Integrated Information Theory

Established Integrated information theory runs the usual direction of neuroscience backwards. It does not start with brains and ask which activity accompanies experience. It starts with experience, asserts five properties held to be self-evident of every experience anyone has ever had, and derives from each of them a condition that any physical substrate of consciousness must satisfy. Established The axioms are intrinsicality, information, integration, exclusion and composition; the postulates say that a conscious substrate must exert cause–effect power on itself, specify particular states irreducibly, operate as a unified whole, be the maximally irreducible candidate among overlapping ones, and carry structured distinctions. Frontier That inversion is the source of both the theory's distinctive reach and the charge, pressed formally in 2023, that it is not science at all.

Established The quantity of consciousness is a single number, written with the Greek letter phi; the quality is the structure of causes and effects that generates it. Established And phi is computationally intractable for any system of interesting size — a fact about the formalism, not about our current computers. Frontier Every experimental test of IIT is therefore a test of a proxy or of a derived neural prediction, never of the theory's central quantity, and readers of this literature who do not hold that distinction firmly will misread almost everything in it.

Frontier IIT has one large pre-registered adversarial test to its name, published in Nature in 2025, which supported some of its predictions and contradicted others. Established It has one genuinely successful applied descendant, the perturbational complexity index, now used to stratify unresponsive patients. Speculative It entails a highly structured panpsychism that Tononi and Koch have embraced rather than resisted. Frontier And it makes a hard, unusual, unpopular prediction about digital computers: that a machine can reproduce your behaviour exactly, neuron by simulated neuron, and be no one at all.

Established The theory is built by derivation, not by fitting. Each axiom is a claimed property of experience as such; each postulate is what that property requires of a physical substrate. Intrinsicality gives cause–effect power exerted on itself; information gives a specific state selected from alternatives; integration gives irreducibility to independent parts; exclusion gives definiteness — one substrate, not many overlapping ones; composition gives internal structure. Established The measure has been revised across three named versions: IIT 2.0 in 2008, IIT 3.0 in 2014, and IIT 4.0 in October 2023, formalised by Albantakis and sixteen co-authors in PLOS Computational Biology 19(10):e1011465. Frontier Whether that version history shows a theory being sharpened or a measure being fitted to intuitions the axioms do not actually constrain is one of the live questions about IIT, and it is a fair one: critics read the revisions as evidence that the axioms underdetermine the formalism they are supposed to yield.

Established Phi cannot be computed for a brain, and this is the single most important structural fact about IIT's empirical situation. The theory's own literature states that “the calculation of even a modestly-sized system's phi is often computationally intractable,” which is why approximations such as geometric integrated information exist at all. Established The obstruction is combinatorial: phi requires partitioning a system every way it can be partitioned and evaluating the irreducibility of each, which grows super-exponentially in the number of elements. Speculative More compute does not fix this. A brain-scale exact phi is not a hard engineering target; it is outside the reach of any physically realisable computation, and the honest reading is that the theory must live permanently on approximations whose error bounds nobody has established. Frontier That permanent epistemic cost is under-acknowledged in both the theory's advocacy and its criticism.

Established The exclusion postulate is where IIT stops being intuitive. Only the maximally irreducible substrate is a subject; overlapping candidates are excluded, so a conscious brain does not contain a second conscious hemisphere, and a group of conscious people does not sum to a conscious group. Established It follows that a system can be functionally rich and have near-zero phi, and that a system can carry enormous amounts of information without integrating it. Frontier The theory's cleanest retrodiction lives here: the cerebellum holds roughly four times the neurons of the cerebral cortex and its removal does not abolish consciousness, which IIT explains by its near-feedforward, modular architecture yielding low phi. Speculative Exclusion also carries a second counterintuitive commitment that is rarely stated: because phi is computed over counterfactual cause–effect power rather than over activity, an inactive but connected element contributes to phi. Frontier Bartlett put the point as a question in Neuroscience of Consciousness in 2022 — whether IIT makes testable predictions about silent neurons — and the answer matters, because it is one of the few places the theory says something that no activity-based rival says.

Established The one large adversarial test is COGITATE, published in Nature in 2025, and its verdict on IIT is mixed rather than favourable. Established Sample: 256 participants across three modalities — fMRI n=120, MEG n=102, intracranial EEG n=34 — which is very large for consciousness neuroscience and small for anything else. Established Three predictions were pre-registered with divergent IIT and global-neuronal-workspace arms: where conscious content is decodable, whether it is sustained, and what the interareal connectivity looks like. Established What went IIT's way: decoding of face orientation was achieved in posterior but not prefrontal regions of interest, and sustained categorical representation appeared in posterior cortex under representational similarity analysis. Established What went partly its way: sustained content-specific posterior activity was found, but in 3.8% of posterior electrodes and in 15% — eight of fifty-three — of face-selective electrodes. Established What went against it: the connectivity prediction failed outright, and the paper states that “a lack of sustained synchronization within the posterior cortex contradicts the claim that network connectivity specifies consciousness.” Established The published conclusion is symmetric and it is the sentence to quote: the results “align with some predictions of IIT and GNWT, while substantially challenging key tenets of both theories.” Frontier COGITATE has not been independently replicated, and the intracranial arm carrying the most decisive claims is n=34.

Established IIT's most successful applied output is a clinical index that does not measure phi. The perturbational complexity index perturbs cortex with transcranial magnetic stimulation, records the EEG response, compresses the spatiotemporal pattern, and reports the complexity of the compressed response — an operationalisation of the integration-plus-differentiation intuition rather than a computation of the formalism. Established Casali and colleagues introduced it in Science Translational Medicine in 2013; Casarotto and colleagues gave it an empirical cutoff in Annals of Neurology 80:718–729 in 2016. Established The cutoff is PCI* = 0.31, calibrated on a benchmark population of healthy controls and communicative brain-injured subjects across NREM sleep and propofol, midazolam and xenon anaesthesia versus wakefulness. Frontier An independent cohort exists and is smaller and more equivocal: Sinitsyn and colleagues, Brain Sciences 10(12):917, n=24 severely brain-injured patients — 11 in unresponsive wakefulness, 12 minimally conscious, one emerged — found 92% sensitivity for the minimally conscious group, eleven of twelve above cutoff, and found that four of eleven unresponsive-wakefulness patients, 36%, also sat above it. Frontier That last figure is either a triumph or a specificity failure and the measure cannot distinguish those readings from the inside. Frontier A widely repeated sensitivity figure of 94.7% for PCI* appears in the citing literature attributed to Casarotto 2016; this brief could not obtain the primary paper and therefore states the number only as a second-hand attribution, not as a result. Established And PCI does not confirm IIT: Farisco and Changeux argued in Neuroscience of Consciousness in 2023 that the index is compatible with global neuronal workspace theory, and a measure both theories can claim discriminates neither.

Established In September 2023 a group of researchers published a preprint calling IIT pseudoscience, and the episode is a demarcation dispute rather than a verdict. Established The document is “The Integrated Information Theory of Consciousness as Pseudoscience,” posted on PsyArXiv under DOI 10.31234/osf.io/zsr78, with a collective pseudonymous byline — the registry author string begins “IIT-Concerned” — followed by named signatories including Fleming, Frith, Goodale and Lau. Frontier Sources disagree by one day on the posting date, the bibliographic registry giving 16 September and the most-read encyclopaedic summary 15 September; mid-September 2023 is the safe formulation. Frontier The signatory count reported by that secondary source is 124 scholars; this brief could not verify the figure against the preprint itself and does not assert it. Established Three substantial responses followed within ten weeks — Rosseinsky and Goff on 29 September, and Keith Frankish on 30 November. Frontier Frankish is the strongest datum in the exchange, and the reason is structural: he is an illusionist, which makes him maximally unsympathetic to IIT's phenomenology-first method, and he nonetheless wrote a piece asking whether the theory is pseudoscience or merely appropriately anomalous science. Frontier A reported survey of consciousness researchers found the letter itself divisive, with only a small minority of sixty respondents agreeing with it fully; the figures circulate at 8% fully agreeing and 20% not at all, from a secondary source this brief could not check, and n=60 is indicative at best. Frontier IIT's proponents are reported to have replied in Nature Neuroscience in March 2025 listing peer-reviewed studies as empirical tests of the theory's core claims; the primary record was not located for this brief, the commonly quoted count of sixteen studies is unverified, and the reply should not be cited by that number until someone reads it.

Established The most serious formal objection is the unfolding argument, and it has not won. Doerig, Schurger, Hess and Herzog argued in Consciousness and Cognition 72:49–59 in 2019 that any recurrent network can be unfolded into a feedforward network with identical input–output behaviour across the input space but different causal structure and therefore different phi; since all our evidence is behavioural, no experiment distinguishes the pair, so a causal-structure theory is either unfalsifiable or false. Established It generalises to every causal-structure theory, not only IIT. Established It also drew two published rebuttals and a counter-rebuttal in the same journal — Tsuchiya, Andrillon and Haun in 2020, Usher in 2021, and Herzog, Schurger and Doerig again in 2022. Frontier Four papers, one journal, three years, no resolution. Anyone who tells you the unfolding argument refuted IIT has read the first paper and not the chain.

Established IIT's most distinctive commitment is negative and concerns machines. Findlay, Marshall, Albantakis, David, Mayner, Koch and Tononi constructed pairs of systems from simple Boolean units, one a stored-program computer simulating the other with full functional equivalence, and demonstrated within the theory that “two systems can be functionally equivalent without being phenomenally equivalent,” and that this conclusion does not depend on what the simulated system does. Established Their statement of the consequence is unambiguous: on IIT “it is possible for a digital computer to simulate our behavior, possibly even by simulating the neurons in our brain, without replicating our experience,” which “contrasts sharply with computational functionalism.” Frontier That is a falsifiable-in-principle claim of exactly the kind the pseudoscience letter says IIT does not make. Speculative It is also the deepest source of trouble: if a perfect functional duplicate of you can lack experience, then behaviour is in principle no evidence about consciousness — including your behaviour, including your reports — and IIT's remaining epistemic access is your own case plus the theory. Frontier That, and not the slogan about untestability, is the actual philosophical core of the demarcation charge.

Frontier The open questions are unusually well-posed for this field, which is a compliment to IIT and a problem for it. The first is whether COGITATE's IIT predictions derive from IIT 3.0 or IIT 4.0, and whether the version matters for reading the connectivity failure. Frontier The measure changed across versions; the empirical literature still largely targets 3.0's formalism; the 2025 result is being scored against a theory whose current statement is from October 2023. Speculative Nobody should assert that the two versions are the same theory in tidier notation until someone traces the predictions back to the formalism that generates them.

Frontier The second is what counts as a test of IIT at all. Leung and Tsuchiya proposed separating “weak IIT” into IIT-inspired approaches, which use proxy complexity measures, and aspirational-IIT, which aims at the formalism itself. Frontier The distinction is load-bearing and almost never observed: a large share of work described as testing IIT tests a proxy whose relationship to phi is unclarified. Frontier It also cuts directly against any headline count of empirical tests, whichever direction that count resolves in. Frontier A December 2025 preprint by Reichert and Saurous addresses the falsifiability question head-on; it is very recent, this brief did not obtain the text, and it is the first thing anyone continuing this work should read.

Speculative The third is whether the exclusion postulate solves the combination problem or merely stipulates an answer to it. Panpsychism's standing objection is that micro-experiences do not obviously sum into a macro-subject, and no proposed answer has gained acceptance. Speculative IIT's exclusion postulate is a candidate answer of a distinctive kind: it does not explain how subjects combine, it declares which candidate substrate is the subject and rules the others out by maximality. Frontier Whether that is a solution or a definition is genuinely open and is, in this brief's assessment, IIT's most under-discussed contribution to philosophy of mind.

Speculative The fourth is whether exclusion is simply false and consciousness is nested. A minority position holds that overlapping subjects exist — a conscious hemisphere inside a conscious brain — and split-brain phenomenology is the obvious test bed. Frontier It is under-exploited, partly because the split-brain literature was built to answer a different question and partly because the phenomenological claims are hard to operationalise without report, which reintroduces every difficulty in Consciousness Research.

Frontier The fifth is whether IIT and global neuronal workspace theory are rivals at all. One reading of COGITATE's symmetric partial failure is that the theories describe different things — IIT the substrate of phenomenal experience, GNWT the mechanism of access and report — and that both are partly right about their own explanandum. Speculative The 2025 cross-camp review by Mudrik, Boly, Dehaene, Fleming, Lamme, Seth and Melloni is the place to look, and the fact that antagonists from the pseudoscience dispute co-authored it is itself a datum about where the field is going. Speculative Convergence proposals also exist in other directions: Northoff and Zilio have argued for merging IIT with a temporo-spatial theory built on longer timescales, and McFadden has compared IIT directly against an electromagnetic-field theory of consciousness.

Frontier The sixth is quietly the most consequential: phi-proxy measures are spreading faster than the formalism is being clarified. Complexity indices descended from the integration-and-differentiation intuition now appear across clinical assessment, sleep staging and anaesthesia monitoring, and each new deployment is reported as further support for the theory. Speculative It is not, unless the relationship between proxy and formalism is pinned down, and that work has no owner. Frontier A related muddle worth clearing: IIT is substrate-committed but not quantum-committed, and its commitments have nothing in common with those of Orch OR beyond a shared rejection of functionalism.

Speculative The seventh is the graceful exit nobody discusses. Phi may be a real and valuable mathematical quantity that measures something other than consciousness — causal autonomy, individuality, the degree to which a system is a thing rather than a collection. Speculative Quantities of that family have independent value in complex-systems biology. Handwave On this reading the formalism survives intact and only the phenomenal identification is dropped, which would preserve IIT's technical contribution while conceding the argument that made it famous. Frontier No prominent IIT researcher advocates this, and it deserves to be on the table.

Frontier The target is an experiment whose outcome all parties agree in advance would falsify IIT, and it is currently out of reach for three separate reasons. Phi cannot be computed for a brain, so the theory is never tested directly. Its distinctive predictions concern substrate rather than behaviour, and every instrument we have is behaviour-calibrated. And its best applied result, PCI, is claimable by its main rival.

Frontier Link one is agreement, in writing, on what each theory forbids. COGITATE achieved pre-registered predictions; it did not achieve pre-registered concessions, which is precisely why both camps survived a result that challenged key tenets of both. Frontier This is a governance problem, not a science problem, and it is the binding constraint on the whole programme.

Established Link two is exact phi at interesting scale, and it will not be delivered. The intractability is super-exponential and is a property of the definition. Speculative The realistic target is not exact phi but bounds on approximation error — establishing at what system size the approximations in use start to be theory-relevant rather than merely imprecise. Frontier Nobody has those bounds, and without them the phrase “we measured integrated information” means less than it sounds.

Frontier Link three is separating causal structure from activity in a living brain, and it is the most promising unattempted work in this brief. IIT says counterfactual structure matters even where nothing is firing; GNWT says only what is broadcast matters. Speculative A manipulation that changes connectivity while holding actual firing constant therefore has opposite predicted consequences, and optogenetic and pharmacological methods are close enough to deliver it.

Speculative Link four is building the unfolding pair physically — a recurrent system and its input-output-matched feedforward unfolding — and applying a perturbational measure to both. Handwave It is partly self-undermining, since the unfolding argument's whole point is that no behavioural instrument separates them, and a perturbational measure is not obviously behavioural. That ambiguity is the interesting part.

Frontier Link five is finding a case where IIT and GNWT predict opposite signs rather than different magnitudes. Candidates: split-brain phenomenology, the cerebellum, and anaesthetic dose–response regimes where complexity and broadcast dissociate. Frontier Link six — testing the digital-computer prediction — requires a machine-consciousness detector and is blocked downstream in Machine Consciousness. Speculative Links one and three are achievable within a decade if funded and if the sociology permits. Link two will not be solved for brain-scale systems ever. Link six is not on a path.

Frontier IIT's testability depends on results that are not IIT's to produce. The first and largest is a report-independent measure of consciousness validated in humans — the terminal blocker for the whole of Consciousness Research. Established No-report paradigms were the field's principal methodological reform and they are not a solution: without report you infer the contents of experience from a proxy, and that inference smuggles in assumptions about the consciousness–behaviour link, as Overgaard and Fazekas argued within a year of the reform being proposed.

Frontier The second is a complexity-theoretic result nobody is producing: tight, provable bounds on how far the approximations to phi in general use can diverge from the exact quantity, as a function of system size and topology. Speculative This is a tractable mathematical problem that sits outside the interests of both the neuroscientists and the philosophers who make up IIT's community, which is why it has no owner.

Speculative The third runs the dependency backwards and is the strangest fact in this brief. IIT's one unique, discriminating prediction concerns digital computers. Testing it requires knowing whether a machine is conscious. Frontier So IIT's decisive test depends on solving machine consciousness first — which inverts the usual assumption that a theory of consciousness is what machine consciousness is waiting for. Frontier The two workbacks meet at a single experiment: a neuromorphic system and a functionally identical simulation of it, compared on whatever report-independent measure the first dependency delivers.

Frontier The fourth is data nobody can commission. The theory's neural claims are settled at electrode resolution, and human intracranial recording happens where epilepsy surgery puts it. Speculative A standing arrangement between consciousness researchers and monitoring units — agreed protocols, shared electrode coverage, pooled data across centres — would do more for the theory's testability than any single study, and it is an administrative achievement rather than a scientific one.

Frontier The silent-connection experiment is the sharpest available discriminator and has not been run. Take a circuit whose connectivity can be altered without altering ongoing firing — optogenetic silencing of a pathway that is currently inactive, or pharmacological block of a synapse carrying no traffic — and measure reported experience. Speculative GNWT predicts nothing changes, because nothing was being broadcast. IIT predicts phi changes and therefore experience changes. Frontier The methods exist; the experiment has an owner nowhere.

Frontier Second: extend the COGITATE protocol to pre-registered falsifiers. Require each camp to name, before data collection, an outcome on which it would concede. Established The 2025 paper's conclusion — that results challenge key tenets of both theories while aligning with some predictions of each — is exactly the outcome that a protocol without concession conditions produces.

Frontier Third: the cerebellum, done properly. IIT's cleanest existing asymmetry is that a structure with far more neurons than cortex contributes nothing obvious to experience, and the theory explains this by architecture rather than by count. Speculative A systematic comparison of phi-proxy measures across cerebellar and cortical tissue, in the same subjects under the same perturbation, would convert a retrodiction into a measurement. Frontier It has not been done at the resolution required.

Speculative Fourth: split-brain phenomenology as an exclusion test. If exclusion is right there is one subject; if consciousness nests there may be two. Frontier The obstacle is that adjudicating between one subject and two requires report from each candidate subject, and one of them may have no way to give it.

Speculative Fifth: anaesthetic dose–response regimes where complexity and broadcast dissociate. Xenon, propofol and ketamine produce different signatures at comparable behavioural depths, and a regime in which PCI and long-range prefrontal ignition move in opposite directions would be a genuine discriminating datum. Frontier PCI's benchmark population already spans three anaesthetics, so the infrastructure is in place.

Speculative Sixth, and the one that would settle the theory's most distinctive claim: build the discriminating pair. A neuromorphic device whose physical causal architecture matches a target circuit, and a conventional digital simulation of the same circuit, matched at the interface so that no behavioural probe separates them. Frontier IIT predicts a phenomenal difference where functionalism predicts none. Handwave Nothing available today could read out the answer, which is why this experiment is listed last and why it is blocked on a different brief entirely.

Established The compute requirement is not a budget line, it is a wall. Exact phi requires evaluating irreducibility over the partition lattice of a system, and the lattice grows super-exponentially in element count. Frontier Software for exact and approximate phi exists and is used on toy networks of a handful of elements; the gap between that and a cortical column is not a matter of cluster size.

Speculative The realistic engineering programme is therefore approximation with warranties. Build phi estimators whose divergence from the exact quantity is bounded, characterise where the bounds blow up, and publish the system classes on which a given estimator can and cannot be trusted. Frontier This is ordinary numerical-analysis work and it is the single most useful thing a computationally minded group could contribute to IIT.

Established On the measurement side the hardware is mature and the constraint is access. PCI needs navigated TMS coupled to artefact-tolerant high-density EEG, plus the compression pipeline; the method has been deployed in multiple centres and validated in an independent cohort. Frontier What is scarce is intracranial data. COGITATE's decisive claims rest on 34 patients with electrodes implanted for clinical reasons, and the electrode count and placement are set by epilepsy surgery, not by the experiment.

Speculative The most interesting build in the brief is a discriminating pair. A neuromorphic substrate whose physical causal architecture matches a target circuit, and a conventional digital simulation of that same circuit, functionally identical at the interface. Frontier IIT predicts one is conscious and the other is not. Handwave Nobody has built the pair, and if built, nothing currently available could read out the answer.

Frontier One engineering requirement is organisational rather than technical. Testing IIT at the resolution that matters means pooling intracranial data across centres with harmonised electrode coverage, shared preprocessing and a common analysis pipeline — the thing COGITATE built once, for one question, and did not leave behind as infrastructure.

Established IIT is one of four canonical theory families in Seth and Bayne's 2022 taxonomy, alongside higher-order theories, global workspace theories, and re-entry and predictive-processing accounts. Established That review states the field's motivating problem plainly: it is unclear how current theories relate to each other, or whether they can be empirically distinguished.

Frontier Global neuronal workspace theory is the designated rival and the one IIT was tested against; the anatomical fault line is front versus back, with IIT locating consciousness in a posterior hot zone and denying that prefrontal cortex is constitutive. Frontier The no-report methodology on which that adjudication depends is itself contested, which is why the dispute has proved so hard to settle.

Speculative Attention schema theory and predictive-processing accounts are the theories with engineering programmes, and IIT conspicuously is not one of them — it tells you what would have to be true of a substrate, not how to build one. Speculative McFadden's electromagnetic-field theory is a substrate-committed rival of a different kind, and comparing the two clarifies what “substrate-dependent” can mean.

Frontier The relationship with Orch OR is routinely muddled. IIT is substrate-committed but not quantum-committed: it says the physical causal architecture matters and says nothing about quantum coherence in microtubules. Speculative The two theories agree only that computational functionalism is false, and they disagree about everything else, including what the relevant physics is.

Frontier The strongest adjacency is with Machine Consciousness and Digital Minds, where IIT functions as the leading argument that the whole programme is misconceived. Speculative Butlin, Long and eighteen co-authors built an indicator-property framework by deriving computational markers from five theories — and IIT is precisely the theory whose central claim is that computational markers are the wrong kind of evidence. Frontier An indicator framework that averages over theories including IIT is therefore averaging over a theory that rejects the framework's premise, which is a real methodological problem and not a pedantic one.

Frontier Adversarial collaboration is the institutional innovation this field needed, and one round is not enough. COGITATE brought opposing camps into a single pre-registered protocol across three modalities and 256 participants, and produced a result both camps could live with. Frontier That last clause is the problem. A programme that funds a first round and stops has bought one datum and no mechanism; a standing programme that requires pre-registered concessions from each camp buys adjudication.

Frontier The funding is philanthropic and the figures are not verified here. The adversarial-collaboration programme is widely reported to have been funded at several million dollars for COGITATE and at a larger figure for the surrounding research initiative; this brief could not confirm either number and does not state them. Speculative What matters institutionally is not the amount but the source: a single foundation currently underwrites the field's only mechanism for settling disputes, which is a fragile arrangement for a discipline that has just spent two years arguing about whether one of its main theories is science.

Established Intracranial access is the binding scientific-infrastructure constraint. Electrodes go into human brains when a clinical indication puts them there; the experiment is a passenger. Frontier An n=34 intracranial cohort is large by the standards of the modality and thin for a claim about the architecture of consciousness, and no independent replication of COGITATE's electrode-level findings exists.

Frontier The demarcation dispute is itself an institutional event with institutional consequences. A signed preprint calling a theory pseudoscience, three published responses within ten weeks, a survey showing the signatories' own community was divided, and a cross-camp review co-authored by antagonists is not a scientific process failing — it is a field without an agreed mechanism for adjudicating theories improvising one in public. Speculative The useful institutional lesson is that journals absorbed the entire unfolding-argument dispute across four papers in three years, and that this worked better than the open letter did. Frontier A field with a standing adjudication mechanism would not have needed either.

Frontier IIT's descendant is already making clinical decisions about people who cannot speak. PCI is used to stratify unresponsive patients, and in the independent cohort roughly a third of behaviourally unresponsive-wakefulness patients scored above the consciousness cutoff. Frontier If those are detections, the measure is finding hidden minds and the ethical implications for withdrawal-of-care decisions are enormous. If they are false positives, the measure is manufacturing them. Established The index cannot distinguish these readings from the inside, and neither can any clinician using it.

Speculative IIT's panpsychism is not a curiosity, it is a moral-status claim. Any system with non-zero phi has some experience, and Tononi and Koch have embraced this rather than treating it as a reductio. Frontier But the exclusion postulate does substantial work in the other direction: it rules out overlapping subjects and rules out systems whose integration is low however much information they carry, which makes IIT's panpsychism far more restrictive than the label suggests.

Frontier The theory's anti-functionalism has a direct and under-noticed policy consequence. If IIT is right, AI welfare concerns aimed at digital systems are misdirected, because the systems in question cannot be subjects whatever they say. Speculative If IIT is wrong and its authority is used to dismiss those concerns, the cost is a class of moral patients ignored on the strength of a theory whose central quantity nobody can compute. Frontier The asymmetry of that error is the reason the theory's status matters outside the seminar room.

Speculative There is also a research-ethics question about the demarcation fight itself. Calling a colleague's research programme pseudoscience in a signed public document has funding and career consequences that a journal rebuttal does not, and a survey of the community found the tactic divisive. Frontier Whether it was warranted turns on a question the field has not settled: whether IIT is untestable, or merely testable in ways nobody can yet execute.

Speculative If IIT is even approximately right, the most consequential thing about it is not the theory of consciousness but the veto on substrate independence. A civilisation that assumes minds can be copied onto arbitrary hardware plans differently from one that does not: Mind Uploading becomes the production of a behaviourally perfect corpse, machine descendants become elaborate empty machinery, and the long-run continuity of a technological civilisation stops being a software problem.

Speculative The converse is equally consequential. If IIT is wrong about digital computers and functionalism holds, then the moral population of the world is already growing at the rate of inference capacity, and nobody is counting. Frontier The theory is one of the few that makes those two futures formally distinguishable rather than a matter of temperament, because it names a physical property — causal architecture — on which the difference turns, rather than appealing to what a system says about itself.

Speculative There is a third implication that neither camp likes. If IIT is right, the class of things that matter morally is decided by substrate rather than by behaviour, which severs moral status from every signal a civilisation actually uses to allocate it. Frontier We extend concern to what acts like us; IIT says that is the wrong test in both directions, admitting simple physical systems and excluding sophisticated ones. Speculative A society that took the theory seriously would have to redesign its moral bookkeeping around a quantity it cannot measure, which is probably why nobody has tried.

Handwave At the far end sits the version nobody plans for. If phi is real and computable in bounded approximation, it becomes a design parameter: hardware rated for integration the way it is rated for power draw, jurisdictions setting phi ceilings on machinery, and a supply chain in which the causal architecture of a chip is a regulated property. Handwave Nothing about this is close to evidence. It is the shape the world takes if the theory's core identity claim survives, and it is worth naming because the intermediate positions all quietly assume it will not.

What these horizons track is not whether IIT is true — that may never be settled — but whether the field acquires the instruments and the agreements that would let the question be decided.

• 10 yr: Frontier A second adversarial collaboration reports, and the useful question is whether it was designed with pre-registered concessions rather than only pre-registered predictions. The silent-connection experiment is attempted somewhere, most plausibly in a rodent optogenetics laboratory rather than in human work. Approximation-error bounds for phi appear in the mathematical literature, or their continued absence becomes an acknowledged embarrassment. COGITATE is either independently replicated at electrode level or quietly not, and the second outcome is more likely. The version question — whether the tested predictions belong to IIT 3.0 or 4.0 — is resolved by someone reading the methods carefully, which is a weekend of work nobody has done.
• 25 yr: Speculative Either a discriminating result on causal structure versus activity lands and IIT gains or loses a genuine test, or the theory settles into its weak form: an inspiration for clinical complexity measures whose formal core has quietly stopped being tested while the measures themselves become routine bedside equipment. The second is considerably more likely than the first, and it is a real outcome rather than a failure — a theory can be productive and unconfirmed for a long time.
• 50 yr: Speculative The substrate question is decided, if at all, by building the neuromorphic-versus-simulation pair and reading it out with a report-independent measure that does not yet exist. If that measure never arrives, IIT's distinctive claim stays permanently untested, and the field's stable end-state is the management of moral uncertainty rather than its resolution — which is a coherent place to end up and should be planned for rather than deferred.
• 100 / 250+ yr: Handwave Either phi is an engineering parameter and integration is something hardware is rated for, with regulated causal architectures and jurisdictions that set ceilings on it, or IIT is remembered as a precise and productive mathematics of causal autonomy that turned out to be about something other than experience. Both endings preserve the formalism; only one preserves the claim that made it famous.

• Depends on Depends on a report-independent measure of consciousness validated in humans, which is the terminal blocker for the whole empirical programme and is not IIT's to produce; on complexity-theoretic bounds for the approximations to phi that all applied work already relies on; and, for its one unique prediction, on the ability to assess consciousness in a machine — a dependency that runs backwards relative to the usual assumption that machine consciousness waits on a theory of consciousness.
• Requires (not on this map) Two institutional facts, not scientific ones, gate the next decade of work on this theory. The first is access: IIT's posterior-hot-zone and connectivity claims are settled at electrode resolution, and electrodes enter human brains only where a clinical indication puts them, which makes epilepsy monitoring units and a small number of primate laboratories the entire supply of decisive data — the intracranial arm of the field's largest test was 34 patients. The second is that adversarial collaboration has to be a standing programme rather than a single grant. One round produced pre-registered predictions but not pre-registered concessions, which is exactly why both camps survived a result that substantially challenged key tenets of both. A funder willing to pay for the second and third rounds, and to require each camp to name in advance the outcome on which it would concede, is the binding requirement on the whole workback plan.
• Enables Enables the perturbational complexity index and its clinical descendants, which are already deployed in disorders-of-consciousness assessment and are the theory's most consequential output by a wide margin; enables a formal case against computational functionalism that gives functionalists a specific mathematical target rather than an intuition to dispute; and enables a structured panpsychism in which exclusion, not summation, decides which system is a subject.
• Adjacent Adjacent to global neuronal workspace theory, its designated rival and co-defendant in the one large adversarial test, on which it may turn out to be a complement rather than a competitor if the phenomenal and access questions really are separate; to higher-order, predictive-processing and attention-schema accounts, which share the explanandum and none of the method, and which unlike IIT convert readily into build specifications; to electromagnetic-field theories, which are substrate-committed in a different way and make the useful point that substrate-dependence is a family of positions rather than one; and to Orch OR, with which it agrees only that computational functionalism is false and disagrees about what the relevant physics is. The clinical measurement literature is adjacent in a stronger sense than any of these: the perturbational complexity index is IIT's descendant, is used without reference to the theory, and would survive the theory's refutation intact.

Frontier “IIT says a big enough computer will be conscious.” This is the most common claim made about IIT and it is the exact inversion of what the theory says. Established Findlay, Marshall, Albantakis, David, Mayner, Koch and Tononi — four of them from Tononi's own group — demonstrated formally that on IIT “it is possible for a digital computer to simulate our behavior, possibly even by simulating the neurons in our brain, without replicating our experience,” a conclusion they state “contrasts sharply with computational functionalism.” Established The reason is architectural, not a matter of scale: phi is computed over the physical cause–effect structure of the substrate, and a von Neumann machine executing a simulation has the causal architecture of a von Neumann machine, not of the thing simulated. Frontier IIT is therefore one of the most anti-machine-consciousness theories in the field, more restrictive than global workspace theory, more restrictive than higher-order theories, and more restrictive than predictive processing. Speculative The inversion happens because IIT is panpsychist, and panpsychism is read as permissive: if a thermostat has a little experience, surely a data centre has a lot. Established But phi is a measure of integration, not of information, and it is subject to exclusion — a system with vast information and low integration has phi near zero, and only the maximally irreducible substrate is a subject at all. Frontier An enthusiast citing IIT in support of machine consciousness has cited the one major theory that rules it out.

Established “COGITATE showed IIT beat global workspace theory, two predictions to nil.” This misreports the published paper and it is the highest-priority correction in this brief. Established The Nature 2025 conclusion is that the results “align with some predictions of IIT and GNWT, while substantially challenging key tenets of both theories.” Established IIT's connectivity prediction was contradicted in terms: “a lack of sustained synchronization within the posterior cortex contradicts the claim that network connectivity specifies consciousness.” Established GNWT's prefrontal category-decoding prediction was supported. Established GNWT took its own hit — none of the 655 electrodes measured the temporal profile it predicted — but that is a different claim from IIT winning. Frontier The two-to-nil framing traces to a preliminary conference announcement made in June 2023, nearly two years before peer review, and it has since migrated into circulation as a summary of the published result, including in the most-read encyclopaedic article on the theory. Established Anyone repeating it is quoting a conference slide.

Established “Phi has been measured in the human brain.” It has not, and it will not be. Established Phi is computationally intractable for systems of any interesting size; every clinical and experimental quantity in circulation is a proxy. Frontier The relationship between the proxies and the formalism is itself disputed, which is what Leung and Tsuchiya's separation of IIT-inspired from aspirational-IIT work is for. Frontier When a paper says it measured integrated information, the correct next question is which estimator, on what system size, with what error bound — and the third of those usually has no answer.

Frontier “PCI's clinical success confirms IIT.” Overstated. Established Farisco and Changeux argued in 2023 that the perturbational complexity index is compatible with global neuronal workspace theory; the title of their paper is the thesis. Established A measure that both theories can claim discriminates neither, and PCI's genuine clinical value is independent of whether the theory that inspired it is true.

Frontier “The 2023 open letter proved IIT is pseudoscience.” It is a signed preprint expressing a view, not a finding. Established It drew three substantial responses within ten weeks, one of them from Keith Frankish, an illusionist with no sympathy at all for IIT's phenomenology-first method, who nonetheless asked whether the theory is pseudoscience or appropriately anomalous science. Frontier A reported survey of sixty consciousness researchers found the letter divisive within the signatories' own community. Speculative The serious version of the charge is not that IIT predicts nothing — it pre-registered predictions in the field's largest test and it makes a specific claim about digital computers. Frontier The serious version is that its distinctive commitments are not currently testable while its testable commitments are not distinctive. State it that way and it is a real problem; state it as untestability and it is refutable in one sentence.

Established “The unfolding argument refuted IIT.” Four papers, one journal, three years, no resolution. Established Doerig, Schurger, Hess and Herzog in 2019, rebutted by Tsuchiya, Andrillon and Haun in 2020 and by Usher in 2021, with a counter-rebuttal by Herzog, Schurger and Doerig in 2022. Frontier Cite the chain or cite nothing.

Frontier “IIT 3.0 and IIT 4.0 are the same theory in tidier notation.” Not established and possibly wrong. Established The measure has changed across versions. Frontier Whether COGITATE's predictions derive from 3.0 or 4.0, and whether the connectivity failure bites both, was not resolved for this brief and should not be asserted either way.

Frontier Two provenance warnings, because this subject's citation trail is worse than it looks. Established The most-repeated sensitivity figure for the PCI cutoff, 94.7%, was read in a citing paper attributing it to the 2016 primary; this brief did not obtain the primary and reports the number only as a second-hand attribution. Frontier And during the research for this brief a publisher's article page served a different paper than the DOI requested — the identifier for the IIT 4.0 paper returning an unrelated article on mouse digit regeneration — while a bibliographic search interface returned five cryobiology papers in response to a query about the adversarial-collaboration protocol, with no error and no low-confidence signal. Speculative Both faults would have produced confident, wholly fabricated citations if left unchecked. Frontier In a literature where the ratio of commentary to data is this high, the failure mode is not that a claim is wrong but that a claim's provenance is invented downstream of the source, and the date discrepancy on the 2023 letter — one registry giving 16 September and one encyclopaedia 15 September for the same document — is the benign version of the same phenomenon.