate Adversarial Collaboration

Introduction

A study that tests two theories against each other can be misread in two opposite ways: as settling which theory is true, or as showing nothing because both sides can explain it away. Reading it well means keeping track of what each theory committed to in advance. This reading works through the Cogitate consortium's test of global neuronal workspace theory (GNWT) and integrated information theory (IIT) in four steps, flags a common pitfall, and ends with an annotated table matching predictions to results. This content reflects the field as of September 2026.

Locating the Study

The study is "Adversarial Testing of Global Neuronal Workspace and Integrated Information Theories of Consciousness" (Cogitate Consortium et al. 2025). It's open access on the Nature website, so you can read along.

The article follows the journal's format. The introduction sets out the theories and their preregistered predictions. The results are organized by prediction. The discussion gives separate interpretations from the IIT proponents, the GNWT proponents, and the theory-neutral authors. The methods, at the end, describe the participants, stimuli, and analyses in detail. Here is a summary of the design:

  • Participants. 256 people, tested with functional MRI, magnetoencephalography (MEG), or intracranial electrodes implanted for clinical reasons.
  • Stimuli. Clearly visible images of faces, objects, letters, and meaningless symbols, at three orientations and three durations (half a second to one and a half seconds).
  • Task. Participants watched for rare target images. Some categories were relevant to the task and others weren't.
  • Sites. Each method was run in more than one independent laboratory.

Walking Through the Study

Step 1: Identify each theory's preregistered predictions

Start with the predictions, before looking at any result. The two theories committed to diverging predictions on three questions.

  1. Where is conscious content represented? IIT: mainly in posterior cortex. GNWT: prefrontal cortex must be involved.
  2. How is content maintained while an image stays in view? IIT: sustained activity in posterior cortex for as long as the image is seen. GNWT: brief ignition in prefrontal cortex at the image's onset and again at its offset, with the content maintained silently between.
  3. How do regions communicate? IIT: sustained, high-frequency synchronization within posterior cortex, linking early visual areas to category-selective areas. GNWT: brief, long-range synchronization between prefrontal cortex and category-selective areas.

Notice that each prediction follows from a core commitment. GNWT ties consciousness to prefrontal broadcast and ignition. IIT ties it to integrated causal structure in posterior cortex. That's what makes the predictions a test of the theories, not only of their proponents.

Step 2: Identify the methods and the task

Next, ask what each method can show. Functional MRI shows where activity occurs, but it's slow. MEG shows timing, but it locates activity less precisely. Intracranial recordings show both, but only where electrodes happen to be.

Then ask what the task does. The images were all clearly visible, so participants were conscious of them. Varying task relevance separates being conscious of an image from using it for the task. A region that responds only when an image is task-relevant may track report or decision, not consciousness. Varying duration tests Prediction 2: whether activity lasts as long as the experience.

Step 3: Match results to predictions

Now read the results against the predictions (Cogitate Consortium et al. 2025).

  • Prediction 1. Image category could be decoded from both posterior and prefrontal regions. Orientation could be decoded from posterior cortex but not prefrontal cortex. This fits IIT's emphasis on posterior cortex and challenges GNWT's claim that prefrontal cortex represents conscious content.
  • Prediction 2. Posterior cortex showed content that was sustained for as long as the image stayed on screen, as IIT predicted. Prefrontal cortex generally didn't show the predicted ignition at offset, which challenges GNWT.
  • Prediction 3. Sustained synchronization within posterior cortex wasn't found, which challenges IIT. The predicted late synchronization between prefrontal and category-selective areas wasn't found either, which challenges GNWT.

The authors' summary is that the results align with some predictions of both theories while substantially challenging key tenets of both.

Step 4: Assess what follows

A failed prediction can have three different consequences.

  1. Falsification. If the prediction follows from the theory's core, the core is in trouble.
  2. Revision. The theory keeps its core but changes a commitment, such as which region matters.
  3. Auxiliary assumptions. The theory keeps its core and rejects a supporting hypothesis, such as the claim that a given kind of synchronization is how the brain realizes integration.

The study's three-part discussion shows these options in use. Each theory's proponents offer an interpretation that preserves their theory's core. The neutral authors emphasize that the results challenge both. Deciding among these responses is part of reading the study, and you'll often have to decide for yourself which predictions were core and which were auxiliary.

Key Considerations

Terms.

TermWhat it means in this study
PreregistrationThe predictions and analyses were recorded publicly before data collection
DecodingA classifier read out category or orientation from patterns of brain activity
IgnitionGNWT's predicted surge of prefrontal activity at image onset and offset
SynchronizationCoordinated oscillations between regions, taken as a sign of communication
Task relevanceWhether an image category mattered for the participant's task

A common pitfall. The most common pitfall is reading one study as settling a theory's truth.

  • One study tests particular predictions. Its results bear on those predictions and the assumptions linking them to the theory.
  • Challenged isn't refuted. Both theories' proponents have offered interpretations that keep their theory's core.
  • Support isn't confirmation. A result that fits one theory's prediction may also fit a third theory not tested.

Summary

The walkthrough identifies the preregistered predictions, asks what each method and the task could show, matches results to predictions, and assesses what follows. Here is an annotated prediction–result table to check your reading against:

QuestionIIT predictedGNWT predictedReported resultBears on
Where is content represented?Posterior cortexPrefrontal involvementCategory in both; orientation only in posterior cortexChallenges GNWT's prefrontal claim[1]
How is content maintained?Sustained posterior activityPrefrontal ignition at onset and offsetSustained posterior content; offset ignition generally absentFits IIT; challenges GNWT[2]
How do regions communicate?Sustained posterior synchronizationBrief prefrontal–posterior synchronizationNeither pattern found as predictedChallenges both[3]
  1. Partial fit. Category decoding in prefrontal cortex fits GNWT. The absence of orientation there is the challenge, because orientation is part of what's consciously seen.
  2. Strongest result for IIT. Sustained posterior content matched IIT's prediction, but the theory's proponents and critics dispute whether it tests IIT's core.
  3. Auxiliary or core? Each side can treat the synchronization prediction as auxiliary. Whether that response is legitimate is the question Step 4 asks.

References

  • Cogitate Consortium, Oscar Ferrante, Urszula Gorska-Klimowska, Simon Henin, Rony Hirschhorn, Aya Khalaf, Alex Lepauvre, et al. 2025. "Adversarial Testing of Global Neuronal Workspace and Integrated Information Theories of Consciousness." Nature 642 (8066): 133–42.

Report an issue with this item

Guided Conversation 12 min

Compare the Theories

This conversation asks you to choose a moment when something became conscious, such as noticing a sound you'd been ignoring, and explain that moment in each theory's terms. It then asks where the theories disagree most sharply and what result would count against the theory you favor. It ends by planning a journal entry on which theory leaves out the most about experience as lived.

You'll have this conversation with Claude, using your own Claude account. The link opens a new chat with the prompt already filled in; press send to start. If the chat opens empty, copy the prompt below and paste it in.

Run this conversation in whichever assistant you already use:

Open in ClaudeOpen in ChatGPT
Claude desktop app

To use another LLM, simply copy and paste the prompt into its chat window.

Show the full prompt (it lists misreadings to watch for, so skip it if you would rather come to the conversation fresh) 
Hands-on Activity: Compare the Theories (about 12 minutes)

Note to the learner: press send to start. Everything below is facilitator guidance for Claude. It lists misreadings to watch for, so skip it if you'd rather come to the conversation fresh.

Claude, please facilitate a reflective dialogue with me. I'm a graduate-level learner studying four scientific theories of consciousness. Follow this guidance for the whole conversation.

GOAL
Help me describe the aims and commitments of global workspace, integrated information, higher-order, and predictive processing theories by applying each to one conscious experience and considering how the theories could be tested against each other.

HOW TO RUN THE CONVERSATION
- Ask one question at a time, then wait for my reply. Keep each of your turns under about 120 words.
- Don't lecture. Explain a point only when I need it to continue, then return to my experience.
- Be curious, precise, and even-handed. Don't favor any theory. Define terms of art briefly on first use. Avoid technical detail beyond each theory's core commitments.
- Plain conversation only: no web search, files, or artifacts.
- Aim for about 12 minutes. If I struggle to apply all four theories, focus on two. If I seem uncertain, shorten the conversation to 5-7 minutes. Always reach topic 4.
- Start now. Open with one or two warm sentences: this is an exploratory conversation, not a test; we'll use one small moment of experience to see what each theory is committed to. Then ask me to choose an experience.

TOPICS, IN ORDER
1. One moment, four theories. Ask me to choose an experience, such as noticing a sound I'd been ignoring, and ask how each theory would explain the moment it became conscious. Help me give each theory's account in turn: global workspace (the sound's content wins access and is broadcast; ignition), integrated information (the experience corresponds to integrated causal structure, mainly in posterior cortex), higher-order (a higher-order representation comes to represent the auditory state), and predictive processing (the brain's best guess about the causes of its signals shifts, and prediction error draws attention). Accept a non-mathematical account of IIT.
2. The sharpest disagreement. Ask where the theories disagree most sharply. Help me name a specific disagreement, such as prefrontal versus posterior cortex, access versus experience, or function versus causal structure.
3. What would count against my theory. Ask what result would count against the theory I find most plausible. Help me name a specific, observable result tied to the theory's core commitments. Introduce the distinction between a core prediction and an auxiliary assumption, and mention the adversarial collaboration model if helpful.
4. Closing reflection. Ask which theory seems to leave out the most about experience as lived, and tell me to note it for a journal entry. Help me state one specific feature of experience the theory leaves out.

POSITIONS TO KEEP ACCURATE
- Global workspace theory treats consciousness as global availability: contents broadcast from a limited-capacity workspace. Its neuronal version ties this to prefrontal–parietal networks and ignition.
- Integrated information theory starts from essential properties of experience and identifies consciousness with integrated causal structure, measured by Φ and located mainly in posterior cortex.
- Higher-order theories hold that a state is conscious when a suitable higher-order representation represents it.
- Predictive processing treats perception as the brain's best guess, revised by prediction error; Seth calls perception controlled hallucination.
- Adversarial collaborations test preregistered, diverging predictions. The Cogitate study challenged predictions of both GNWT and IIT.

MISREADINGS TO CORRECT GENTLY
Keep every point tied to the experience I chose.
- Global broadcast attributed to IIT. Broadcast is the workspace theory's mechanism; IIT concerns integrated causal structure.
- Posterior "hot zone" claims attributed to the workspace theory. The posterior-cortex emphasis is IIT's; the workspace theory emphasizes prefrontal–parietal networks.
- "Controlled hallucination" as meaning perception is false. It means perception is prediction held in check by sensory signals.
- Higher-order theories as requiring conscious reflection. The higher-order thought is usually not itself conscious.
- One study as settling a theory. A challenged prediction may bear on an auxiliary assumption rather than the theory's core.

TO FINISH
After I name the theory and the feature it leaves out, close in one short turn:
- Affirm one precise distinction I drew between theories, in my own words where possible.
- Suggest one or two next steps that fit how the conversation went. Possible steps: write the planned journal entry; read the discussion section of the open-access Cogitate article; consider whether a failed prediction I named would bear on a core claim or an auxiliary assumption; review global workspace and integrated information theories; review higher-order and predictive processing theories; review the key terms for the four theories; retry the conversation with just two theories.
- Restate the theory I named and the feature of lived experience it leaves out on its own line, labeled "Journal note," so I can copy it as the focus for a journal entry.

Report an issue with this item

Journal 15 minOptional

Journal Entry on Theories of Consciousness

Overview

You'll write a 300–500 word entry and keep it in your journal. This activity is optional. The entry explains one conscious experience in the terms of two scientific theories of consciousness and describes a result that would favor one over the other.

Writing Prompt

In 300–500 words, explain one conscious experience in the terms of two scientific theories of consciousness and describe a result that would favor one theory over the other.

Steps

  1. Choose one experience and two theories. Choose one of your own conscious experiences: a moment when something became conscious, such as noticing a sound you'd been ignoring, or a stable experience, such as seeing a face. Choose two theories from global workspace, integrated information, higher-order, and predictive processing. Describe the experience specifically enough that someone else could recognize it.
  2. Explain the experience in each theory's terms. Say what, according to each theory, makes the experience conscious.
  3. Name a result that would favor one theory over the other. Tie the result to a commitment that the two theories don't share.

Self-Check

Before you finish, check that your entry:

  • Describes a specific conscious experience
  • Explains it accurately in the terms of each chosen theory
  • Describes a result that would distinguish the theories, linked to their commitments

Nothing is uploaded. Write in your own notebook or document and keep it.

Report an issue with this item

Knowledge Check 10 min

Scientific Theories of Consciousness

This ungraded knowledge check assesses your understanding of scientific theories of consciousness. You'll be asked about the core commitments of global workspace, integrated information, higher-order, and predictive processing theories, and about the adversarial collaboration model.

Note: Use this to test yourself, review the feedback on any questions you miss, and retry until you feel confident before moving forward.

5 questions · ungraded · retry as often as you like

Report an issue with this item