In 2025, the Cogitate Consortium published the results of an unusual experiment in Nature. Rather than letting rival theories of consciousness talk past each other, the researchers had proponents of Integrated Information Theory and Global Neuronal Workspace Theory pre-register competing predictions. They then tested those predictions with fMRI, MEG and intracranial recordings. Neither theory emerged unscathed. Posterior cortex failed to show the sustained synchronisation IIT anticipated, while prefrontal cortex lacked the ignition at stimulus offset that workspace theory required.

The result is instructive less for which theory lost points than for what both theories share. Each treats consciousness as something to be explained in terms of information: how it is integrated, broadcast, or made globally available. The disagreement concerns structure and location, not the underlying vocabulary. Information has quietly become the lingua franca of consciousness science. It is the common currency in which neuroscientists, AI researchers and philosophers express their otherwise incompatible commitments.

That consensus deserves scrutiny. Information is a slippery concept. It is sometimes a precise mathematical quantity, sometimes a metaphor, and sometimes a candidate for the fabric of reality itself. Whether consciousness is information, depends on information, or merely correlates with information processing are three very different claims. This piece examines the strongest versions of each, from integrated information to Wheeler's 'it from bit', and asks what it would take to measure the relationship at all.

Giulio Tononi's Integrated Information Theory makes the boldest claim in the field. It does not say consciousness correlates with integrated information, or emerges from it. It says consciousness is integrated information, specifically the maximally irreducible cause-effect structure a physical system specifies over itself. The quantity Φ (phi) measures how far a system's causal powers exceed those of its parts taken separately. Where Φ is maximal, there is a subject. Where it is zero, there is nobody home.

IIT's method is distinctive. Rather than starting from neural data, it begins with axioms about experience: that experience exists intrinsically and is structured, specific, unified and definite. From these axioms it derives physical postulates that any conscious substrate must satisfy. IIT 4.0, published by Albantakis, Tononi and colleagues in 2023, refines this into a formal calculus of causal distinctions and relations. The ambition is to read phenomenal structure directly off physical architecture, so that the shape of an experience corresponds to the shape of its cause-effect structure.

The consequences are striking, and for many critics, disqualifying. Because Φ depends on causal architecture rather than function, a purely feedforward network has zero Φ however sophisticated its behaviour. On IIT's account, a conventional digital computer that perfectly simulated a human brain could be negligibly conscious. Scott Aaronson argued that simple, inert lattice-like systems could in principle have Φ exceeding a human brain's, and Tononi accepted the implication. In 2023, over a hundred researchers signed a letter labelling IIT pseudoscience, a charge that provoked sharp counter-responses.

Alternative information-based theories locate consciousness differently. Global Workspace Theory, developed by Bernard Baars and given neural form by Stanislas Dehaene, treats consciousness as information becoming globally available. That information is broadcast across a fronto-parietal network to consumer systems for report, memory and planning. Dennett's notion of 'fame in the brain' captures a similar intuition. Predictive processing frameworks, building on Karl Friston's free energy principle, describe experience in terms of the brain's generative models and the precision-weighted prediction errors that update them.

The crucial divide is between intrinsic and extrinsic information. Workspace and predictive theories use information in roughly Shannon's sense: what a state tells some consumer about something else. IIT insists on information for the system itself, independent of any observer. This is not a technical quibble. It determines whether consciousness depends on what a system does with information, which would make it realisable in any suitably organised machine, or on what a system physically is.

Takeaway

When a theory says consciousness is information, ask whose information it means. Information for an external observer and information for the system itself lead to radically different verdicts about brains, computers and machines.

In 1989, John Archibald Wheeler proposed that every physical 'it' derives its existence from 'bits', the answers to yes-or-no questions posed by measurement. 'It from bit' was partly a provocation, but it crystallised a trend. Rolf Landauer insisted that 'information is physical'. The Bekenstein bound limits how much information a region of space can contain, and the holographic principle in quantum gravity points the same way. Together they suggest information is not a bookkeeping device layered over physics but part of its foundations.

If information is fundamental, the explanatory gap might narrow. David Chalmers explored this possibility in his double-aspect theory of information. Perhaps information has two aspects: a physical one, captured by its structural and causal roles, and a phenomenal one, the way it is experienced. On this view, consciousness would not need to be squeezed out of matter. It would be the intrinsic face of the very information that physics describes extrinsically. The view connects naturally with Russellian monism and with IIT's own claims about intrinsicality.

The difficulty is that information, in the sense physics and engineering actually use, appears thoroughly relational. Shannon information is defined relative to a probability distribution over possible states, a choice of alphabet, and a channel connecting source to receiver. John Searle pressed the point bluntly: apart from observer-relative interpretation, there is no fact about what information a physical system carries. If information requires an interpreter, it cannot ground the interpreter's own consciousness without circularity.

Defenders reply that physics does supply observer-independent quantities, such as entropy, mutual information between subsystems, and the counterfactual structure of causal dynamics. IIT goes further and redefines information as intrinsic causal power, sidestepping Shannon's framework entirely. Critics counter that this simply rebrands causation as information without explaining why causation should feel like anything. The same worry recurs throughout informational ontologies. Structural descriptions, however rich, describe relations, and it remains unclear how relations alone yield the qualitative character they are meant to explain.

Dennett offered a deflationary alternative in his account of 'real patterns'. On this account, information patterns are objectively there, discoverable and predictively indispensable, without being ontologically basic. Asking whether information is fundamental may then be the wrong question for consciousness science. What matters is which patterns of information processing reliably accompany report, discrimination and flexible control. The deflationist trades metaphysical depth for empirical tractability. That bargain looks better or worse depending on whether one believes the hard problem is real.

Takeaway

Information can be real without being fundamental. Much of the debate turns on whether consciousness must be grounded in the bedrock of physics, or merely in patterns that are objectively there.

Even granting a deep link between consciousness and information, the relationship must be measured before it can be tested. Here information theories face severe practical limits. Computing Φ exactly requires evaluating every possible partition of a system and every candidate mechanism, a search that grows super-exponentially with system size. Exact Φ has been calculated only for toy networks of a handful of nodes. For a brain of eighty-six billion neurons, the quantity IIT identifies with consciousness is, for now, practically incomputable.

Researchers have therefore turned to proxies. The most successful is the perturbational complexity index (PCI), developed by Marcello Massimini, Adenauer Casali and colleagues. A pulse of transcranial magnetic stimulation perturbs the cortex, EEG records the response, and an algorithm compresses the resulting spatiotemporal pattern. Conscious brains produce responses that are both widespread and differentiated. Unconscious brains respond either locally or with stereotyped slow waves. PCI separates wakefulness from deep sleep, general anaesthesia and vegetative states with impressive reliability.

Yet PCI's success is ambiguous evidence. The measure was inspired by integration and differentiation, but it does not measure Φ. Its results fit equally well with workspace accounts that emphasise recurrent cortical dynamics. More fundamentally, every such measure is calibrated against cases where we already know, typically through report, whether someone is conscious. The calibration problem is unavoidable. We validate the meter against behaviour, then hope it generalises to cases where behaviour is absent, such as covert awareness, infants, other species or machines.

The 'unfolding argument', advanced by Adrien Doerig and colleagues in 2019, sharpens this into a dilemma. Any recurrent network can, in principle, be unfolded into a feedforward network with identical input-output behaviour. A theory that assigns consciousness by causal structure must predict that these functionally identical systems differ in experience. Yet no behavioural experiment could ever detect the difference. Either the theory is unfalsifiable on that point, or it must collapse into a functional account in which structure no longer matters independently.

This is why adversarial collaborations like Cogitate matter. They force theories to commit to predictions within the reportable domain, where verification is possible. But they cannot touch the claims that make information theories philosophically ambitious: the verdicts about systems that cannot report. As large language models produce increasingly fluent claims about inner states, the gap widens. Behaviour has become cheap evidence, while structure-based measures remain either uncomputable or uncalibrated outside the human case.

Takeaway

Every consciousness meter is calibrated on beings who can tell us what they experience. The hardest questions arise precisely where that calibration runs out.

The relationship between consciousness and information remains genuinely open, but its shape is clearer. Three questions are often run together: whether consciousness is identical to information, whether information is fundamental to reality, and whether informational measures track consciousness. Each can be answered independently, and progress depends on keeping them apart.

For science, the implication is methodological. Information-based measures like PCI are already clinically valuable, however the metaphysics resolves. Adversarial testing is pushing theories toward sharper, riskier predictions. The next frontier is tractable measures that discriminate between integration, broadcast and prediction, rather than confirming all three at once.

For philosophy, the lesson cuts deeper. If information is intrinsically relational, informational theories must either supply an intrinsic notion that does real explanatory work or accept Dennett's deflation. That choice will shape how we assess artificial systems, and it may matter sooner than either neuroscience or philosophy is prepared for.