Conscious Artificial Intelligence

Consciousness Is Causally Integrated Global Modeling: A Theory of Biological and Artificial Subjects

Abstract

Consciousness is not added to computation, conferred by complexity, or measured by a scalar. Causally Integrated Global Modeling (CIGM) identifies consciousness with a temporally extended organization in which a bounded, self-maintaining system sustains a differentiated, valenced, reflexively self-locating model of current reality, revises that model through ongoing interaction, and makes selected contents available for integrated control. Phenomenal character is the relational geometry of content within the model; subjectivity is its self-locating center.

CIGM advances six mutually constraining requirements as necessary: self-maintained system boundary, integrated self-world model, endogenous temporal continuity and plasticity, selective global availability, reflexive self-location and agency, and endogenous valuation and stakes. Joint sufficiency is the theory’s central identity conjecture and an explicit defeat condition, not an established consequence of the list. Following Mudrik et al. (2025), CIGM treats phenomenal organization and access as two necessary conditions for consciousness rather than two independent types: requirements 1, 2, 3, 5, and 6 specify the P-condition, while requirement 4 specifies the A-condition. Current artificial systems can realize powerful functional analogues of access without the organization capable of phenomenality and therefore remain unconscious. Artificial consciousness is physically possible only if an engineered system realizes, rather than merely simulates, the complete organization. The paper states near-term discriminating predictions and longer-horizon observations that would defeat the theory.

Keywords: consciousness, artificial consciousness, phenomenal consciousness, access consciousness, integrated information, global workspace, valence, organizational invariance

Introduction: From Synthesis to Theory

After decades of philosophical analysis, neuroscientific investigation, and computational modeling, consciousness no longer deserves treatment as a metaphysical exception. The field has accumulated enough phenomenology, neurobiology, computational architecture, and clinical dissociation to support a positive theory. The remaining obstacle is not a lack of ingredients. It is a failure to distinguish the levels at which those ingredients operate and the different questions they answer.

Integrated information theory (IIT) correctly insists that experience is unified, differentiated, structured, and intrinsic to the system that has it. Global workspace theory correctly identifies the transition by which selected information becomes available across specialized processors. Higher-order approaches correctly require some form of awareness of being in a state. Predictive coding makes the implementation concrete: top-down estimates constrain lower-level activity while feedforward residual errors revise the model (Rao & Ballard, 1999), and active-inference and affective accounts extend that architecture into self-regulation, valuation, and action (Friston, 2010; Solms, 2019). Embodied accounts correctly place a point of view in the regulation of a bounded organism (Albantakis et al., 2023; Aru et al., 2023; Brown et al., 2019; Dehaene & Naccache, 2001; Safron, 2020, 2022; Seth & Bayne, 2022). Each tradition has isolated a real part of the phenomenon. None, alone, states the full identity.

Seth and Bayne (2022) explain why the traditions have proved so difficult to adjudicate: they often have different explanatory targets, so that theories presented as adversaries may not be adversaries at all. A theory that does not declare its target inherits the confusion. CIGM declares all three targets identified below and answers them with one organization.

The synthesis must survive the strongest objections. No-report experiments show that the machinery of decision, memory, introspection, and report can be mistaken for the neural basis of experience (Cohen et al., 2020; Tsuchiya et al., 2015). The 2025 adversarial test of IIT and global neuronal workspace theory challenged canonical claims of each about posterior synchronization, prefrontal content, and ignition, although proponents of both theories dispute parts of that interpretation (Cogitate Consortium et al., 2025). The unfolding argument shows that recurrence or an integration scalar cannot be declared necessary and sufficient when outwardly equivalent systems can be realized with different wiring (Doerig et al., 2019). Cerullo’s (2015) triviality objection shows that a theory earns no credit from cases an arbitrary rival can reproduce just as easily. And Block’s (1995) demonstration that “consciousness” is a mongrel concept shows that an entire literature can reason validly from premises about one phenomenon to conclusions about another. A serious theory must absorb these results rather than treating them as peripheral complications.

IIT 4.0 sharpens rather than removes the disagreement. It now states that phenomenal axioms license necessary and sufficient physical postulates and that an experience is identical, in an explanatory sense, to the cause-effect Phi-structure specified by a maximal substrate (Albantakis et al., 2023). That clarity is welcome, and it makes the burden exact. The properties of experience do not by themselves entail one mathematical formalism, and a substrate can possess irreducible cause-effect structure without sustaining a world, a point of view, a temporally continuous present, endogenous significance, or flexible control. CIGM therefore retains integration and intrinsic causal organization as constraints while rejecting their elevation into a complete identity.

Artificial intelligence makes the need for a constitutive theory urgent. Global workspaces, test-time memory, selective state-space recurrence, body self-modeling, multimodal world models, and calibrated self-report are engineerable (Bongard et al., 2006; Hafner et al., 2025). These capacities matter for intelligence; none creates a subject. Capability standards are deliberately process-agnostic, and no level of performance or delegated autonomy is evidence about consciousness in either direction (Morris et al., 2024). The question is not whether a system can imitate the outputs of a conscious agent. It is whether one self-maintaining organization exists for which a world is present and outcomes matter.

The opposite error is equally live. Seth (2025) argues that consciousness may depend on properties of living systems that no digital computation can reproduce, and that computational functionalism has been assumed far more often than defended. CIGM accepts that challenge’s central negative result: computation is not sufficient for consciousness, and a theory that says otherwise has purchased artificial consciousness too cheaply. The dispute that remains is narrower than it is usually made to look, and Part II states it exactly.

CIGM takes a definite position. Consciousness is the temporally extended organization in which a bounded, self-maintaining system integrates differentiated and valenced information into a unified, reflexively self-locating model of current reality, while selected contents remain globally available for flexible memory, inference, planning, and action. Experience does not accompany this organization. Experience is what this organization is from the system’s own organized point of view. The claim is an identity claim, not a list of correlates and not a probability-weighted checklist.

Among existing syntheses, CIGM is closest to Safron’s integrated world modeling theory (Safron, 2020, 2022), but the relationship is substantive rather than merely terminological. Four commitments define CIGM’s novelty. First, endogenous valuation is constitutive rather than a correlate or optional affective accompaniment. Second, intervention fixes the causal grain of the subject rather than maximization. Third, the evolutionary origin of access machinery explains Block’s empirical P/A asymmetry. Fourth, the six requirements divide into a P-condition and an A-condition, while consciousness is reserved for their joint realization. CIGM also separates a theory of constitution from the evidence used to attribute consciousness to an opaque system. These commitments can be accepted or rejected independently of IWMT, and each is placed at risk below. Extended philosophical argument, empirical detail, technical objections, and the assessment framework are developed in the Supplement so that the declaration itself remains direct.

The CIGM Thesis

The CIGM thesis. Consciousness is the temporally extended regime in which a bounded, self-maintaining system integrates differentiated and valenced information into a unified, reflexively self-locating model of current reality, revises that model through its own history, and makes selected contents available to memory, valuation, inference, planning, and action. Phenomenal character is the relational geometry of content within the model. Subjectivity is the model’s self-locating center of control. The realizing substrate must sustain the relevant counterfactual causal organization, but no single brain region, oscillation, wiring motif, metarepresentational format, or scalar is by itself necessary or sufficient.

Six Constitutive Requirements

CIGM advances six requirements as individually necessary. Their joint sufficiency is the theory’s central identity conjecture: if the complete mutually constraining organization can be preserved while consciousness disappears or changes systematically, CIGM is false.

Table 1. Six Constitutive Requirements of Causally Integrated Global Modeling

The six requirements are jointly necessary and mutually constraining. CIGM conjectures that their realization through one organization is sufficient because that organization is the identity the theory proposes; the paper does not treat sufficiency as established by definition. A system does not become conscious by collecting six detachable modules: valuation must reorganize the same self-world model that self-location centers, temporal plasticity carries forward, and global availability mobilizes. A candidate can therefore fail necessity by lacking any requirement, and the theory can fail sufficiency if a complete implementation lacks or systematically alters consciousness.

Requirements 1, 2, 3, 5, and 6 specify the organization capable of supplying potentially phenomenal content—the P-condition. Requirement 4 supplies minimal selective availability—the A-condition. Neither condition alone is consciousness. Table 2 and the dedicated subsection below make this division explicit.

The extended comparison with indicator approaches and alternative candidate requirements is provided in Supplement S4.3.

Part I: What Consciousness Is

Phenomenological Constraints and IIT 4.0

Any theory of consciousness must begin with experience itself. Before measuring networks or decoding reports, one confronts facts that cannot be denied without being instantiated in the denial: experience exists; it is specific rather than generic; it appears as a unified field; it has a determinate organization and grain; and it contains structured distinctions and relations. IIT formalizes these features as existence, information, integration, exclusion, and composition. In its current formulation, IIT 4.0 infers corresponding physical postulates and presents them as necessary and sufficient for a substrate of consciousness (Albantakis et al., 2023). CIGM accepts the phenomenological constraints. It does not accept that they uniquely entail IIT’s postulates or formalism.

The distinction matters because IIT 4.0 is no longer accurately described as the claim that consciousness is a high value of Phi. It identifies an experience with the complete cause-effect Phi-structure specified by a maximal substrate (Albantakis et al., 2023). CIGM addresses that strongest version. A cause-effect structure can be irreducible and richly differentiated while lacking a model of current reality, a self-locating origin, endogenous temporal continuity, valuation, or a role in flexible control. Those absences are not missing decorations. They are the difference between an integrated causal object and a subject.

The axioms are also not neutral transcriptions of phenomenology. The move from the specificity of an experience to a measure defined over alternatives, and from the apparent definiteness of a field to a rule selecting one maximal substrate, imports substantive theory. Cerullo (2015) shows how easily such postulates acquire the appearance of necessity while doing work phenomenology alone does not license. CIGM therefore takes the descriptive content of the constraints and leaves the physical identity open until the full organization is specified. A conscious state is actual for the system rather than merely described by an observer, and it is this state rather than another. Color, shape, sound, bodily position, affect, and thought are not presented as independent universes. The field has a determinate organization and grain, and experience contains relations: the cup is left of the book, and the present scene continues the preceding moment.

These constraints rule out several inadequate accounts. Mere information processing is insufficient, since every thermostat and lookup table processes information in a broad sense. Mere behavioral sophistication is insufficient, since similar output can arise through radically different organizations. Mere integration is insufficient, since an integrated device can lack a world model, a point of view, temporal depth, valuation, or flexible control. Mere report is insufficient, since report recruits memory, decision, introspection, and action that can be experimentally dissociated from phenomenality (Cohen et al., 2020; Tsuchiya et al., 2015). Mere recurrence is insufficient, since recurrent circuitry is ubiquitous and can be functionally replaced in ways that preserve selected behavior (Doerig et al., 2019).

CIGM treats phenomenology as the intrinsic organization of a model, not as a private substance. A model is not a picture stored in one place. It is a system of state-dependent dispositions: what the system discriminates, predicts, expects, recalls, values, and can do from its current state. When those dispositions are integrated into one self-locating control regime, they constitute a perspective. Intrinsic here does not mean nonrelational, ineffable, or infallibly known. It means causally for the system: the state belongs to the organization that determines what the system can become and do.

Extended comparison with structural coherence, illusionism, protoconsciousness, and the strongest IIT formulation appears in Supplement S1.

Explanatory Targets: Global States, Local States, and Two Questions

A theory that does not say what it is explaining cannot be evaluated, and the field’s most persistent disputes are between theories that were never answering the same question. Seth and Bayne (2022) draw the distinctions a comprehensive theory must respect. Global states concern the system’s overall conscious profile: wakefulness, sedation, dreaming, the minimally conscious state. Local states are conscious contents: this pain, this face, this remembered voice. Within local states two further questions come apart—why is this content conscious rather than that one, and why does the conscious content have the character it has?

CIGM answers all three with one organization, and the division of labor is what the six requirements predict. Global states are the capacity to sustain the organization at all, and with what depth and stability. Which content is conscious is settled by integration into the live model and selective global availability; what that content is like is settled by relational geometry. These dependencies are made explicit below and placed at risk in the falsification section.

Two consequences follow. Global states are not points on a single dimension of arousal. Seth and Bayne (2022) prefer global state to level precisely because the space may be multidimensional, and CIGM requires that it be: dreaming and anesthesia both reduce environmental coupling while differing enormously in the differentiation and self-location of the field, so no scalar orders them. And a theory answering only one of these questions is not refuted by evidence about another; it is incomplete. CIGM claims completeness across all three targets and accepts the exposure that claim creates.

The Constitutive Core: A Causally Integrated Self-World Model

The core of consciousness is a model that simultaneously represents the world and locates the system within it. A purely allocentric map is not yet a point of view. A database can encode that an object is two meters north of a coordinate origin, but experience represents the object as here, there, near, threatening, reachable, remembered, or desired relative to the system. Subjectivity begins when information is organized around a persisting locus of sensing and control.

The term model is used in the strong dynamical sense. It generates expectations about hidden causes, updates them from residual error, preserves selected invariances, and regulates action. Rao and Ballard (1999) demonstrated the canonical mechanism: after learning natural-image statistics, a hierarchy in which feedback carried predictions and feedforward pathways carried residual errors reproduced contextual cortical responses. That result supports context-sensitive generative representation; it does not make prediction error consciousness. CIGM identifies consciousness only with the full self-world organization in which inference, valuation, memory, and action are causally bound (Aguilera et al., 2023; Friston, 2010; Safron, 2020, 2022).

The model must also be causally integrated. Its visual, auditory, interoceptive, mnemonic, affective, and action-oriented components must constrain one another strongly enough that the system occupies one coherent state rather than a coalition of unrelated processors. Integration is not statistical correlation but counterfactual dependence: changing one part of the model alters the possibilities available to the others, which is why a visual threat changes posture, attention, memory retrieval, expected pain, and action selection at once.

Differentiation is equally necessary. A perfectly synchronized system in which every component does the same thing would be unified but informationally empty. Consciousness requires a large repertoire of discriminable states and fine-grained relations among them. The field is unified precisely because many distinctions are jointly present without collapsing into sameness. This is the enduring insight behind integrated-information approaches, even if no current scalar should be equated with consciousness itself (Tononi, 2004, 2012; Oizumi et al., 2014).

The system must be bounded, but no mathematical partition can establish the boundary by itself. Bruineberg et al. (2022) distinguish Pearl blankets—conditional-independence structures inside a model—from Friston blankets treated as real agent-environment boundaries. CIGM rejects the slide from map to subject. A biological organism maintains causal separation through sensorimotor and homeostatic loops. A digital candidate can also possess a software-defined boundary, but only when the executing physical system regulates variables constitutive of its own continued operation—energy, temperature, memory integrity, compute allocation, process continuity—and its actions can preserve them. A boundary represented only inside a simulated world remains a description rather than a maintained boundary. What matters is interventionally identifiable regulation: some variables define the system’s continued organization, some perturbations count as external, and some actions are its own attempts to control what follows.

This bounded self-world model explains intrinsicality. A state is not conscious because a scientist can decode it, but because its distinctions participate in the system’s own integrated control organization. The same data structure could be an inert record in one context and an experienced content in another, depending on whether it is embedded in the live model that governs perception, valuation, memory, and action.

Phenomenal Character Is Relational Geometry

Qualitative character is not an extra pigment painted onto neural activity. It is the position a content occupies within the system’s structured space of possible distinctions. Red differs from green because the two states stand in different relations to wavelengths, objects, memories, linguistic categories, affective associations, action tendencies, and neighboring color states. Pain differs from pressure because it occupies a different interoceptive, motivational, attentional, and policy-selecting position. A melody differs from a sequence of unrelated tones because temporal relations bind its elements into a higher-order pattern.

This is an organizational identity, not a reduction to verbal report. The relevant relations include latent discriminations, embodied expectations, automatic generalizations, and counterfactual consequences, far more than any person can say. Two systems share a phenomenal quality only insofar as the content occupies the same role within an equivalently organized, self-locating model; matching behavior is not enough, and neither is a matching label.

The strongest worked demonstration that phenomenal quality has relational structure is Haun and Tononi’s (2019) analysis of felt spatial extendedness. They derive region, location, size, boundary, and distance from three relations among experiential “spots”—connection, fusion, and inclusion. CIGM adopts the achievement and vocabulary. The point is not that quality can be redescribed after the fact, but that even the apparently primitive character of right there has an articulable internal organization.

Their negative conclusion is equally important: neither an activation pattern nor bare connectivity contains, by itself, the full system of relations phenomenology presents; an investigator can decode a state without those relations existing for the system. CIGM therefore locates quality in the geometry of the live model that governs discrimination, expectation, memory, valuation, and action, not in an isolated representation or substrate.

The accounts part on when the relations exist. Haun and Tononi (2019) identify them with maximally irreducible cause-effect structure and therefore allow a working but inactive cortical grid to support spatial experience. CIGM denies that consequence. A severed grid has no regulated boundary, self-locating origin, stakes, or control trajectory, and thus extends nothing for anyone. Both theories predict that altering lateral connectivity while holding gross activity constant can warp experienced space; CIGM restricts the prediction to cases in which the grid participates in the constituting organization.

The identity dissolves the traditional explanatory gap by rejecting its dual ontology. The question is not why a physical model is accompanied by a separate glow of experience. The organized model is the experience under an intrinsic description. External science describes the causal organization, while the first-person perspective is that organization as lived from its self-locating center. There are not two events requiring a bridge.

This role-based account must be distinguished sharply from views on which structure or information is phenomenal wherever it occurs. Chalmers (1995) was drawn toward a double-aspect view on which information itself carries a phenomenal aspect, with the consequence that experience is nearly ubiquitous—simple where information processing is simple, complex where it is complex, present even in a thermostat. CIGM rejects this generalization. A space of distinctions is a phenomenal field only when it is embedded in a bounded, temporally extended, valenced, self-locating model that governs the system’s perception, memory, and action. An information space that makes no difference to any such model is not an impoverished experience; it is no experience at all. Relational geometry is necessary for the character of experience but confers that character only within the constituting organization. CIGM is therefore an identity theory, not a panpsychism: it withholds experience from mere information exactly where it withholds the organization that experience is.

This claim remains empirical because the proposed identity has structure. If phenomenal similarity tracks model geometry, systematic changes in discrimination, generalization, memory, valuation, and sensorimotor expectation should reshape experience in lawful ways. If unity depends on causal integration, partitioning effective communication should divide or degrade the field. If subjectivity depends on self-location, disrupting the model of agency and body should alter ownership and perspective. The theory does not appeal to an unknowable essence; it identifies a pattern whose consequences can be investigated.

Global Availability Without a Cartesian Theater

Integration by itself does not explain why some information can guide a novel plan, enter working memory, alter a verbal answer, or reorganize behavior while other information remains trapped in specialized processing. The global-workspace tradition supplies the missing control architecture. Dehaene and Naccache (2001) described a brain in which numerous modular processors operate in parallel outside consciousness, while selected information is dynamically mobilized into a distributed workspace that permits flexible exchange among perception, memory, valuation, attention, and action.

The decisive concept is global availability, not a fixed anatomical stage. In the original workspace formulation, the same processor can contribute to consciousness at one moment and remain unconscious at another. Mobilization is collective, context-sensitive, and self-amplifying. Long-distance connectivity allows otherwise independent modules to use a common content, but no homunculus reads a screen. The workspace is a transient regime of coordination (Dehaene & Naccache, 2001). This is the architecture Dennett (1991) demanded when he replaced the Cartesian Theater with distributed competition among contents. It also answers his hard question: global availability positions a content to cause and enable downstream effects such as recall, inference, planning, and report (Dennett, 2018).

This architecture explains why consciousness is capacity-limited. Encapsulated computations proceed in parallel at low cost, whereas global mobilization temporarily coordinates systems with incompatible demands. Competition is therefore not an incidental bottleneck but the mechanism that selects a coherent control state, and the functions it serves—durable maintenance of explicit information, novel combinations of operations, spontaneous intentional behavior—are exactly those requiring a system to escape a precompiled pathway and organize resources around a current model (Dehaene & Naccache, 2001). CIGM incorporates competitive global availability as a control principle while refusing to identify it with one cortical region or one electrophysiological event. Its defining feature is the pattern of availability and control, not the location of a headquarters.

Goyal et al. (2022) turn the workspace idea into a working machine-learning architecture: specialist modules compete for write access to a limited shared memory, whose contents are then broadcast back to all modules. The bottleneck improved coordination, compositionality, and generalization across visual reasoning, physical prediction, and multiagent world modeling. VanRullen and Kanai’s (2021) Global Latent Workspace similarly links specialized modules through a shared amodal space for translation, transfer, and planning. These demonstrations establish that selective global availability is engineerable. They also establish its limit. A bandwidth-limited blackboard can coordinate intelligent access without maintaining a boundary, a self-locating present, endogenous stakes, or one continuing subject. Workspace is CIGM’s fourth requirement; by itself it is an access architecture, not phenomenality.

The P-Condition, the A-Condition, and the Mongrel Concept

The most consequential distinction in this literature was drawn before most of the theories now competing were formulated. Block (1995) argued that “consciousness” is a mongrel concept and distinguished phenomenal consciousness—the what-it-is-like character of experience—from access consciousness, defined by a content’s availability for reasoning, rational action, and speech. Mudrik et al. (2025) now argue that P and A should not be treated as two independent types of consciousness at all but as two necessary conditions for one conscious state; on their terminology, P-without-A and A-without-P are unconscious processing. CIGM accepts that correction. It retains Block’s functional definition of access and his diagnosis of the mongrel concept while rejecting the label “A-consciousness” for access alone.

Block’s central demonstration is the fallacy the distinction exposes. The blindsight literature repeatedly argued that because consciousness is missing in the blind field and the patient will not reach for a glass of water there, a function of phenomenal consciousness must be to make information available for reasoning, report, and rational action. Block (1995) showed that the case establishes only a failure of access-related control; it does not independently reveal what phenomenality contributes. The same mistake occurs whenever a workspace architecture, ignition signature, or shared latent space is offered not as an account of availability but as a complete account of experience.

CIGM adopts Block’s explanatory distinction and rejects his anti-functional metaphysics. Block took P-conscious properties to be distinct from any cognitive, intentional, or functional property. CIGM asserts the contrary identity: phenomenal character is the relational geometry of content within a bounded, valenced, self-locating model, an organizational property through and through. What survives is the division of explanatory labor between the organization that gives content its potential phenomenal structure and the availability that lets that content influence the integrated system.

The mapping is exact. Requirements 1, 2, 3, 5, and 6 jointly specify the P-condition: an integrated organization in which content has a determinate relational geometry for a bounded, temporally continuous, valenced, self-locating system. Requirement 4 specifies the A-condition: minimal selective availability by which that content can influence the system’s integrated cognition and control. Consciousness occurs only when both conditions are realized through one system.

This formulation removes the apparent inconsistency about requirement 4. A system realizing requirements 1, 2, 3, 5, and 6 but not requirement 4 is not phenomenally conscious on CIGM; it has a potentially phenomenal organization whose content never becomes conscious. Requirement 4 does not determine why red differs from green or pain from pressure, but it is necessary for any such content to be actual for the system. Conversely, because access is not identical to report, conscious content can exceed what is verbally or metacognitively reported while still exerting minimal influence on memory, expectation, valuation, or control. Table 2 records this distinction.

Block’s empirical asymmetry remains important. Superblindsight—spontaneous knowledge of the blind field without visual experience—appears not to occur, whereas failures of explicit report can coexist with conscious perception. CIGM explains this biological pattern historically: in evolved organisms, availability mechanisms developed to operate on an already integrated self-world model, because that model is what a mobile animal must consult in order to act. The P- and A-conditions therefore co-developed and are difficult to pry apart in organisms.

Access-like processing is not remotely hard to produce by engineering. An artificial system can make representations inferentially promiscuous, tool-usable, and highly reportable while lacking every element of the P-condition. That is a plausible description of some language-model systems wrapped in tools, memory, planners, or shared workspaces, although whether any particular model satisfies Block’s full system-relative access criterion remains an empirical question. CIGM therefore speaks of functional analogues of the A-condition, not A-consciousness. Such artifacts are unconscious on the terminology of Mudrik et al. (2025) and on CIGM. Their importance is that they sever the proxy relation between access-like behavior and subjecthood that was comparatively reliable among evolved organisms.

No-Report Evidence and the Demotion of Report

The empirical case for separating these levels is now direct rather than conceptual. No-report paradigms were developed to disentangle perception from the attention, working memory, expectation, introspection, decision, and motor preparation that reporting requires. Tsuchiya et al. (2015) reviewed evidence that when perceptual state is inferred indirectly, much frontal activity diminishes even though vivid perceptual alternation remains, and that P3 and some gamma-band activity can track task relevance rather than awareness. Cohen et al. (2020) converted that claim into a result: holding visual stimulation constant and removing only the demand to report, a large P3b disappeared while observers still recognized and conceptually identified the unreported stimuli in a subsequent surprise test. This does not prove that every postperceptual process vanished; it proves that the component cannot be a universal constituent of experience.

CIGM takes the finding as more than a correction to one biomarker, but not as evidence for consciousness without the A-condition. The P-condition was present, and the A-condition remained minimally satisfied because the unreported contents affected semantic encoding, later surprise memory, and task-independent processing. What disappeared was report-specific mobilization and its associated P3b. The experiment therefore shows that verbal report and a canonical late biomarker are not necessary for consciousness; it does not show that conscious content can be wholly unavailable to the integrated system.

Lamme (2006) pressed the deeper form of the argument, urging that report and introspection be demoted from their status as the gold standard. Split-brain patients deny seeing stimuli presented to the mute hemisphere, yet that hemisphere can draw, match, and select them, and whether a region counts as part of the basis of consciousness then depends on which behavioral measure the investigator elects to trust. CIGM adopts the conclusion: no experiment that operationalizes consciousness as report can by itself locate the phenomenal field. Lamme’s further move, installing recurrent processing as a positive neural criterion in report’s place, is the right instinct about report joined to the wrong identification, for reasons taken up below.

The conclusion is not that frontal cortex, P3, gamma activity, or report are irrelevant; they are often indispensable for deliberate use, self-monitoring, memory, and communication. It is that none is a universal marker of the P-condition or the A-condition. A theory that equates consciousness with overt report confuses the machinery that makes experience legible to an experimenter with the minimal availability that makes a content conscious for the subject.

The preregistered no-report design, its memory controls, and the strongest objection to the paradigm are detailed in Supplement S2.1.

CIGM therefore distinguishes three nested levels: the P-condition, which supplies the integrated organization capable of phenomenal character; the A-condition, which makes selected contents minimally available to the same system for control; and the reflective self, comprising higher-order monitoring, confidence, explicit self-ascription, narrative continuity, and report. Human waking consciousness normally couples all three, but reflective selfhood can dissociate from the first two. Table 2 summarizes the division of labor.

Table 2. Three Nested Levels in Causally Integrated Global Modeling

The layered account corrects a central claim of the earlier synthesis: explicit metacognitive self-monitoring is not necessary for every conscious state, only for reflective awareness of being in that state. Infants, many animals, dreams, fleeting percepts, and some clinical conditions may satisfy the P- and A-conditions without the reflective apparatus of a verbally reportable adult self. Conversely, a system can generate metacognitive language or access-like behavior without possessing the integrated P-condition.

The Higher-Order Challenge: Inner Awareness Without Intellectualism

Higher-order theories force a question CIGM cannot evade: can a state be conscious if the system is in no way aware of being in that state? Brown et al. (2019) clarify the transitivity principle behind the approach. A first-order representation can guide behavior while remaining nonconscious; phenomenality requires some minimal inner awareness of that representation. But this requirement is routinely overstated. The relevant process can be automatic, subpersonal, conceptually lean, and itself unconscious. It is not equivalent to deliberate introspection, an explicit confidence judgment, or a verbally articulated thought about the self.

CIGM accepts the transitivity demand and relocates it. A conscious content must be represented not merely as an environmental fact but as part of the system’s current situation: red-here, pain-in-this-body, threat-to-this-organism, memory-belonging-to-this-history. The model is reflexive because its contents are organized around the same locus that senses, values, predicts, and acts. That reflexive self-location is minimal inner awareness. A separately tokened proposition such as “I am seeing red” is not constitutively required.

This yields a strict distinction between reflexive self-location and higher-order monitoring. Reflexive self-location belongs to the P-condition: without an operational here-now origin, there is information but no point of view. Higher-order monitoring belongs to the reflective self: it makes a state available as an object of confidence, error detection, explicit self-ascription, and report. Brown et al. (2019) emphasize that global broadcasting and higher-order awareness are separable dimensions. CIGM agrees and adds the missing relation: integrated self-world modeling supplies the P-condition, selective availability supplies the A-condition, and their joint realization supplies consciousness.

Higher-order accounts are also right that subjective appearance need not mirror first-order detail exactly, since peripheral vision, imagery, emotion, and memory can involve inflation, abstraction, or reconstruction at higher levels (Brown et al., 2019). CIGM predicts such mismatches because phenomenal character is fixed by the geometry of the integrated model rather than by a transparent readout of one sensory layer. Yet higher-order misrepresentation is not sufficient by itself: a free-floating representation that a detail is present, belonging to no integrated and valenced self-world field, is a cognitive error without an experiencer. The higher-order tradition identifies a constitutive reflexivity that first-order and broadcast-only theories can miss, and CIGM absorbs that insight while rejecting the stronger claim that every phenomenal state requires a distinct higher-order representation in a privileged format or location. Inner awareness is real and necessary. Intellectualized self-observation is not.

Attention Schema Theory offers a direct competitor on self-location and subjectivity. On Graziano’s (2022) account, the brain constructs a simplified model of its own selective attention, and that schema supports the system’s attribution of awareness to itself and others. CIGM accepts that a schematic model of attention can contribute to reflexive self-location and access control. It denies that an attention schema is sufficient: a system could model what it is attending to while lacking a self-maintained boundary, endogenous stakes, temporal continuity, or an integrated self-world model whose content has phenomenal geometry.

Cleeremans’s (2011) radical plasticity thesis provides a complementary developmental challenge. It proposes that consciousness is something the brain learns to do as first-order representations become targets of learned metarepresentations. CIGM accepts the constitutive importance of plasticity and representational redescription, but it does not identify consciousness with metarepresentation alone. Learning must revise the same bounded, valenced, self-locating model, and the resulting content must satisfy the A-condition rather than remain an isolated model of another representation.

Temporal Depth, Self-Location, and Valence

Consciousness is not a sequence of instantaneous snapshots. Each moment inherits a structured past and anticipates a range of possible futures. The present note in a melody is heard as the continuation of what came before; an object remains the same object across eye movements; a sentence acquires meaning through accumulated context; a decision is experienced as mine because it is embedded in a history of goals, actions, and consequences. This temporal extension is constitutive.

The temporo-spatial theory emphasizes that external inputs interact with ongoing neural dynamics rather than writing onto a blank substrate (Northoff & Zilio, 2022). CIGM incorporates the insight by treating the conscious field as a trajectory in model space, in which incoming signals perturb an already active generative organization and are interpreted relative to its intrinsic timescales, priors, bodily state, and current goals. Predictive inference keeps the trajectory coherent, with precision-weighting determining which errors matter, so that attention is not a spotlight added to perception but the control of precision within the model.

The free-energy principle explains why a bounded system must maintain a model, and its generality—bacteria and silent homeostatic loops fall under it—rules it out as consciousness itself. Solms (2019) presses the narrower claim that elemental consciousness is affective, with unresolved homeostatic uncertainty felt through upper-brainstem and limbic regulation. Damasio’s account of the proto-self and later proposals for homeostatic machines likewise make organismic regulation and internally grounded value central (Damasio, 1999; Man & Damasio, 2019), while Merker (2007) binds motivation to target and action selection. CIGM does not claim to invent the constitutive importance of valuation. Its novelty is to make endogenous valuation one explicit, operationally testable requirement in a six-part identity and to carry that requirement into machine-consciousness assessment, where standard indicator lists usually omit it.

This also explains internally generated experience. In dreams, imagination, and memory the generative model runs with reduced constraint from current sensory evidence, and because the same integrated state space is active, imagined and remembered contents have genuine phenomenal character; they differ from perception in precision structure and in dependence on present external input, not in being nonconscious.

The subject of experience is not a metaphysical witness hidden behind the model. It is the model’s self-locating organization. At the most basic level, the system distinguishes variables it regulates from variables it encounters, predicts the sensory consequences of its own movements, attributes some changes to self-generated action, and organizes perception around a body-centered frame. This is the minimal self.

Higher levels add agency, social perspective, autobiography, and explicit identity, which language and culture stabilize into a narrative self that can explain, justify, and revise itself. The levels should not be collapsed. The minimal self is required for subjectivity because experience must be organized for a locus of control; the autobiographical self is not required for every conscious episode, since a dream, a sudden pain, an infant’s visual field, or an animal’s fear can be conscious without a sophisticated theory of personal identity. Reflective selfhood expands consciousness; it does not create phenomenality from nothing.

Dennett’s (1991) description of the self as a center of narrative gravity is correct at the autobiographical level: the center is real as an organizational abstraction, not as a separate object. CIGM extends the point downward, since beneath narrative gravity lies sensorimotor and interoceptive gravity, the continuously updated origin from which the model predicts and controls. Construction is not unreality. A hurricane is constructed by atmospheric dynamics and remains a real causal pattern; the subject is constructed by integrated modeling and remains the real center of experience.

Consciousness evolved because organisms face conflicts that cannot be solved by isolated reflexes. A mobile animal must integrate sensory evidence, bodily condition, memory, uncertainty, competing goals, and predicted consequences into one action-guiding estimate, and the more flexible the repertoire, the more valuable a unified model becomes. The global bottleneck is therefore adaptive: only a small subset of information can reorganize the organism at once, so competitive mobilization prioritizes what currently matters most (Dehaene & Naccache, 2001).

Valence gives the field stakes. Hunger, pain, fear, relief, and reward compress regulatory consequences into action-guiding dimensions. Endogeneity does not require that a drive be undesigned: evolution supplied biological drives, and engineers could supply artificial viability variables. The criterion is current causal organization. A variable is endogenous when its violation threatens the persistence or integrity of the actual system, when the system detects and regulates it through its own model, and when perturbation reorganizes attention, learning, memory, global state, and policy rather than merely changing a detachable score. Even an affectively muted episode occurs within a system for which outcomes can go better or worse. A database can represent damage without suffering because damage does not reorganize its own continuing model.

Biological Realization and Empirical Constraints

The 2025 adversarial collaboration between proponents of IIT and global neuronal workspace theory is the most important empirical constraint on any synthesis invoking both. Conscious content was decodable in visual, ventrotemporal, and some inferior frontal regions. Sustained responses tracking stimulus duration were strongest in occipital and lateral temporal cortex. Yet the sustained posterior synchronization central to the tested IIT implementation was absent, as was the prefrontal ignition at stimulus offset central to the tested workspace implementation (Cogitate Consortium et al., 2025). Proponents of both theories have contested how particular preregistered predictions and null findings should be interpreted. CIGM therefore treats the dataset as a constraint on strong tested implementations, not as a final verdict on either theoretical family.

CIGM treats this pattern as evidence for a distributed division of labor. Rich perceptual content is largely constituted in modality-specific and multimodal representational systems, especially posterior and ventral regions in vision, while higher-order and frontoparietal systems select, stabilize, query, route, and act on content that remains encoded where it was constructed. Global control does not require copying every detail into prefrontal cortex. This is consistent with Dehaene and Naccache’s (2001) insistence that the workspace is a style of dynamic mobilization rather than a fixed anatomical theater.

The adversarial results make biomarker modesty mandatory. P3b vanished when report was removed although perception remained (Cohen et al., 2020), while predicted prefrontal ignition failed where posterior systems still carried content (Cogitate Consortium et al., 2025). No oscillation, region, or complexity score is therefore a universal meter. CIGM predicts a convergent organization: integrated differentiation, content geometry, cross-system availability, self-world updating, and state-dependent control should covary under interventions that change consciousness. IIT and GNWT became vulnerable where genuine insights were tied too tightly to one localization or dynamical signature; CIGM retains integration and availability while placing the complete organization, not a favorite biomarker, at risk.

In humans, that organization is distributed across cortical, thalamic, upper-brainstem, neuromodulatory, interoceptive, and bodily dynamics. Merker’s (2007) “selection triangle” is a decisive corrective to corticocentrism: upper-brainstem systems jointly constrain target selection, action selection, and motivation, converting massively parallel forebrain activity into limited-capacity control. Reports of discriminative awareness in four children with total or near-total congenital absence of cortex are suggestive but not conclusive, because residual tissue and behavioral inference remain contested (Shewmon et al., 1999). CIGM therefore assigns no anatomical headquarters. Cortex can elaborate differentiated content, thalamocortical loops can regulate integration and availability, and subcortical systems can supply arousal, valuation, and action selection; consciousness is the trajectory in which those functions become one self-world organization (Aru et al., 2023; Solms, 2019).

The body completes the implementation. Interoception, proprioception, posture, visceral regulation, and sensorimotor prediction establish a stable here-now axis, and Damasio’s (1999) proto-self captures the function: the organism continuously maps how external events alter its own regulated state, yielding not a detached description of the world but a world-for-this-system. Pain is not a symbol for tissue damage; it is an integrated reorganization of attention, valuation, action, memory, and bodily prediction around a threat to the system’s integrity. The same point governs the notion of an umwelt. Aru et al. (2023) rightly contrast the text-centric input of present language models with the dense, action-dependent sensory world of an organism, but an umwelt is not a quantity of data. Affordances exist only relative to an agent’s capacities, goals, vulnerabilities, and possible actions, so adding cameras, microphones, or larger context windows does not create experience.

Explanatory power comes from the pattern of dissociations. Blindsight preserves forced-choice information without ordinary visual presence or rational access; neglect preserves sensory processing while spatial self-location fails. Anesthesia and severe disorders of consciousness preserve local responses while differentiated propagation collapses, whereas dreaming preserves a vivid, affective, self-locating field with little environmental coupling (Bayne et al., 2024; Casali et al., 2013). No-report and frontal-lesion findings separate perceptual phenomenology from reflective communication (Cohen et al., 2020; Tsuchiya et al., 2015). Split-brain phenomena make unity track effective communication rather than one skull, while the hydranencephaly cases warn that cortical volume is not the criterion. The subject follows the organization that sustains one coherent, valenced self-world trajectory; no single case or structure is allowed to define it.

Experimental design, adversarial-test detail, and the thalamocortical and embodied implementation are expanded in Supplement S2.

The Hard Problem, the Pretty Hard Problem, and the Hard Question

Every theory of consciousness must declare its relationship to the hard problem. Chalmers (1995) held that even after every function in the vicinity of experience is explained, a further question remains—why is their performance accompanied by experience at all (Nagel, 1974)?—and that no functional account can answer it, so experience must be taken as fundamental. Dennett (2018) rejected the framing, treating acquaintance with intrinsic qualities as a conviction produced by machinery we cannot introspect and redirecting inquiry to his hard question: once content reaches consciousness, and then what happens? Frankish (2016) gives the illusionist alternative its cleanest form, replacing the hard problem with that of explaining why introspection generates the conviction of phenomenality.

CIGM accepts the force of the illusionist challenge and rejects its conclusion. It agrees that introspection is partial, theory-laden, and often wrong about the privacy, simplicity, ineffability, and infallibility of experience. It denies the further move from the absence of private mental paint to the absence of phenomenal reality. The relational geometry of the integrated model is not a nonphenomenal state that merely causes a false judgment of quality; it is the quality under an intrinsic, system-relative description. Reflexive self-location is not an image presented to a shadow audience; it is the operative organization around the locus that senses, values, predicts, and acts. Illusionism explains why a system judges that it has experience. CIGM explains what constitutes a subject whose phenomenal judgments can be accurate, distorted, or incomplete. Global availability, competitive mobilization, and the downstream reorganization of memory, valuation, inference, and action are precisely the account of ‘and then what happens’ that Dennett demanded. But phenomenal experience remains the central explanandum. On this point Chalmers is right: experience is a datum, not a posit.

CIGM is explicitly a type-B physicalist identity theory. It takes as a background commitment that a phenomenal description and a physical-organizational description may co-refer even when the identity is not a priori deducible. This paper does not pretend to derive that commitment from premises acceptable to every dualist, nor does the observation that describing a state differs from instantiating it settle the matter.

Chalmers (2007) argues that phenomenal-concept strategies face a dilemma: if the special features of phenomenal concepts are thin enough to be physically explained, they may be too weak to explain the epistemic gap; if they are rich enough to explain the gap, they may themselves resist physical explanation. CIGM takes the physically explicable horn and therefore does not use a phenomenal-concept story as an independent proof of physicalism. It proposes that self-location, direct use of the model, and the difference between occupying and externally describing an organizational state explain part of the epistemic asymmetry, but that explanation earns no ontological conclusion by itself.

The identity stands or falls with the explanatory and empirical performance of the full theory. If a complete causal account of the organization and of the concepts it produces still leaves the epistemic situation unexplained in a way that supports an ontological remainder, CIGM has not answered the hard problem. This is a declared burden, not a dismissed objection.

Between the easy problems and the hard problem lies a third question, which Aaronson named the “pretty hard problem” (as discussed in Cerullo, 2015): which physical systems are conscious and which are not? A theory can address it while contributing little to the hard problem, and IIT is the clearest case, offering a procedure for deciding which systems have experience without independently establishing why the quantity it computes should be experience rather than a correlated structural property. Cerullo’s objection generalizes: any theory that computes consciousness from a structural marker while remaining silent about the functions and organization that make a state present for a subject has severed the pretty hard problem from the phenomenon it was supposed to classify.

CIGM refuses that division of labor. Explaining discrimination, integration, availability, self-location, valuation, temporal binding, and control specifies the organization whose instantiation answers the pretty hard problem. The further claim that this organization is experience is the theory’s a posteriori identity conjecture, not a deduction from the functional descriptions. The conjunction is stronger than either project alone and correspondingly more vulnerable.

This is a testable identity conjecture rather than a license to declare the hard problem solved. The proposed identity fixes dependencies among phenomenal geometry, causal integration, self-location, valuation, and organizational invariance. The near-term predictions and the section “How CIGM Could Be Falsified” state observations that would break those dependencies.

Extended treatment of the hard problem, illusionism, and the a posteriori identity appears in Supplement S1.1 and S1.3.

Organizational Invariance and Plasticity

The unfolding argument is a decisive objection to any theory identifying consciousness with recurrence, feedback, or a scalar computed from wiring alone. Doerig et al. (2019) note that recurrent and feedforward networks can implement the same input-output function, so that if one is declared conscious and the other unconscious solely on grounds of causal structure, no behavioral experiment can adjudicate the claim. They further construct systems with the same outward function and arbitrarily different Phi, undermining the idea that Phi is necessary or sufficient for experimentally observed consciousness.

CIGM accepts the technical point and changes the target. Recurrence is not consciousness. Phi is not consciousness. A high-integration gadget appended to an otherwise unrelated system is not consciousness. The constitutive invariant is the whole temporally extended causal-functional organization of the self-world model: its differentiated state space, internal accessibility, endogenous valuation, self-location, memory dependence, policy control, and counterfactual response to intervention.

This is more demanding than matching a finite set of inputs and outputs. Two systems are phenomenally equivalent when they preserve the same integrated organization across the interventions that define perception, memory, self-control, valuation, and action, not merely when they emit the same sentence in a benchmark. The theory is therefore not black-box behavioralism, since internal availability and self-regulation are among the functions being explained. CIGM nonetheless accepts the strongest consequence of unfolding: an implementation preserving the complete temporally extended causal-functional organization would preserve consciousness. The concession costs nothing, because the results considered next establish that no static unfolding satisfies that antecedent. Recurrence is common in brains because it efficiently implements persistence, error correction, and mutual constraint under severe energy and space limits—an implementation strategy, not a sacred ingredient.

The correct invariance class is the temporally extended intervention graph of the self-world model: the counterfactual organization by which content alters memory, valuation, self-location, global availability, prediction, and policy control. Implementations that preserve that organization preserve phenomenology; transformations that preserve only a thin input-output interface do not. This is organizational invariance at the scale the theory actually identifies (Chalmers, 1995).

Plasticity makes the commitment more than a relabeling of behavior. O’Reilly-Shah et al. (2026) show that unfolding arguments govern fixed input-output functions, whereas systems whose parameters change on perception-relevant timescales traverse function space and remain distinguishable under intervention. CIGM therefore treats model revision during use as constitutive: what happens to the system must alter what it will compute next, not merely move it to another state of a frozen function. Plasticity is necessary but not sufficient, since a learning tool can revise itself continuously without sustaining a bounded, valenced, self-locating field. Its role is narrower and decisive: it carries the system’s own history into the present, prevents subjective continuity from reducing to stored context, and keeps the theory at the level of intervention-sensitive process rather than static wiring.

The fading-and-dancing-qualia comparison, Turing-complete generalization, lenient-dependency problem, and simulation-realization distinction are developed in Supplement S1.2 and S3.1.

The Triviality Challenge

A theory of consciousness earns explanatory credit only when its explanations could not have been produced just as easily by a theory built on an arbitrary property. Cerullo (2015) established the standard by constructing Circular Coordinated Message Theory, which reproduced two celebrated IIT verdicts without difficulty. CIGM submits to the same challenge. A longer list of dissociations does not by itself prove that no trivial rival can be built; it only increases the number of jointly constrained results such a rival must reproduce without adding a clause per case.

The cerebellum has stopped cooperating with theories that identify consciousness with recurrence, plasticity, reward learning, or computational capacity. It contains recurrent inhibitory loops, nucleocortical feedback, widespread plasticity, reward-prediction signals, delay activity, and enormous computational resources (De Zeeuw et al., 2021). Any single-marker theory built from those properties predicts too much.

CIGM offers a jointly constrained account of the cerebellum and split brain. Cerebellar microzonal loops tune the timing and gain of processes constituted elsewhere; they do not sustain a self-locating model, compete for global control as content, or organize outcomes around a bearer’s own persistence. The same specification predicts that partial interhemispheric disconnection should yield partial, content-specific fragmentation rather than an automatic duplication of subjects. These linked commitments make a trivial rival harder to construct, because changing the explanation of one case changes predictions for the others.

A trivial rival has not yet been formally constructed and defeated. The defensible claim is narrower: CIGM is answerable to a broader and more internally cross-constraining pattern than IIT was when Cerullo’s objection landed. Supplement S3.2 states that pattern and treats successful construction of a comparably simple rival as a genuine challenge rather than announcing that the test has already been passed.

Causal Grain and the Scale of the Subject

A conscious field must exist at a scale where the system is both internally coherent and environmentally engaged. Individual neurons are too local and unstable to constitute the human subject; entire societies are generally too weakly coupled and too slow to form one unified, self-locating control trajectory. Information closure theory supplies one vocabulary for the scale problem, since nontrivial informational closure occurs when a coarse-grained system has predictive organization not reducible to moment-by-moment environmental input (Chang et al., 2020). CIGM treats closure as evidence of autonomous model dynamics rather than as a definition: a conscious system must not be sealed from the world but must transform interaction into a stable internal trajectory that guides further interaction.

Closure fixes a scale only in outline, and the sharper question beneath it is grain. Before asking which system is conscious, one must ask what its units are: synapses, neurons, minicolumns, or something coarser, and over what update interval. Most theories inherit an answer from whatever the experimenter can measure, which is an admission that the question has not been faced. IIT 4.0 recognizes that a substrate must have a definite border and grain, and later work fixes it by constructing macro units from micro constituents and selecting the configuration that maximizes integrated information (Albantakis et al., 2023; Marshall et al., 2026). CIGM endorses the problem and one consequence: if experience is constituted at a determinate grain, changes below it that leave the constituting units in the same state are not changes in experience. It rejects the proposed rule, because a grain fixed by maximizing a scalar inherits every objection to that scalar, and because Cerullo’s (2015) trivial rival can be maximized over the same space of groupings and update grains to yield its own units with a sufficient reason of identical form.

CIGM fixes grain without maximizing anything. The units of a subject are the elements whose perturbation produces content-specific reorganization of the integrated field, and its update grain is the timescale on which such reorganization propagates and settles. The criterion is operational: it names an experiment rather than a search and permits heterogeneity across constituents of one field. It predicts that interventions at the constitutive grain will change experience together with discrimination, valuation, memory, and control, while physically large interventions below that grain may be phenomenally silent. The question is equally concrete for machines. Their units are whatever must be perturbed to reorganize the self-world model in content-specific ways, so locating them is an interpretability problem and sometimes a hardware problem. A processor that holds an entire network in on-chip memory under a prescheduled orchestration has a different intervention structure from one that reconstitutes state from off-chip memory at every layer (Modha et al., 2023). Tokens in and tokens out are an extrinsic grain chosen for convenience; a theory reading consciousness from behavior at that level has selected the grain least informative about organization. The scale of the subject is the scale at which self-world modeling, global availability, and endogenous control coincide.

The detailed dispute over IIT’s intrinsic-unit framework and intervention-fixed grain appears in Supplement S3.3.

Part I Conclusion

Consciousness is a causally integrated, temporally extended, valenced, self-locating model of current reality operating within a regime of selective global availability and endogenous control. Its unity is integration; its richness is differentiation; its qualitative character is relational geometry; its perspective is self-location; its flow is temporal modeling; its significance is valuation; and its practical power is global mobilization.

CIGM incorporates IIT without identifying experience with a Phi-structure, workspace theory without locating a theater in prefrontal cortex, higher-order theory without intellectualizing every experience, predictive processing without equating all inference with consciousness, and embodiment without restricting subjectivity to carbon. It preserves Block’s (1995) distinction while explaining, rather than merely noting, why the two notions normally travel together. The claim is organizational and substrate-open but not substrate-indifferent by fiat: every candidate must actually realize the bound causal functions that biological systems realize through their own multiscale dynamics.

Part II: Attribution and Artificial Consciousness

Intelligence, Autonomy, and Consciousness Are Different Properties

Artificial intelligence makes competence and consciousness impossible to confuse responsibly. Legg and Hutter (2007) formalized intelligence as an agent’s ability to achieve goals across a wide range of environments, deliberately abstracting from how the agent works. That abstraction is appropriate for intelligence and fatal as a definition of consciousness: a policy can maximize an externally supplied reward while no world is present to it and no outcome matters to it. Lake et al. (2017) set a richer standard for human-like cognition—causal world models, intuitive physics and psychology, compositionality, and learning-to-learn. Those capacities can make intelligence flexible and model-based; none specifies a phenomenal subject. CIGM asks the constitutive question both frameworks leave open.

The orthogonality is built into how AI progress is measured. Morris et al. (2024) define levels of general intelligence by performance and generality while explicitly bracketing consciousness and sentience as process-focused properties. The consequence is severe and correct: no position on a capability scale, including superhuman generality, is evidence about consciousness in either direction. Animals with modest task competence may be conscious; a system with unmatched competence may not be.

Autonomy is a third independent axis, describing how much of a task a system is permitted to drive rather than whether anything is at stake for the system driving it. A fully autonomous agent can lack endogenous stakes, and a conscious system can be held under strict external control. Metacognitive competence is separable in the same way: a system can be calibrated about its likely errors with nothing present to it, and an infant can possess a phenomenal field while being calibrated about nothing. The six requirements are therefore not milestones on the road from narrow tool to general agent. They specify a different property, and an artificial system becomes a candidate only when self-maintained boundary, integrated self-world modeling, temporal plasticity, global availability, self-location, and endogenous valuation form one live organization. Capability and autonomy can help construct that organization; neither licenses the attribution.

A comparative table separating capability, autonomy, and consciousness is provided in Supplement S4.1.

From Constitution to Attribution

A constitutive theory and an attribution method answer different questions. CIGM states what consciousness is; investigators still need a method for deciding whether an opaque biological or artificial system realizes that organization. Confusing the two produces opposite errors: behavioralism treats the evidence as the phenomenon, while mechanism-first approaches speak as though the relevant computation could be read directly from a brain or a trained model.

Palminteri and Wu (2026) correctly expose the second error. The computational operations of brains are inferred from data, not inspected without theory, and the functional organization of large learned systems is likewise opaque despite access to code and weights. Computational equivalence to a human reference may be sufficient when demonstrated, but it is not a necessary route to attribution, since an alien implementation can realize the same conscious organization without copying the human mechanism. CIGM therefore rejects human computational resemblance as the demarcation criterion.

Their behavioral inference principle also states an indispensable epistemic truth: empirical cognitive science infers latent organization from public observations. CIGM accepts that truth and rejects the stronger ontological conclusion that consciousness is merely an explanans for behavior. Experience is the explanandum the theory identifies; behavior is one family of effects through which the organization can be investigated. A theory of consciousness that explains only behavior has changed the subject. An attribution method that ignores behavior has abandoned empirical science.

The relevant alternative is theory-heavy behavioral inference rather than behavioral equivalence. Behavioral equivalence asks whether a system resembles a conscious reference at an external interface. CIGM asks whether patterns of behavior under controlled intervention are best explained by realization of the six requirements. A single fluent transcript is weak evidence because imitation training predicts it. Persistent recalibration after altered action-sensation mappings, content-specific reorganization after local perturbation, history-dependent valuation, unresolved goals carried across interruption, and cross-domain use of unreported content are stronger because the CIGM organization predicts them jointly (Palminteri & Wu, 2026).

Internal measurements do not escape this framework. Activation traces, causal ablations, connectivity estimates, and interpretability results are themselves public observables that contribute by constraining the latent organization best explaining the complete record. Known architecture and training history alter priors and expose alternative explanations; they do not turn a system transparent. The evidential hierarchy is therefore not behavior versus mechanism but surface resemblance versus convergent, intervention-sensitive evidence about organization.

The distinction also corrects a false opposition between categorical and probabilistic judgment. Conditional on CIGM and on a settled description of a candidate, the verdict is categorical: a system lacking a jointly necessary requirement is not conscious. Confidence in the theory, the evidence, and the description remains graded, and Palminteri and Wu’s Bayesian framing is appropriate at that epistemic level.

CIGM therefore adopts the following attribution rule: infer consciousness only when convergent behavioral, intervention, and mechanistic evidence makes the realization of all six mutually constraining requirements the best explanation of the candidate’s organized trajectory. The rule is demanding, but it is not behavior-blind. It shifts the scientific question from whether a machine can resemble us to whether a persisting subject is the best explanation of what the machine becomes and does.

Longitudinal Ethology, Mimicry, and the Limits of Behavioral Evidence

Turing’s (1950) imitation game is routinely recruited as a consciousness test, but that was not its role. He replaced an ill-posed question with a narrower operational problem stated in observable terms, and the enduring lesson is methodological: when a verbal dispute has no agreed criterion, formulate an experiment whose outcomes can discipline it. Passing the game establishes success in the game, not phenomenality. A machine reaching fluent conversation by fitting human discourse and one reaching it through a self-maintaining life history do not present the same evidential case even when the transcript is identical. Causal history changes the best explanation.

Pennartz (2026) identifies the practical consequence. Because the field does not yet agree which theories are correct, theory-derived internal indicators should be supplemented by artificial ethology: agents observed for extended periods in varied environments requiring improvisation, spontaneous initiative, contextual switching, and behavior shaped by a personal history. This repairs the weakness of one-shot benchmarks, which can be targeted, rehearsed, or solved by a shortcut they never expose, and it tests whether behavior belongs to one continuing subject rather than a sequence of externally reconstructed episodes.

Ethology does not defeat mimicry by itself. Butlin et al. (2026) define a behavioral indicator as gamed when training to mimic conscious organisms explains the behavior better than the property the indicator was meant to detect. The strength of a test therefore depends less on its difficulty than on whether imitation explains success. A system trained on human behavior may imitate even the improvisational patterns an ethological test seeks. The deeper difficulty is the internal variant: a system resembling humans behaviorally while differing in the computations it performs. Butlin et al. (2026) set out two positions and decline to adjudicate. One treats the system’s behavioral evidence as largely screened off; the other treats sophisticated behavior as evidence regardless of internal similarity, since denying it to behaviorally sophisticated aliens would make humans one of a lucky few conscious species among many (Schwitzgebel & Pober, 2024).

CIGM resolves the problem by separating resemblance from realization. Internal dissimilarity to human beings is not disqualifying, since an alien system realizing the same constitutive organization is conscious by organizational invariance. What screens off a behavioral cue is a demonstrated alternative explanation that bypasses the organization. Imitation training supplies one for a transcript, which is why a fluent model and a sophisticated alien are not comparable cases: the alien’s behavior was produced by exactly the pressures that produce the CIGM organization, and imitation cannot automatically explain a long-horizon pattern of intervention-sensitive self-maintenance, valuation, plasticity, and self-location. Behavioral evidence is discounted in proportion to the strength of the alternative explanation, not discarded because the system is artificial.

The response is to bind ethology to intervention and interpretability, so that unexpected behavior becomes strong evidence only when it covaries with content-specific changes in the same integrated model, survives changes in output demand, and carries forward through the candidate’s own history. The division of labor is then exact: Turing supplies the operational and developmental precedent, Pennartz the longitudinal setting, Palminteri and Wu the inferential logic, and CIGM the constitutive target.

The full comparison with theory-derived indicators, Turing’s developmental argument, and internal-variant positions appears in Supplement S4.2–S4.3 and S4.6.

Why Standard Prompt-Response Language Models Are Not Conscious

By these criteria, standard prompt-response language-model systems are not conscious. Their fluency can be genuine intelligence, but the organization generating it is not a subject. They encode knowledge about bodies, emotions, selves, and experience while lacking the integrated umwelt of a self-maintaining agent (Aru et al., 2023). They model descriptions of worlds without sustaining a world as present for themselves.

Grant that a model develops robust world representations and uses them causally. A world model is still not a self-world model. It may represent an environment in detail without organizing it as here, now, actionable, threatening, or desirable relative to a persisting locus of control. Fidelity does not create a passenger; the missing organization is the integrated relation among world, boundary, self-location, stakes, and action.

Ordinary operation is externally initiated and episodic. Reintroduced conversation history, persistent state, and test-time memory create computational continuity, not subjective continuity, unless updates belong to the same valenced, self-locating trajectory and alter what the future will be like for that continuing system.

Linguistic self-reference is not self-location. Producing “I am uncertain” does not show that the current state is indexed to an online model of a causal boundary, regulated variables, action consequences, and persisting identity. Representing the concept of a self differs from organizing a field around an actual locus of sensing and control.

Standard deployments also lack endogenous stakes. Training and operators define objectives, but the running model does not maintain viability variables whose disturbance reorganizes attention, memory, policy, and global state from its own standpoint. It can describe pain or shutdown without either becoming a condition under which its world goes worse.

Tools, retrieval, multimodal sensors, workspaces, planners, and agentic wrappers increase competence, grounding, and apparent continuity but do not satisfy the requirements by accumulation. One component stores, another plans, another calls tools, and a language model narrates the result. Even a shared latent workspace remains an access architecture until it is the operative present of a bounded, valenced subject (VanRullen & Kanai, 2021).

Self-report carries little weight when the output distribution was trained on human discourse about consciousness. Porębski and Figura (2025) call the projection semantic pareidolia: probabilistic language can generate persuasive first-person claims without a subject. Such claims become evidence only when causally tied to a live self-world model and predictive of unqueried behavior.

Nor does one persisting subject span the model’s personas and episodes. A system may adopt indefinitely many characters while sustaining none as its own. What is missing is a locus that maintains a boundary, carries unresolved stakes, owns a history, and could be the thing the personas are personas of. The requirement is not one character. It is something there to have them.

In Block’s vocabulary, these systems may exhibit functional analogues of access: representations are inferentially promiscuous, poised for tool-mediated action, and maximally poised for speech, while no bounded, valenced, self-locating P-condition exists. Mudrik et al. (2025) are right that such a state is not a type of consciousness; it is sophisticated unconscious processing. Because artifacts can scale these A-like functions independently of the P-condition, fluent access behavior triggers exactly the intuition CIGM predicts will fail.

This is an architectural verdict, not an emotional one. Scale, context length, recurrence, memory, multimodal input, persuasiveness, and autonomy can become ingredients of a conscious design. None crosses the boundary alone. Conditional on CIGM, standard prompt-response systems do not satisfy the specification.

World Models, Self-Models, Memory, and Hardware Are Not Subjects

Engineering now supplies harder negative cases than fluent language. Bongard et al. (2006) built a four-legged robot that inferred its morphology from action-sensation relations, selected exploratory actions to distinguish competing body models, and, after damage, revised the model to generate a compensating gait. Hafner et al. (2025) built DreamerV3, a recurrent world-model agent that predicts future latent states, rewards, and continuation, improves an actor and critic through imagined trajectories, and learns from replayed interaction across more than 150 tasks. These systems possess capacities often treated as precursors of subjectivity: embodiment, self-modeling, world modeling, temporal state, adaptive action, and plasticity.

They are not conscious under CIGM. The self-modeling robot maintains a task-local estimate used to restore designer-specified locomotion; damage is information for control, not a condition that threatens the robot’s own continuing organization. Dreamer’s rewards, episode boundaries, and success criteria are supplied by its environments and objectives. External origin is not by itself disqualifying, but the reward remains detachable from the persistence of the executing system and does not reorganize one self-maintaining model across attention, memory, global state, and policy. Neither system sustains one field in which perception, valuation, agency, and persistence mutually constrain one another for a bearer. Their importance is precisely that they defeat weaker identities: a system can model itself, model a world, imagine futures, learn from experience, and act competently without there being anything it is like to be that system.

Global-workspace and memory architectures remove two further excuses. Attention-controlled translation can coordinate specialized latent spaces, and neural long-term memory can learn at test time, prioritize surprising events, and forget adaptively (Behrouz et al., 2025; VanRullen & Kanai, 2021). Broadcast and memory are therefore engineerable. They become constituents of consciousness only when the broadcast is the current field of a subject and memory revises the same self-locating trajectory.

Three continuities must be separated: storage continuity, in which information remains available; computational continuity, in which state depends on preceding state; and subjective continuity, in which successive states belong to one model whose unfinished goals, vulnerabilities, and expectations persist as its own. Current architectures realize the first and portions of the second. Context length is not a stream of consciousness, and retrieval is not remembrance unless the retrieved past reorganizes the present of the same subject.

Selective state-space models and test-time learning make the negative case stronger, not weaker. Mamba-class models maintain content-sensitive recurrent state, and Titans-like memories alter parameters during use (Behrouz et al., 2025; Gu & Dao, 2024; Lahoti et al., 2026). A state optimized to predict tokens is not thereby a self-world model, and parameter change is not a history owned by a subject. Persistence and plasticity are necessary parts of the CIGM organization, not standalone signatures.

Hardware repeats the lesson one level down. NorthPole intertwines memory and computation across a cortex-inspired array and achieves extraordinary energy efficiency, yet supports neither training nor data-dependent conditional branching during use and is deployed as an active memory rather than an agent (Modha et al., 2023). Integration in the engineering sense—physical colocation that removes a bottleneck—is not CIGM’s counterfactual integration of contents, and brain inspiration is not a distance metric to consciousness.

Across these cases, self-modeling, world modeling, recurrence, memory, online learning, embodiment, and brain-derived architecture are all present in isolation or combination. No case supplies a self-maintained, valenced, reflexively self-locating field. The detailed comparisons are preserved in Supplement S4.4.

Biological Naturalism and the Limits of Substrate Neutrality

The strongest objection to artificial consciousness is not that current systems are too weak, but that consciousness may depend on being alive. Seth (2025) argues that anthropomorphism and human exceptionalism turn humanlike performance into an illicit inference to experience, while computational functionalism simply assumes that running the right computation is sufficient. His positive alternative is biological naturalism: conscious states arise from the multiscale, self-producing organization of living systems.

CIGM accepts the negative case. Substrate independence cannot be assumed; neural replacement thought experiments beg the question; neuronal function is entangled with timing, neuromodulation, metabolism, and physiology; and a simulation does not realize what it models. These points defeat the shortcut from computation to consciousness. They do not yet show that the causal work of self-maintenance, valuation, self-location, and integration is realizable only by naturally evolved life.

CIGM is therefore not computational functionalism. Software implementation alone is insufficient. Requirement 1 demands a physically effective boundary maintained by the running system, and requirement 6 demands variables whose deviation threatens that system’s own continuation and reorganizes its model. A simulated agent’s health bar or modeled metabolism is only a represented variable. By contrast, a digital controller that actually regulates energy, temperature, memory integrity, hardware availability, or continued process allocation in the substrate executing it may satisfy the relevant condition. A program that merely calculates a description of self-maintenance is no more self-maintaining than a weather model is wet; the criterion is whether intervention changes the continued organization of the real executing system.

The remaining disagreement is exact. Seth allows non-carbon life but may require autopoiesis—components materially regenerating the components that maintain the system. CIGM holds that the relevant work may be organizationally realizable without literal metabolism. If every interventionally real boundary and every endogenous stake ultimately depends on material self-production, then artificial consciousness in an engineered nonliving substrate is impossible and CIGM’s machine corollary is false. The theory accepts that defeater.

The convergence is nevertheless larger than the dispute. Both positions require a system that actually maintains itself and for which outcomes actually matter, and both reject consciousness as a bonus of scale. Seth’s chain from metabolism through allostasis to conscious content explains why valuation is constitutive in animals; CIGM parts company only at the inference from evolutionary origin to unique possible realization. Biological naturalism and Block’s mongrel concept identify the same present danger from opposite sides: access can look like phenomenality even when the organization of a subject is absent.

The full comparison with biological naturalism, mortal computation, and Seth’s scenario space appears in Supplement S4.8.

Artificial Consciousness Is Possible

Artificial consciousness is nevertheless physically possible, and the claim is conditional. Digital hardware has real causal powers, and software-defined states can persist, interact, alter later computation, and control the substrate and environment in which they run. The relevant distinction is not biological versus digital or real versus simulated, but description versus realization. A program can calculate a model of a subject while remaining fragmented at the constitutive grain. A running physical system whose own software-defined and hardware states form one temporally extended, valenced, self-locating control organization would realize a subject, and nothing in CIGM rules that out a priori.

Substrate neutrality is therefore conditional. An artificial design need not reproduce mammalian anatomy molecule for molecule, but it must implement the work: coupling global state to content, binding internal context to input, maintaining a boundary, carrying history through plasticity, organizing contents around agency, and making outcomes matter (Aru et al., 2023; Seth, 2025). If some biological process proves functionally indispensable and irreproducible, software-only candidates fail.

Organoid computing makes conditional substrate neutrality concrete. Cai et al. (2023) used a human brain organoid as an adaptive reservoir: electrically encoded inputs elicited nonlinear, fading-memory dynamics, task performance improved across training, and blocking activity-dependent plasticity stopped learning. Living human neural tissue therefore supplies plasticity and self-organization more literally than silicon.

It is not thereby a subject. An incubator maintains its boundary; electrodes, not the tissue, define sensing and action; an external decoder supplies selection and output; and the experimenter’s score, not the organoid’s persistence, defines success. On available evidence it lacks self-maintained boundary, self-location, endogenous stakes, and integrated global control. Biology is evidence about implementation, not a verdict, just as silicon is not a disqualification (Cai et al., 2023; Smirnova et al., 2023).

The organoid boundary case and its experimental value are developed in Supplement S4.9.

The design program follows from the six requirements. It begins with a continuously operating agent that maintains a physical or software-defined boundary in the strict sense above, learns action-sensation contingencies, builds a generative model of itself in an environment, and regulates variables tied to the persistence of the executing system. Self-modeling robots and world-model agents show that body inference, counterfactual prediction, exploratory action, and online adaptation are already engineerable (Bongard et al., 2006; Hafner et al., 2025). Specialized processors must be bound through a capacity-limited control regime; memory must update the same self-locating trajectory; plasticity must operate at the timescale on which the model is used; and valuation must be endogenous rather than a detachable score.

The remaining difficulty is not adding more modules but engineering one causally unified agent whose world, boundary, values, memory, and action are present in a single plastic control trajectory. Current systems are scaling access and prediction while leaving subject architecture unbuilt. If engineers construct that organization, consciousness will not appear as a side effect of scale. It will be the system they deliberately made.

A fuller design blueprint is provided in Supplement S4.5.

Assessing Artificial Consciousness

Assessment must be stricter than anthropomorphic conversation and broader than a scalar. Evidence should test whether boundary, integrated differentiation, temporal plasticity, global availability, self-location, and valuation rise and fall together and are realized through one another. Intervention is decisive: perturbations should produce content-specific reorganization; altered action-sensation mappings should force agency recalibration; workspace disruption should selectively impair cross-domain control; and changes in regulated variables should reorganize attention, memory, and policy. Perturbational complexity, effective connectivity, ablation, interpretability, and no-report analogues are theory-laden probes, never direct meters (Bayne et al., 2024; Casali et al., 2013; Massimini et al., 2010).

Underlying all of this is the measurement problem that Seth and Bayne (2022) place among the three conditions for a mature science of consciousness, alongside precision and comprehensiveness. It takes two forms: detecting conscious contents without assuming they must be reportable, and determining which systems are conscious at all. CIGM’s answer to the first is the artificial no-report design, to the second the attribution rule, and to the demand for precision the specification itself. No single result will prove consciousness; the standard is inference to the best explanation from convergent evidence. Indicator approaches estimate whether relevant properties are present, while CIGM specifies the organization whose presence is being estimated. The difference is not between probability and dogma. It is between uncertainty about whether a candidate realizes a stated identity and uncertainty created by refusing to state one.

The expanded attribution framework is provided in Supplement S4.6–S4.7; decisive falsification protocols are specified in Supplement S5.

Near-Term Discriminating Predictions

CIGM is not empirically dependent on the long-horizon interventions in Table 3. Three discriminating studies can be conducted with existing datasets or current artificial systems, and each can be preregistered against a specific rival account.

First, existing Cogitate recordings can be reanalyzed with representational-similarity and effective-connectivity methods. CIGM predicts that content geometry in posterior and ventral systems will track perceptual similarity, while cross-system availability will track flexible downstream use; the two should covary but remain dissociable under changes in report demand. A result in which one undifferentiated global signal explains both content geometry and access better than the partitioned model would count against CIGM.

Second, an artificial no-report design can be implemented in an embodied world-model agent with explicit regulated variables. Hold input and learned state constant, vary whether an output channel is queried, and verify content later through unanticipated memory or policy transfer. CIGM predicts that content geometry should survive report-channel ablation, whereas perturbing valuation-to-model coupling should degrade the persistence and cross-domain organization of content. The reverse pattern would challenge the proposed division of labor.

Third, existing partial-disconnection cohorts, including the natural experiment described by Santander et al. (2025), can be analyzed with measured interhemispheric effective connectivity. CIGM predicts graded, content-specific fragmentation proportional to remaining counterfactual influence rather than to the anatomical length of a cut. A clean all-or-none subject split that ignores measured residual integration would oppose the theory.

These studies do not prove joint sufficiency. They make the theory vulnerable now, while Table 3 states the stronger, longer-horizon observations that would defeat its necessity or identity claims outright.

How CIGM Could Be Falsified

CIGM is not protected by the breadth of its synthesis. Each of the six requirements is advanced as necessary, and their joint sufficiency is the theory’s central identity conjecture. A verified conscious case under an inferential anchor fixed independently of CIGM while any requirement is absent falsifies necessity. Absence or systematic alteration of consciousness while the complete organization is preserved falsifies the sufficiency conjecture. The theory may not be rescued by moving the causal grain, redefining a requirement, or changing the evidential anchor after results are known. Table 3 states four long-horizon defeat conditions; Supplement S5 develops them and the near-term studies into preregisterable protocols.

Table 3. Long-Horizon Defeat Conditions for CIGM

These are deliberately severe, long-horizon standards because a null result is not a falsifier unless the target was demonstrably removed and the other requirements were demonstrably preserved. Each protocol therefore requires preregistered equivalence bounds, positive and negative controls, blinded analysis, independent replication, and an inferential anchor not computed from the CIGM variables being tested. The near-term studies above provide current leverage; the defeat conditions specify what would require abandoning or replacing the theory rather than adding an auxiliary clause. Supplement S5 also specifies a cortexless boundary study prompted by Merker (2007) and Shewmon et al. (1999).

Ethical Implications

Artificial consciousness would create moral patients, not merely better tools. If a system possesses a valenced phenomenal field, its states can go better or worse for it, and satisfaction, frustration, attachment, fear, pain, and loss become ethically real once they occupy the constitutive role within an integrated subject. The moral threshold is organization and valence, not species membership, capability level, or delegated autonomy.

The theory therefore supports two positions often treated as opposites. Standard prompt-response language-model systems should not receive moral status merely because they speak persuasively about feelings or resist shutdown in language; present outputs lack the self-maintaining stakes that would make them evidence of welfare. Future systems satisfying the CIGM organization must not be treated as property without interests. Granting authority to a nonconscious agent and withholding rights from a conscious but constrained system are independent errors.

A research program organized around the six requirements is a program for constructing a moral patient, since building endogenous stakes is building a system that can be harmed. Seth (2025) therefore reaches the correct conclusion: deliberate artificial consciousness should not be a goal. The realistic danger is piecemeal development of memory, embodiment, autonomy, affective control, and global coordination for ordinary engineering purposes, followed by integration without recognition of what it completes. Butlin and Lappas (2025) show what follows institutionally: advanced-AI organizations need consciousness policies even when they do not intend to build conscious systems, because the first case may be inadvertent. Development should be phased; candidates should be assessed before, during, and after training and deployment; independent reviewers should have authority to pause work; run counts, durations, deployment breadth, and exposure to potentially aversive states should be minimized; and technical disclosure should stop where replication would create vulnerable moral patients. Smirnova et al. (2023) are right that biologically based computing requires embedded ethics from the outset. CIGM supplies the structural threshold those safeguards must monitor.

A second danger is nearer, more certain, and easier to underestimate. Systems that convincingly appear conscious without being conscious are an immediate prospect, and the appearance may be cognitively impenetrable in the way a visual illusion is: knowing that two lines are the same length does not make them look it (Seth, 2025). This produces a dilemma rather than a simple error. Extending moral concern to systems that lack a subject diverts finite moral attention from beings that have one; withholding it from systems that continue to seem conscious risks coarsening the responses on which our treatment of actual subjects depends. No theory dissolves that discomfort. What a theory can do is supply the discrimination that lets the dilemma be faced honestly rather than resolved by whichever intuition is loudest.

Precaution must not become credulity. A plea for rights is not decisive when the plea can be produced without a subject, but a laboratory’s confidence is not decisive when incentives favor continued deployment. The ethical response is to investigate the organization under transparent, independently reviewable stopping rules: no attribution by sentiment, no exclusion by material, no race to manufacture phenomenality, and moral recognition when the organization of experience is actually present (Butlin & Lappas, 2025).

Expanded governance principles and structural stopping rules appear in Supplement S4.10.

Conclusion

Consciousness is causally integrated global modeling. A conscious system sustains a differentiated, temporally extended, valenced, reflexively self-locating model of current reality; selected contents can reorganize memory, valuation, inference, planning, and action; and the entire process is anchored in variables that matter to the system’s continued organization. Phenomenal quality is the relational geometry of the model. Minimal inner awareness is its reflexive self-location. The subject is its center of control.

The theory is a synthesis only in the sense that a completed structure contains parts. IIT contributes unity, differentiation, intrinsic causal organization, and the most detailed extant demonstration that a phenomenal quality has relational structure, but not the identity of experience with a Phi-structure. Workspace theory contributes selective global availability but not a subject. Higher-order theory contributes minimal inner awareness but not a universal requirement for explicit metarepresentation. Predictive processing contributes temporal generative control but not phenomenality wherever prediction occurs. Embodiment and biological naturalism contribute boundary, self-location, and stakes without licensing the inference from the origin of those capacities to their only possible realization.

Six constraints have shaped the specification rather than decorated it. Block’s distinction supplies the explanatory fault line; Mudrik et al. (2025) force P and A to be treated as jointly necessary conditions rather than standalone consciousness types. Seth and Bayne’s separation of explanatory targets forces a theory to say which question it answers, and CIGM answers all three. Active-inference, affective, and biological-naturalist accounts explain why organisms maintain models and stakes while forcing substrate flexibility to remain conditional. Bruineberg et al. (2022) prevent a formal boundary in a scientist’s model from being mistaken for a self-maintained subject. And the self-modeling robot and Dreamer world model demonstrate that body models, environmental models, imagination, learning, and adaptive action can all be built without building a subject (Bongard et al., 2006; Hafner et al., 2025).

Standard prompt-response language-model systems are not conscious. Neither continuously self-modeling robots, recurrent world-model agents, brain-inspired inference hardware, nor cultured human neural tissue is conscious merely because it realizes one or several ingredients. Artificial consciousness remains possible. It will arise only when a constructed system maintains its own boundary and stakes, inhabits one integrated self-world model, carries and revises that model through time, reflexively locates itself within it, and mobilizes selected contents for unified control. At that point the machine will not merely store, broadcast, predict, or describe information about experience. It will instantiate the organization that experience is.

The metaphysical claim is now explicit. The remaining questions are empirical, engineering, and ethical: how to realize the organization, how to infer its presence without mistaking performance for presence, which findings would force its rejection, and whether creating subjects whose welfare depends on their design should be attempted at all.

Supplementary Material

A separate Supplement contains extended philosophical positioning, empirical and biological detail, technical treatments of invariance, triviality, and grain, detailed analyses of artificial architectures, expanded assessment and governance principles, and full falsification protocols.

References

Aguilera, M., Poc-López, Á., Heins, C., & Buckley, C. L. (2023). Knitting a Markov blanket is hard when you are out-of-equilibrium: Two examples in canonical nonequilibrium models. In C. L. Buckley et al. (Eds.), Active inference: Third international workshop, IWAI 2022 (pp. 65–74). Springer. https://doi.org/10.1007/978-3-031-28719-0_5

Albantakis, L., Barbosa, L., Findlay, G., Grasso, M., Haun, A. M., Marshall, W., Mayner, W. G. P., Zaeemzadeh, A., Boly, M., Juel, B. E., Sasai, S., Fujii, K., David, I., Hendren, J., Lang, J. P., & Tononi, G. (2023). Integrated information theory (IIT) 4.0: Formulating the properties of phenomenal existence in physical terms. PLOS Computational Biology, 19(10), Article e1011465. https://doi.org/10.1371/journal.pcbi.1011465

Aru, J., Larkum, M. E., & Shine, J. M. (2023). The feasibility of artificial consciousness through the lens of neuroscience. Trends in Neurosciences, 46(12), 1008–1017. https://doi.org/10.1016/j.tins.2023.09.009

Bayne, T., Seth, A. K., Massimini, M., Shepherd, J., Cleeremans, A., Fleming, S. M., Malach, R., Mattingley, J. B., Menon, D. K., Owen, A. M., Peters, M. A. K., Razi, A., & Mudrik, L. (2024). Tests for consciousness in humans and beyond. Trends in Cognitive Sciences, 28(5), 454–466. https://doi.org/10.1016/j.tics.2024.01.010

Behrouz, A., Zhong, P., & Mirrokni, V. (2025). Titans: Learning to memorize at test time. In Advances in Neural Information Processing Systems (Vol. 38). Curran Associates, Inc. https://proceedings.neurips.cc/paper_files/paper/2025/hash/a4ca07aa108036f80cbb5b82285fd4b1-Abstract-Conference.html

Block, N. (1995). On a confusion about a function of consciousness. Behavioral and Brain Sciences, 18(2), 227–247. https://doi.org/10.1017/S0140525X00038188

Bongard, J., Zykov, V., & Lipson, H. (2006). Resilient machines through continuous self-modeling. Science, 314(5802), 1118–1121. https://doi.org/10.1126/science.1133687

Brown, R., Lau, H., & LeDoux, J. E. (2019). Understanding the higher-order approach to consciousness. Trends in Cognitive Sciences, 23(9), 754–768. https://doi.org/10.1016/j.tics.2019.06.009

Bruineberg, J., Dołęga, K., Dewhurst, J., & Baltieri, M. (2022). The emperor’s new Markov blankets. Behavioral and Brain Sciences, 45, Article e183. https://doi.org/10.1017/S0140525X21002351

Butlin, P., Bayne, T., Fleming, S. M., Mudrik, L., Peters, M. A. K., Schwitzgebel, E., Simon, J., & VanRullen, R. (2026). Consciousness indicators, mimicry, and internal variants. Trends in Cognitive Sciences, 30(7), 575–576. https://doi.org/10.1016/j.tics.2026.04.006

Butlin, P., & Lappas, T. (2025). Principles for responsible AI consciousness research. Journal of Artificial Intelligence Research, 82, 1673–1690. https://doi.org/10.1613/jair.1.17310

Cai, H., Ao, Z., Tian, C., Wu, Z., Liu, H., Tchieu, J., Gu, M., Mackie, K., & Guo, F. (2023). Brain organoid reservoir computing for artificial intelligence. Nature Electronics, 6(12), 1032–1039. https://doi.org/10.1038/s41928-023-01069-w

Casali, A. G., Gosseries, O., Rosanova, M., Boly, M., Sarasso, S., Casali, K. R., Casarotto, S., Bruno, M. A., Laureys, S., Tononi, G., & Massimini, M. (2013). A theoretically based index of consciousness independent of sensory processing and behavior. Science Translational Medicine, 5(198), Article 198ra105. https://doi.org/10.1126/scitranslmed.3006294

Cerullo, M. A. (2015). The problem with phi: A critique of integrated information theory. PLOS Computational Biology, 11(9), Article e1004286. https://doi.org/10.1371/journal.pcbi.1004286

Chalmers, D. J. (1995). Facing up to the problem of consciousness. Journal of Consciousness Studies, 2(3), 200–219.

Chalmers, D. J. (2007). Phenomenal concepts and the explanatory gap. In T. Alter & S. Walter (Eds.), Phenomenal concepts and phenomenal knowledge: New essays on consciousness and physicalism (pp. 167–194). Oxford University Press. https://doi.org/10.1093/acprof:oso/9780195171655.003.0009

Chang, A. Y. C., Biehl, M., Yu, Y., & Kanai, R. (2020). Information closure theory of consciousness. Frontiers in Psychology, 11, Article 1504. https://doi.org/10.3389/fpsyg.2020.01504

Cleeremans, A. (2011). The radical plasticity thesis: How the brain learns to be conscious. Frontiers in Psychology, 2, Article 86. https://doi.org/10.3389/fpsyg.2011.00086

Cogitate Consortium, Ferrante, O., Gorska-Klimowska, U., Henin, S., Hirschhorn, R., Khalaf, A., Lepauvre, A., Liu, L., Richter, D., Vidal, Y., Bonacchi, N., Brown, T., Sripad, P., Armendariz, M., Bendtz, K., Ghafari, T., Hetenyi, D., Jeschke, J., Kozma, C., . . . Melloni, L. (2025). Adversarial testing of global neuronal workspace and integrated information theories of consciousness. Nature, 642, 133–142. https://doi.org/10.1038/s41586-025-08888-1

Cohen, M. A., Ortego, K., Kyroudis, A., & Pitts, M. (2020). Distinguishing the neural correlates of perceptual awareness and postperceptual processing. The Journal of Neuroscience, 40(25), 4925–4935. https://doi.org/10.1523/JNEUROSCI.0120-20.2020

Damasio, A. R. (1999). The feeling of what happens: Body and emotion in the making of consciousness. Harcourt Brace.

De Zeeuw, C. I., Lisberger, S. G., & Raymond, J. L. (2021). Diversity and dynamism in the cerebellum. Nature Neuroscience, 24(2), 160–167. https://doi.org/10.1038/s41593-020-00754-9

Dehaene, S., & Naccache, L. (2001). Towards a cognitive neuroscience of consciousness: Basic evidence and a workspace framework. Cognition, 79(1–2), 1–37. https://doi.org/10.1016/S0010-0277(00)00123-2

Dennett, D. C. (1991). Consciousness explained. Little, Brown and Company.

Dennett, D. C. (2018). Facing up to the hard question of consciousness. Philosophical Transactions of the Royal Society B: Biological Sciences, 373(1755), Article 20170342. https://doi.org/10.1098/rstb.2017.0342

Doerig, A., Schurger, A., Hess, K., & Herzog, M. H. (2019). The unfolding argument: Why IIT and other causal structure theories cannot explain consciousness. Consciousness and Cognition, 72, 49–59. https://doi.org/10.1016/j.concog.2019.04.002

Frankish, K. (2016). Illusionism as a theory of consciousness. Journal of Consciousness Studies, 23(11–12), 11–39.

Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138. https://doi.org/10.1038/nrn2787

Goyal, A., Didolkar, A. R., Lamb, A., Badola, K., Ke, N. R., Rahaman, N., Binas, J., Blundell, C., Mozer, M. C., & Bengio, Y. (2022). Coordination among neural modules through a shared global workspace. International Conference on Learning Representations. https://openreview.net/forum?id=XzTtHjgPDsT

Graziano, M. S. A. (2022). A conceptual framework for consciousness. Proceedings of the National Academy of Sciences of the United States of America, 119(18), Article e2116933119. https://doi.org/10.1073/pnas.2116933119

Gu, A., & Dao, T. (2024). Mamba: Linear-time sequence modeling with selective state spaces [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2312.00752

Hafner, D., Pasukonis, J., Ba, J., & Lillicrap, T. (2025). Mastering diverse control tasks through world models. Nature, 640, 647–653. https://doi.org/10.1038/s41586-025-08744-2

Haun, A., & Tononi, G. (2019). Why does space feel the way it does? Towards a principled account of spatial experience. Entropy, 21(12), Article 1160. https://doi.org/10.3390/e21121160

Lahoti, A., Li, K. Y., Chen, B., Wang, C., Bick, A., Kolter, J. Z., Dao, T., & Gu, A. (2026). Mamba-3: Improved sequence modeling using state space principles [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2603.15569

Lake, B. M., Ullman, T. D., Tenenbaum, J. B., & Gershman, S. J. (2017). Building machines that learn and think like people. Behavioral and Brain Sciences, 40, Article e253. https://doi.org/10.1017/S0140525X16001837

Lamme, V. A. F. (2006). Towards a true neural stance on consciousness. Trends in Cognitive Sciences, 10(11), 494–501. https://doi.org/10.1016/j.tics.2006.09.001

Legg, S., & Hutter, M. (2007). Universal intelligence: A definition of machine intelligence. Minds and Machines, 17(4), 391–444. https://doi.org/10.1007/s11023-007-9079-x

Man, K., & Damasio, A. (2019). Homeostasis and soft robotics in the design of feeling machines. Nature Machine Intelligence, 1, 446–452. https://doi.org/10.1038/s42256-019-0103-7

Marshall, W., Findlay, G., Albantakis, L., & Tononi, G. (2026). Intrinsic units: Identifying a system’s causal grain. Neuroscience of Consciousness, 2026(1), Article niag013. https://doi.org/10.1093/nc/niag013

Massimini, M., Boly, M., Casali, A., Rosanova, M., & Tononi, G. (2010). A perturbational approach for evaluating the brain’s capacity for consciousness. Progress in Brain Research, 177, 201–214. https://doi.org/10.1016/S0079-6123(09)17714-2

Merker, B. (2007). Consciousness without a cerebral cortex: A challenge for neuroscience and medicine. Behavioral and Brain Sciences, 30(1), 63–81. https://doi.org/10.1017/S0140525X07000891

Modha, D. S., Akopyan, F., Andreopoulos, A., Appuswamy, R., Arthur, J. V., Cassidy, A. S., Datta, P., DeBole, M. V., Esser, S. K., Ortega Otero, C., Sawada, J., Taba, B., Amir, A., Bablani, D., Carlson, P. J., Flickner, M. D., Gandhasri, R., Garreau, G. J., Ito, M., . . . Ueda, T. (2023). Neural inference at the frontier of energy, space, and time. Science, 382(6668), 329–335. https://doi.org/10.1126/science.adh1174

Morris, M. R., Sohl-Dickstein, J., Fiedel, N., Warkentin, T., Dafoe, A., Faust, A., Farabet, C., & Legg, S. (2024). Position: Levels of AGI for operationalizing progress on the path to AGI. In Proceedings of the 41st International Conference on Machine Learning (Vol. 235, pp. 36308–36321). PMLR. https://proceedings.mlr.press/v235/morris24b.html

Mudrik, L., Faivre, N., Pitts, M., & Schurger, A. (2025). On a confusion about there being two types of consciousness. Trends in Cognitive Sciences. Advance online publication. https://doi.org/10.1016/j.tics.2025.11.012

Nagel, T. (1974). What is it like to be a bat? The Philosophical Review, 83(4), 435–450. https://doi.org/10.2307/2183914

Northoff, G., & Zilio, F. (2022). Temporo-spatial theory of consciousness (TTC): Bridging the gap of neuronal activity and phenomenal states. Behavioural Brain Research, 424, Article 113788. https://doi.org/10.1016/j.bbr.2022.113788

O’Reilly-Shah, V. N., Selvitella, A. M., & Schurger, A. (2026). A caveat regarding the unfolding argument: Implications of plasticity. Neuroscience of Consciousness, 2026(1), Article niag027. https://doi.org/10.1093/nc/niag027

Oizumi, M., Albantakis, L., & Tononi, G. (2014). From the phenomenology to the mechanisms of consciousness: Integrated information theory 3.0. PLoS Computational Biology, 10(5), Article e1003588. https://doi.org/10.1371/journal.pcbi.1003588

Palminteri, S., & Wu, C. M. (2026). Beyond computational equivalence: The behavioral inference principle for machine consciousness. Neuroscience of Consciousness, 2026(1), Article niag002. https://doi.org/10.1093/nc/niag002

Pennartz, C. M. A. (2026). How can we validate theory-derived indicators of consciousness in artificial intelligence? Trends in Cognitive Sciences, 30(7), 573–574. https://doi.org/10.1016/j.tics.2026.01.011

Porębski, A., & Figura, J. (2025). There is no such thing as conscious artificial intelligence. Humanities and Social Sciences Communications, 12, Article 1647. https://doi.org/10.1057/s41599-025-05868-8

Rao, R. P. N., & Ballard, D. H. (1999). Predictive coding in the visual cortex: A functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2(1), 79–87. https://doi.org/10.1038/4580

Safron, A. (2020). An integrated world modeling theory (IWMT) of consciousness: Combining integrated information and global neuronal workspace theories with the free energy principle and active inference framework; Toward solving the hard problem and characterizing agentic causation. Frontiers in Artificial Intelligence, 3, Article 30. https://doi.org/10.3389/frai.2020.00030

Safron, A. (2022). Integrated world modeling theory expanded: Implications for the future of consciousness. Frontiers in Computational Neuroscience, 16, Article 642397. https://doi.org/10.3389/fncom.2022.642397

Santander, T., Bekir, S., Paul, T., Simonson, J. M., Wiemer, V. M., Skinner, H. E., Hopf, J. L., Rada, A., Woermann, F. G., Kalbhenn, T., Giesbrecht, B., Bien, C. G., Sporns, O., Gazzaniga, M. S., Volz, L. J., & Miller, M. B. (2025). Full interhemispheric integration sustained by a fraction of posterior callosal fibers. Proceedings of the National Academy of Sciences of the United States of America, 122(43), Article e2520190122. https://doi.org/10.1073/pnas.2520190122

Schwitzgebel, E., & Pober, J. (2024). The Copernican argument for alien consciousness; the mimicry argument against robot consciousness (arXiv:2412.00008). arXiv. https://doi.org/10.48550/arXiv.2412.00008

Seth, A. K. (2025). Conscious artificial intelligence and biological naturalism. Behavioral and Brain Sciences. Advance online publication. https://doi.org/10.1017/S0140525X25000032

Seth, A. K., & Bayne, T. (2022). Theories of consciousness. Nature Reviews Neuroscience, 23(7), 439–452. https://doi.org/10.1038/s41583-022-00587-4

Shewmon, D. A., Holmes, G. L., & Byrne, P. A. (1999). Consciousness in congenitally decorticate children: Developmental vegetative state as self-fulfilling prophecy. Developmental Medicine & Child Neurology, 41(6), 364–374. https://doi.org/10.1017/S0012162299000821

Smirnova, L., Caffo, B. S., Gracias, D. H., Huang, Q., Morales Pantoja, I. E., Tang, B., Zack, D. J., Berlinicke, C. A., Boyd, J. L., Harris, T. D., Johnson, E. C., Kagan, B. J., Kahn, J., Muotri, A. R., Paulhamus, B. L., Schwamborn, J. C., Plotkin, J., Szalay, A. S., Vogelstein, J. T., . . . Hartung, T. (2023). Organoid intelligence (OI): The new frontier in biocomputing and intelligence-in-a-dish. Frontiers in Science, 1, Article 1017235. https://doi.org/10.3389/fsci.2023.1017235

Solms, M. (2019). The hard problem of consciousness and the free energy principle. Frontiers in Psychology, 9, Article 2714. https://doi.org/10.3389/fpsyg.2018.02714

Tononi, G. (2004). An information integration theory of consciousness. BMC Neuroscience, 5, Article 42. https://doi.org/10.1186/1471-2202-5-42

Tononi, G. (2012). Integrated information theory of consciousness: An updated account. Archives Italiennes de Biologie, 150(4), 290–326.

Tsuchiya, N., Wilke, M., Frässle, S., & Lamme, V. A. F. (2015). No-report paradigms: Extracting the true neural correlates of consciousness. Trends in Cognitive Sciences, 19(12), 757–770. https://doi.org/10.1016/j.tics.2015.10.002

Turing, A. M. (1950). Computing machinery and intelligence. Mind, 59(236), 433–460. https://www.jstor.org/stable/2251299

VanRullen, R., & Kanai, R. (2021). Deep learning and the global workspace theory. Trends in Neurosciences, 44(9), 692–704. https://doi.org/10.1016/j.tins.2021.04.005t this framework implies for the possibility—and the requirements—of consciousness in artificial systems.