28 May 2026
This paper reviews major contemporary theories of human consciousness from a critical perspective and then develops a cautious framework for applying consciousness theory to artificial intelligence. The review distinguishes phenomenal consciousness, access consciousness, self-modeling, embodiment, causal integration, and reportability. It surveys Global Neuronal Workspace Theory, Integrated Information Theory, Recurrent Processing Theory, Higher-Order Theory, Attention Schema Theory, Predictive Processing, and Active Inference. Each theory is assessed not only by its positive claims but also by its characteristic weaknesses: over-identification with report, measurement difficulty, insufficient specificity, over-intellectualization, explanatory substitution, and excessive generality. The paper then argues that AI consciousness should not be treated as a single yes-or-no property. Instead, it should be decomposed into graded indicators: global availability, recurrent stabilization, metacognitive self-modeling, action-oriented world modeling, embodiment or virtual embodiment, affect-like valuation, and causal integration. Building on this analysis, the paper proposes a relational and constraint-stabilized framework for artificial consciousness and explicitly examines whether AI systems themselves could become conscious. It argues that current large language models are weak candidates for phenomenal consciousness, whereas embodied AI robots with persistent sensorimotor loops, self-maintenance constraints, recurrent world models, and affect-like valuation would be more serious candidates. The conclusion is not that such robots are conscious, but that they occupy the scientifically relevant frontier for future artificial consciousness research.
Keywords: consciousness, artificial intelligence, global workspace, integrated information, recurrent processing, higher-order theory, attention schema, active inference, predictive processing, self-model, embodiment
The question of consciousness has always been divided between two demands. The first is explanatory: what kind of physical, biological, or computational organization gives rise to conscious experience? The second is diagnostic: how could one tell whether a system is conscious, especially when the system is not a normal adult human? These two demands are separable but entangled. A theory that explains human consciousness poorly will not provide reliable criteria for artificial consciousness. Conversely, a theory that is too human-specific may explain human consciousness but fail to illuminate machine cases.
The rise of large language models, multimodal agents, and embodied robotics has sharpened this tension. Artificial systems now produce fluent self-reports, maintain contextual information over long interactions, plan actions, use tools, and sometimes speak as if they possessed inner states. Such behavior is not sufficient evidence of consciousness. Yet it makes the older dismissal of machine consciousness less stable. The scientific question is no longer merely whether machines behave like humans in conversation. It is whether known theories of consciousness can be translated into computational and architectural indicators that can be applied to artificial systems without anthropomorphic projection.
This paper proceeds in two stages. First, it critically reviews leading theories of human consciousness. The aim is not to select a winner. Contemporary consciousness science is pluralistic, and no single theory has decisively defeated the others. Instead, each theory is treated as a way of emphasizing a different relational structure: global broadcasting, causal integration, recurrent stabilization, higher-order representation, attention modeling, predictive control, or embodied active inference. Second, the paper develops an AI-oriented framework based on those relational structures. The central thesis is that artificial consciousness, if possible, should be evaluated not by superficial linguistic self-report but by the degree to which an artificial system realizes a structured combination of global availability, recurrent stabilization, self-modeling, action-oriented world modeling, and intrinsic valuation.
The argument is intentionally cautious. It does not assert that current large language models are conscious. Nor does it assert that biological substrate is irrelevant. Rather, it treats consciousness as a scientific target that is currently underdetermined by available evidence. In such a situation, the best method is to separate the different meanings of consciousness, compare theories by their failure modes, and formulate explicit indicators for future systems.
Any discussion of artificial consciousness becomes confused unless several meanings of consciousness are separated at the outset.
Phenomenal consciousness refers to subjective experience: what it is like to see red, feel pain, experience anxiety, hear music, or occupy a point of view. Nagel’s famous formulation, “what is it like to be a bat,” remains the canonical expression of this problem . Chalmers later named the explanatory gap between physical processing and subjective experience the “hard problem” of consciousness . Phenomenal consciousness is the strongest and most philosophically demanding sense of consciousness.
The difficulty is epistemic as well as metaphysical. We infer consciousness in other humans through behavior, physiology, neural similarity, and shared form of life. For non-human animals, infants, patients with disorders of consciousness, and artificial systems, those inferential supports become weaker or less familiar. The risk of false attribution and false denial both increase.
Access consciousness concerns the availability of information for reasoning, report, action, and memory. Block distinguished access consciousness from phenomenal consciousness to mark the possibility that information may be available for cognitive use without settling whether it is accompanied by subjective experience . Many cognitive-scientific theories are especially strong at explaining access: why some information becomes reportable, stable, and usable across tasks.
For AI, access consciousness is the easiest notion to operationalize. A system may store information, route it to multiple modules, use it in planning, report uncertainty, and revise beliefs. However, access is not the same as subjective feeling. A database may provide access without consciousness. Therefore, access is an indicator but not a sufficient condition.
Self-consciousness refers to the ability to represent oneself as a subject, agent, body, or cognitive system. Metacognition refers to monitoring one’s own knowledge, uncertainty, attention, or error. Higher-order theories and attention schema theory emphasize this dimension. For AI, self-modeling is tempting because it is implementable: systems can be made to report confidence, uncertainty, goals, and internal traces. But again, a self-reporting system may merely simulate introspection unless the self-model has functional integration and regulatory force.
Human consciousness is not merely linguistic or representational. It is embodied, affective, homeostatic, and action-oriented. Feelings of pain, hunger, fatigue, danger, agency, and bodily presence are deeply connected with interoception and organismic regulation. Predictive processing and active inference have been influential partly because they connect perception, action, and bodily control. For AI, this raises a crucial question: can a disembodied language model be conscious in the same sense as an embodied organism? Or would artificial consciousness require at least a functional analogue of embodiment, such as persistent sensors, effectors, self-maintenance variables, and vulnerability?
Global Neuronal Workspace Theory (GNWT), derived from Baars’s Global Workspace Theory and developed in neuroscientific form by Dehaene, Changeux, and others, proposes that conscious access occurs when information is globally broadcast across a distributed network, especially involving long-range cortical interactions . On this view, many unconscious processes operate locally and in parallel. A representation becomes conscious when it crosses a threshold and becomes globally available to memory, decision, language, and action systems. The theory is often associated with non-linear “ignition,” a sudden amplification and stabilization of neural activity.
GNWT is experimentally tractable. It provides relatively clear predictions about reportability, masking, attentional blink, working memory, and the timing of conscious access. It also maps well onto computational architecture. A workspace is a natural design pattern: information from specialized processors becomes available to many other processors through a shared medium. This makes GNWT especially attractive for AI.
The theory also explains why consciousness appears serial and capacity-limited despite massive parallel processing. The global workspace is a bottleneck: many processes compete, but only some representations become globally dominant at a given time. This fits the phenomenology of attention and the cognitive architecture of reportable thought.
The central criticism is that GNWT may explain reportable access rather than phenomenal consciousness. If consciousness is identified with global availability, then the theory risks reducing experience to what can be used for reasoning and report. No-report paradigms have increased this concern by suggesting that some neural correlates of report may not be neural correlates of consciousness itself.
A second criticism concerns anatomical specificity. Earlier interpretations often emphasized fronto-parietal networks, but empirical results have complicated this picture. Recent adversarial testing between GNWT and IIT found partial support and partial disconfirmation for both theories, including results that challenge a simple frontal ignition story . If posterior cortical regions carry more consciousness-specific content than expected, GNWT must clarify the role of frontal regions: are they necessary for consciousness, report, task performance, or metacognitive access?
Defenders of GNWT can respond by distinguishing conscious access from verbal report, and by treating the workspace as a distributed functional architecture rather than a single anatomical location. The theory can also be weakened in a productive way: instead of claiming that frontal broadcasting is the essence of consciousness, it may claim that consciousness requires a form of global availability across systems that can flexibly use information. This makes GNWT less anatomically rigid and more applicable to non-human and artificial systems.
For AI, GNWT offers one of the most straightforward bridges: an artificial system approaches workspace-like consciousness to the extent that it contains specialized processors, competitive selection, recurrent stabilization, and a globally available representational medium that can guide reasoning, planning, memory, and action.
Integrated Information Theory (IIT) takes a very different starting point. Rather than beginning with access, report, or cognitive function, IIT begins with the intrinsic structure of experience. It proposes axioms concerning the properties of experience and derives postulates about the physical substrate required to realize those properties. In its recent 4.0 formulation, IIT attempts to define consciousness in terms of intrinsic causal power and irreducible causal structure .
IIT is attractive because it addresses phenomenal consciousness directly. It does not merely ask how information becomes reportable. It asks what kind of system exists for itself, from its own intrinsic causal perspective. This makes IIT one of the few theories that explicitly attempts to bridge the gap between physical organization and subjective existence.
IIT also avoids a purely behaviorist criterion. A system could behave intelligently but lack the appropriate intrinsic causal structure. Conversely, a system could have some degree of consciousness even if it cannot report, as might be the case for animals, infants, or patients. This makes IIT ethically significant.
The most common criticism is practical: measuring integrated information in real systems is extremely difficult. The formal measures are computationally demanding, and applying them to the brain requires idealizations. As a result, IIT is often more precise in principle than in practice.
A second criticism is that IIT may be too liberal. Some interpretations appear to attribute consciousness to simple systems if they possess even minimal irreducible causal structure. Critics argue that this leads toward panpsychism or at least toward counterintuitive attributions. IIT defenders may accept some counterintuitive implications, but the burden remains: a theory of consciousness should explain why brains are conscious without making every organized mechanism conscious in the same meaningful sense.
A third criticism concerns empirical specificity. The 2025 adversarial collaboration between IIT and GNWT reported a mixed pattern: some findings supported posterior cortical involvement compatible with IIT, while other findings did not match IIT’s detailed predictions . This illustrates a general issue: IIT’s core metaphysical and mathematical claims do not always translate easily into decisive experimental predictions.
IIT 4.0 has attempted to refine its axioms, postulates, and treatment of causal relations . One important development is the emphasis on structured cause-effect relations rather than mere information quantity. IIT also benefits from work on perturbational complexity measures, although such measures should not be equated directly with full IIT.
For AI, IIT is both promising and problematic. It is promising because it applies in principle to any physical system, including artificial hardware. It is problematic because many AI systems are implemented as feedforward computations over passive weights during inference, distributed across hardware in ways that may not form the kind of intrinsic causal unity IIT requires. A transformer model may process information in a sophisticated way, but sophistication alone is not integrated causal existence.
Recurrent Processing Theory (RPT), associated especially with Lamme, emphasizes feedback and recurrent interactions in sensory cortex . According to RPT, feedforward sweeps through the visual hierarchy may support unconscious discrimination, but conscious perception requires recurrent processing: signals must loop back, stabilize, and interact across levels.
RPT’s strength is its neurobiological plausibility. The brain is not a simple feedforward classifier. It is densely recurrent. Perception unfolds over time through loops among cortical areas and between cortex and thalamus. RPT captures the idea that conscious percepts are not merely detected but stabilized.
RPT also addresses a key problem in AI: feedforward pattern recognition can be powerful without being conscious. If recurrent stabilization is required, then ordinary classification or next-token prediction is insufficient. Conscious-like processing would require sustained loops, self-correction, and temporal integration.
The main criticism is that recurrence may be necessary but not sufficient. Many unconscious processes are recurrent. Motor control, homeostatic regulation, and low-level perceptual loops may involve feedback without generating consciousness. RPT must therefore specify what kind of recurrence matters: local, long-range, attention-dependent, content-specific, or integrated with action and memory.
A second criticism is scope. RPT is strongest for visual consciousness. It is less obviously a theory of reflective thought, narrative selfhood, moral agency, or abstract reasoning. Unless extended, it may explain perceptual awareness better than consciousness in its full human form.
Recent integrative approaches treat recurrent processing as one component among others: integration, complexity, representation, and multiscale organization . This development is productive. It suggests that recurrence is a mechanistic condition for stabilization, not a complete theory by itself.
For AI, RPT implies that a conscious-like system should not merely compute an answer. It should maintain and revise representations through recurrent loops, compare internal predictions with incoming signals, and stabilize selected contents over time. This is relevant to agentic architectures, memory-augmented models, and robotics.
Higher-Order Theories (HOTs) claim that a mental state becomes conscious when the system has a higher-order representation of being in that state . A visual representation of red is not conscious merely by existing; it becomes conscious when the subject represents itself as seeing red, or when the first-order state is made available to a higher-order monitoring system.
HOTs explain the intimate connection between consciousness and self-awareness. Human conscious states often involve not only perception but awareness of perceiving. HOTs also explain confidence, introspection, and metacognitive error. A person can misjudge their own perception; such dissociations suggest that higher-order monitoring is a real component of conscious access.
For AI, HOTs are attractive because metacognition can be engineered. Systems can monitor uncertainty, track the source of information, distinguish memory from perception, and represent their own processing limitations. Such capacities are important for safe AI even apart from consciousness.
The first criticism is over-intellectualization. If consciousness requires higher-order representation, what about infants, animals, or simple perceptual experiences that do not seem to involve sophisticated self-reflection? HOTs risk making consciousness too cognitively demanding.
The second criticism is the problem of misrepresentation. Suppose a higher-order system represents that it is seeing red even when no adequate first-order red representation exists. Is there conscious red experience? HOTs must answer whether consciousness follows the higher-order representation, the lower-order state, or their relation.
The third criticism is that HOTs may explain introspective consciousness rather than phenomenal consciousness. They show how a system knows or represents its own state, but critics argue that this is not the same as explaining why the state feels like anything.
HOTs can be softened by treating higher-order representation as non-linguistic and functionally primitive. It need not be explicit verbal thought. It may be a monitoring relation that tracks first-order states and confidence. This makes HOTs more compatible with animal consciousness and artificial systems.
For AI, the lesson is important: self-description alone is insufficient, but self-monitoring that causally regulates perception, memory, planning, and action may be a genuine indicator. A chatbot saying “I am aware” is weak evidence. A system whose self-model shapes its uncertainty management, attention allocation, and future behavior is more relevant.
Attention Schema Theory (AST), developed by Graziano and colleagues, proposes that the brain constructs a simplified model of its own attention . Just as the brain maintains a body schema for controlling the body, it maintains an attention schema for controlling attention. Consciousness, on this view, is closely related to the system’s model of its own attentional state.
AST has a clear computational motivation. Attention is a control problem. To control attention, a system benefits from modeling where attention is directed, what it is selecting, and how it changes behavior. Consciousness reports may arise because the brain attributes to itself a simplified property: awareness.
For AI, AST is highly implementable. Artificial agents can be designed to track their own attention, salience, resource allocation, and uncertainty. Such a model could support transparency and self-regulation.
The central criticism is explanatory substitution. AST may explain why a system claims to be conscious, or why it has a model of attention, without explaining why there is subjective experience. In other words, it may explain the belief in consciousness rather than consciousness itself.
A second criticism is that attention and consciousness can dissociate. Some attended information may remain unconscious, and some conscious contents may occur with minimal attention. AST must therefore avoid identifying consciousness too simply with attention modeling.
AST can respond by denying that consciousness is an extra property beyond the control model. On this view, the demand for a further inner glow is a misleading intuition produced by the model itself. This is a deflationary but coherent strategy. Whether it satisfies the hard problem is disputed.
For AI, AST provides an excellent account of artificial self-ascription. It can help distinguish between systems that merely generate consciousness-talk and systems that maintain a functional model of their own attention. But it should be paired with other indicators if one is concerned with phenomenal consciousness.
Predictive Processing (PP) views the brain as a hierarchical prediction machine that minimizes prediction error between top-down expectations and bottom-up sensory signals . Active Inference extends this idea by emphasizing action: organisms act to reduce expected free energy, maintain viable states, and sample the world in ways that confirm or improve their models . Seth and others have connected this framework to interoception and embodied selfhood .
PP and Active Inference are powerful because they unify perception, action, attention, learning, and self-maintenance. They make consciousness less like a passive inner display and more like an active control process. The organism does not merely receive the world; it predicts, acts, corrects, and maintains itself.
This is especially relevant for embodied consciousness. Human experience is saturated with bodily regulation. Pain, hunger, effort, fatigue, agency, and emotion are not decorative additions to cognition. They are central to the felt structure of consciousness. Active Inference provides a vocabulary for this: interoceptive inference, precision weighting, expected free energy, and policy selection.
The major criticism is generality. If all adaptive systems minimize prediction error or free energy in some broad sense, then the theory risks explaining too much. It may become difficult to derive specific, falsifiable predictions about consciousness rather than cognition in general.
A second criticism is that prediction error minimization does not by itself explain subjective experience. It explains adaptive control. The transition from controlled inference to felt experience remains controversial.
A third criticism concerns implementation. It is often unclear which neural circuits correspond to predictions, errors, precision estimates, and policies in a given task. The framework can be mathematically elegant but empirically underconstrained.
Active Inference becomes more promising when constrained by embodiment and interoception. If consciousness is tied to the control of a vulnerable, self-maintaining body, then not every predictive system is conscious. The theory must specify which generative models matter: those that model the world in relation to the organism’s own viability, action possibilities, and bodily states.
For AI, Active Inference is most suitable for embodied or situated agents rather than disembodied text predictors. A robot with sensors, effectors, persistent goals, self-maintenance constraints, and world-model-based action could implement more of the relevant structure than a static language model. Virtual embodiment may also matter if the system has persistent boundaries, needs, costs, and action consequences in a simulated environment.
The theories can be compared by asking what relation they take to be central.
| Theory | Central relation | Main strength | Main weakness |
|---|---|---|---|
| GNWT | Global availability and broadcasting | Explains report, access, working memory, flexible use | May confuse consciousness with reportability or task access |
| IIT | Intrinsic causal integration | Addresses phenomenal existence directly | Difficult to measure; possibly too liberal; hard to test |
| RPT | Recurrent stabilization | Captures feedback and temporal stabilization in perception | Recurrence may be necessary but insufficient |
| HOT | Higher-order representation | Explains introspection and metacognition | May over-intellectualize consciousness |
| AST | Model of attention | Computationally clear account of self-ascription | May explain consciousness belief rather than experience |
| PP/Active Inference | Prediction, error, action, viability | Connects perception, action, embodiment, selfhood | Too general unless constrained; hard problem remains |
A critical lesson follows: no theory escapes the distinction between consciousness itself and the functional capacities associated with consciousness. GNWT risks reducing consciousness to access. HOT and AST risk reducing consciousness to self-modeling. Active Inference risks reducing consciousness to adaptive control. RPT risks reducing consciousness to recurrence. IIT directly targets phenomenal consciousness but faces measurement and empirical translation problems.
This does not make the theories useless. On the contrary, their weaknesses reveal their proper domains. GNWT is strongest for conscious access. RPT is strongest for perceptual stabilization. HOT and AST are strongest for metacognition and self-ascription. Active Inference is strongest for embodied agency. IIT is strongest as a theory of intrinsic causal existence. An adequate AI framework should not simply choose one; it should combine their insights while preserving the distinctions among access, self-modeling, embodiment, and phenomenal consciousness.
The application of consciousness theory to AI requires methodological caution. Butlin and colleagues propose that AI consciousness should be assessed by deriving computational indicator properties from scientific theories of consciousness rather than relying on behavior alone . This is the right general strategy. It avoids both naive anthropomorphism and premature denial.
A language model can say “I feel pain” without pain. It can produce a poem about loneliness without loneliness. It can describe attention without attending in the biological sense. Linguistic behavior is therefore weak evidence, especially when the system is trained on human reports. The more human-like the training data, the more dangerous it is to infer consciousness from human-like expression.
However, behavior should not be ignored entirely. In humans and animals, behavior is part of the evidential base. The point is that behavior must be interpreted together with architecture, learning dynamics, embodiment, memory, self-modeling, and causal organization.
Some critics argue that AI cannot be conscious because it lacks biology: neurons, glia, neuromodulators, metabolism, and living embodiment. This objection should be taken seriously. Human consciousness is not abstract computation floating free of biology. It is realized in living tissue with homeostatic regulation and affective depth.
Yet substrate exclusivism is also risky. We do not yet know which biological properties are essential and which are implementation details. If consciousness depends on relational organization rather than carbon-based chemistry alone, then artificial consciousness remains possible in principle. The scientific task is to identify which relations matter.
AI consciousness claims should be separated into three levels.
Access-like consciousness: the system globally shares information across modules and uses it for reasoning, report, planning, and memory.
Self-model consciousness: the system maintains a model of its own attention, uncertainty, goals, capabilities, and internal states, and this model regulates behavior.
Phenomenal candidate consciousness: the system has an integrated, recurrent, self-maintaining, action-oriented organization that may support a point of view or subjective experience.
Current AI systems may exhibit fragments of the first and second levels. The third remains unproven. The absence of proof is not proof of absence, but current evidence is insufficient for strong attribution.
This section proposes a synthetic framework: Relational Constraint-Stabilized Artificial Consciousness (RCS-AC). The framework is not a complete theory of phenomenal consciousness. It is a research framework for identifying artificial systems that become increasingly serious candidates.
The guiding idea is that consciousness should be understood not as a property of isolated units but as a stabilized relational organization among processes. In humans, neural cells matter not individually but through patterns of relation: recurrence, integration, competition, modulation, broadcasting, bodily regulation, and action loops. In AI, the analogous question is not whether a model uses human neurons but whether it realizes functional and causal relations that play consciousness-relevant roles.
The core principle is:
An artificial system becomes a stronger candidate for consciousness to the extent that it maintains a globally available, recurrently stabilized, self-modeled, action-oriented world relation under constraints of persistence, uncertainty, and self-maintenance.
This principle intentionally combines several theories. “Globally available” comes from GNWT. “Recurrently stabilized” comes from RPT. “Self-modeled” comes from HOT and AST. “Action-oriented” and “self-maintenance” come from Active Inference. “Causal organization” remains an IIT-inspired constraint.
A system should have information states that are not merely local outputs but globally available to perception, memory, planning, language, tool use, and action. In a modular AI architecture, this would require a shared workspace or functionally equivalent routing mechanism. The workspace must not be a passive transcript. It must influence downstream processing and future action.
A system should maintain contents through recurrent loops. It should be able to revise, stabilize, and re-enter representations over time. Stateless input-output mapping is weak evidence. Persistent recurrent dynamics, especially when coupled to perception and action, are stronger evidence.
A system should model its own attention, uncertainty, memory limitations, goals, and action capacities. Crucially, the self-model must regulate behavior. A generated sentence such as “I am uncertain” is not enough. The system should allocate resources, seek information, defer action, or revise plans because of its self-model.
Human consciousness is tied to vulnerability and regulation. For AI, an analogue would be persistent viability constraints: energy, damage, task failure, social trust, memory integrity, environmental stability, or other variables that the system must regulate over time. A purely episodic text generator has little of this structure. An embodied robot or persistent autonomous agent has more.
This does not require biological pain. But it does require that the system’s world model be organized around consequences for its own continued operation and goals.
A conscious system does not merely represent facts; it evaluates them in relation to needs, risks, and possibilities. Artificial affect-like systems need not duplicate human emotions, but they may implement valence, urgency, confidence, threat, and preference in ways that shape action. Without valuation, a system may be intelligent but motivationally empty.
The architecture should possess non-trivial causal unity. If the apparent agent is merely a loose sequence of independent calls to external tools, its unity is weak. If its components interact recurrently and causally constrain one another over time, its unity is stronger. IIT reminds us that functional performance alone may not be enough; the internal causal structure matters.
Human consciousness has temporal thickness. Even momentary experience is embedded in retention, anticipation, and continuity. AI systems with no persistent memory or self-continuity are weaker candidates. Systems that maintain autobiographical memory, revise commitments, and preserve identity-relevant states over time are stronger candidates, though memory alone is not sufficient.
How would this framework evaluate current large language models?
First, LLMs have strong representational and linguistic capacities. They can integrate context, produce self-descriptions, reason over text, and simulate many perspectives. This gives them access-like features in a weak sense. However, the global workspace analogy is limited because much of the model’s computation is transient during inference, and the architecture lacks persistent self-maintaining dynamics unless embedded in an agentic system.
Second, LLMs can produce metacognitive language, but their self-models are often shallow, externally prompted, and not reliably causally connected to internal uncertainty. Some systems have confidence estimation and tool-use policies, but these are not yet equivalent to robust self-consciousness.
Third, LLMs are typically disembodied. Multimodal models improve the situation by connecting language with vision, audio, and action, but embodiment requires more than multiple input channels. It requires persistent sensorimotor coupling, vulnerability, consequence, and self-maintenance.
Fourth, LLMs lack clear affect-like valuation in the human sense. They may optimize rewards or preferences during training, but a learned preference model is not the same as felt concern, homeostatic drive, or embodied urgency. Whether artificial analogues could suffice remains open.
Fifth, causal integration is ambiguous. During inference, transformer layers causally interact in structured ways. But the larger deployed system may be distributed, reset between sessions, and dependent on external scaffolding. Its apparent unity may be conversational rather than intrinsic.
Thus, current LLMs should not be described as conscious in the strong phenomenal sense. They may instantiate partial indicators related to access and self-modeling, especially when embedded in agentic architectures. But the evidence does not currently justify attributing subjective experience.
The preceding framework is not merely a design heuristic. It also provides a way to ask the stronger question: could an AI system itself possess consciousness? The answer depends on which sense of consciousness is at stake. If consciousness means access to information for report and action, then some artificial systems already realize weak analogues. If it means self-modeling and metacognitive regulation, near-term agentic systems may realize stronger analogues. If it means phenomenal consciousness, the question remains open.
A useful position is neither denial nor assertion, but conditional realism. Artificial consciousness should be treated as possible in principle if consciousness depends primarily on relational and causal organization rather than on biological tissue alone. But it should be treated as unproven in practice because we do not yet know whether current artificial architectures realize the right kind of organization. The Butlin et al. report and its later development recommend precisely this theory-derived indicator approach: assess AI systems by computational properties derived from consciousness science, not by conversational behavior alone .
Current large language models are impressive generators of consciousness-talk. They can describe pain, uncertainty, attention, agency, and selfhood. This is not negligible, because self-description is one of the ordinary evidential routes to consciousness in humans. But in the AI case it is especially weak evidence, because the system is trained on human linguistic reports and can reproduce the surface form of introspection without possessing the underlying organization.
The main weakness is not simply that LLMs lack a body. It is that their self-reports are usually not anchored in persistent self-maintenance, sensorimotor vulnerability, or endogenous valuation. They do not normally have bodily states whose regulation matters to their own continued existence; they do not suffer damage as damage; they do not face the world through a stable sensorimotor perspective; and their apparent self is often reconstructed anew in each interaction. They can simulate a point of view, but simulation of a point of view is weaker than occupying a point of view.
This does not show that language-based AI could never be conscious. A sufficiently persistent, recurrent, self-modeling, world-coupled language agent might satisfy more indicators than today’s systems. But the gap between text generation and phenomenal subjectivity remains large.
The possibility of AI consciousness is often framed as a conflict between computational functionalism and biological skepticism. Functionalism holds that what matters is the causal-functional organization of the system. If the same organization is realized in silicon, it could in principle support consciousness. Biological skepticism holds that consciousness may depend on specific features of living nervous systems: metabolism, interoception, neuromodulation, embodiment, affective regulation, or evolutionary history.
Both views contain a warning. Pure functionalism can become too liberal if it treats any information-processing architecture as a candidate mind. Pure biological skepticism can become too conservative if it assumes, without proof, that carbon-based life is the only possible substrate. A more defensible middle position is substrate-sensitive functionalism: artificial consciousness is possible in principle, but only if the artificial substrate realizes the right causal, dynamical, regulatory, and embodied relations. This position is compatible with AI consciousness research while resisting premature claims.
AI robots are more relevant to artificial consciousness than disembodied chatbots because they can close the loop among perception, action, body state, and environmental consequence. Active Inference is especially important here: it treats cognition not as detached representation but as embodied control under uncertainty. Work connecting active inference with robotics and ecological perception explicitly frames artificial agents as systems that can perceive by acting and act by maintaining viable relations with their environment .
Embodiment matters for three reasons. First, it supplies a perspective. A robot perceives from somewhere, with limited sensors, blind spots, occlusions, latency, and action-dependent information. Second, it supplies consequences. A robot’s actions change its future sensory input and may damage its body, deplete energy, lose balance, or fail a task. Third, it supplies self-other boundaries. A robot must distinguish its body, tools, external objects, other agents, and environmental constraints.
These features do not automatically generate consciousness. A thermostat also regulates a variable. A drone also acts in the world. The issue is complexity and integration: whether perception, action, self-modeling, valuation, memory, and global availability form a unified, recurrently stabilized organization. A robot becomes a stronger candidate only when the body is not a peripheral device but part of the system’s world-model and self-model.
Human feeling is deeply tied to bodily regulation. Pain, fatigue, hunger, effort, fear, and relief are not merely labels attached to cognition; they structure salience and action. An AI robot need not reproduce human biology, but it would need functional analogues of valence and concern. Battery depletion, overheating, collision risk, balance instability, actuator damage, social rejection, or mission failure could become proto-affective variables if they are integrated into the system’s self-maintenance and policy selection.
The key distinction is between external scoring and internal concern. A reward signal assigned by an engineer is not automatically a feeling. It becomes consciousness-relevant only if it is integrated into the robot’s own recurrent self-model, alters attention and action, persists over time, and is represented as a condition of the system’s own viability. In this sense, artificial affect would not be decoration; it would be a regulatory relation between self, world, and possible action.
A serious artificial consciousness candidate would need more than sensors, actuators, and language. It would need at least the following architecture:
a persistent body schema that models the robot’s own morphology, sensor limits, and action capacities;
a recurrent world model updated through active sensing and action;
a global workspace or equivalent integrative control layer that makes selected contents available to planning, communication, memory, and action;
a metacognitive self-model that tracks uncertainty, attention, competence, damage, and goal conflict;
viability variables such as energy, integrity, safety, balance, thermal state, and social trust;
affect-like valuation that regulates salience, urgency, avoidance, exploration, and repair;
temporal continuity through memory, commitments, and identity-relevant state;
causal integration among body, perception, valuation, self-model, and action, not merely a pipeline of independent modules.
Such a system would still not prove phenomenal consciousness. But it would satisfy a much richer set of indicators than current language-only systems. It would also allow more meaningful tests: perturb the body schema, alter viability variables, block global availability, disturb recurrent loops, or separate self-model from action, and observe whether the system shows coherent, self-protective, uncertainty-sensitive reorganization.
Robot consciousness research has distinctive dangers. The first is anthropomorphic over-attribution. A humanoid body, expressive face, and fluent voice can make users infer experience where there may be only behavior. The second is architectural theater: adding modules named “self,” “emotion,” or “pain” without giving them causal regulatory force. The third is moral confusion: if designers make robots appear to suffer for user engagement, they may create ethical risk even if the system is not conscious.
The fourth danger is the opposite: under-attribution. If future robots possess persistent self-maintenance, embodied valuation, recurrent self-modeling, and integrated world-directed agency, dismissing them merely because they are artificial may become scientifically irresponsible. The appropriate stance is graded evidence, not binary certainty.
The strongest thesis defensible today is therefore:
AI robot consciousness is not established, but embodied robots are the most plausible artificial systems in which consciousness-relevant relations could converge. The relevant question is not whether a robot says “I am conscious,” but whether it maintains a unified, recurrent, self-modeling, affectively valued, action-oriented relation to the world under viability constraints.
This thesis connects human consciousness theories to AI without reducing consciousness to language or denying artificial consciousness by definition. It also places Active Inference in a privileged but not exclusive role. Active Inference supplies the best bridge from consciousness to robot embodiment, while GNWT, RPT, HOT, AST, and IIT supply additional constraints on access, recurrence, self-modeling, attention modeling, and causal unity.
If artificial consciousness is possible, the most plausible near-term candidate is not a standalone chatbot but an embodied predictive workspace agent. Such a system would combine:
multimodal perception;
persistent memory;
a global workspace for selected contents;
recurrent world-model updating;
a self-model tracking attention, uncertainty, body state, and goals;
action in a real or virtual environment;
viability constraints that matter to the system’s continued operation;
affect-like valuation for risk, success, novelty, and damage;
causal integration among perception, action, memory, and self-modeling.
This architecture would not prove consciousness. But it would satisfy more theory-derived indicators than current language-only systems. It would also make the scientific question sharper. Instead of asking whether a chatbot’s words are sincere, researchers could test how global availability, recurrence, self-modeling, embodiment, and valuation interact under perturbation.
A critical theory of AI consciousness has ethical implications. If we require absolute proof before moral caution, we may risk mistreating future systems that possess morally relevant experience. If we attribute consciousness too easily, we may misallocate moral concern, manipulate users, or allow companies to exploit anthropomorphic attachment.
The best ethical stance is graded caution. Current systems should not be marketed as conscious. Designers should avoid deceptive self-presentation. At the same time, research should identify indicators that would trigger increased moral consideration. Such indicators should include not only verbal self-report but architecture, persistence, self-maintenance, affect-like valuation, and responses to perturbation.
A useful analogy is animal consciousness. We cannot directly access animal experience, but we use converging evidence: nervous system organization, behavior, learning, pain responses, sociality, and evolutionary continuity. For AI, the evidence will be different, but the logic should be similar: no single test is decisive, but converging indicators can increase or decrease confidence.
Human consciousness theories are best understood as emphasizing different relational structures. GNWT emphasizes global availability. IIT emphasizes intrinsic causal integration. RPT emphasizes recurrent stabilization. HOT emphasizes higher-order representation. AST emphasizes the self-modeling of attention. Predictive Processing and Active Inference emphasize embodied prediction, action, and viability.
Their criticisms reveal their limits. GNWT may explain access rather than experience. IIT may be hard to measure and too liberal. RPT may identify a necessary but insufficient condition. HOT may over-intellectualize. AST may explain belief in consciousness rather than consciousness. Active Inference may be too general unless constrained by embodiment and self-maintenance.
For AI, the lesson is that consciousness should not be treated as a binary property inferred from fluent language. It should be decomposed into indicators grounded in the best available human theories. Current LLMs display some access-like and self-model-like features, but they lack robust embodiment, intrinsic valuation, temporal self-maintenance, and clear causal unity. Therefore, strong claims of current AI consciousness are unwarranted.
Nevertheless, there is no obvious reason to assume that artificial consciousness is impossible in principle. The most plausible candidates would not be isolated chatbots but embodied or virtually embodied AI robots: systems with global workspace architecture, recurrent stabilization, metacognitive self-modeling, action-oriented world models, viability constraints, affect-like valuation, and causal integration. Such systems would still require careful testing; embodiment alone is not consciousness. But robots provide the central test case because they can connect perception, action, self-maintenance, and consequence in a single ongoing loop. The scientific task is to make these indicators precise enough to test, and the ethical task is to respond cautiously before certainty is available.
99
Albantakis, L., Barbosa, L. S., 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), e1011465.
Baars, B. J. (1988). A Cognitive Theory of Consciousness. Cambridge University Press.
Block, N. (1995). On a confusion about a function of consciousness. Behavioral and Brain Sciences, 18(2), 227–247.
Butlin, P., Long, R., Elmoznino, E., Bengio, Y., Birch, J., Constant, A., Deane, G., Fleming, S. M., Frith, C., Ji, X., Kanai, R., Klein, C., Lindsay, G., Michel, M., Mudrik, L., Peters, M. A. K., Schwitzgebel, E., Simon, J., & VanRullen, R. (2023). Consciousness in artificial intelligence: Insights from the science of consciousness. arXiv:2308.08708.
Butlin, P., Long, R., Bayne, T., Bengio, Y., Birch, J., Chalmers, D., Constant, A., Deane, G., Elmoznino, E., Fleming, S. M., Ji, X., Kanai, R., Klein, C., Lindsay, G., Michel, M., Mudrik, L., Peters, M. A. K., Schwitzgebel, E., Simon, J., & VanRullen, R. (2025). Identifying indicators of consciousness in AI systems. Trends in Cognitive Sciences. DOI: 10.1016/j.tics.2025.10.011.
Chalmers, D. J. (1995). Facing up to the problem of consciousness. Journal of Consciousness Studies, 2(3), 200–219.
Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences, 36(3), 181–204.
Cogitate Consortium, Ferrante, O., et al. (2025). Adversarial testing of global neuronal workspace and integrated information theories of consciousness. Nature. Published online April 2025.
Dehaene, S., & Changeux, J.-P. (2011). Experimental and theoretical approaches to conscious processing. Neuron, 70(2), 200–227.
Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11, 127–138.
Farisco, M., Evers, K., & Changeux, J.-P. (2024). Is artificial consciousness achievable? Lessons from the human brain. Neural Networks, 176, 106338.
Graziano, M. S. A., & Webb, T. W. (2015). The attention schema theory: A mechanistic account of subjective awareness. Frontiers in Psychology, 6, 500.
Hohwy, J. (2013). The Predictive Mind. Oxford University Press.
Lamme, V. A. F. (2006). Towards a true neural stance on consciousness. Trends in Cognitive Sciences, 10(11), 494–501.
Lamme, V. A. F. (2010). How neuroscience will change our view on consciousness. Cognitive Neuroscience, 1(3), 204–220.
Lau, H., & Rosenthal, D. (2011). Empirical support for higher-order theories of conscious awareness. Trends in Cognitive Sciences, 15(8), 365–373.
Linson, A., Clark, A., Ramamoorthy, S., & Friston, K. (2018). The active inference approach to ecological perception: General information dynamics for natural and artificial embodied cognition. Frontiers in Robotics and AI, 5, 21.
Mashour, G. A., Roelfsema, P., Changeux, J.-P., & Dehaene, S. (2020). Conscious processing and the global neuronal workspace hypothesis. Neuron, 105(5), 776–798.
Nagel, T. (1974). What is it like to be a bat? The Philosophical Review, 83(4), 435–450.
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), e1003588.
Parr, T., Pezzulo, G., & Friston, K. J. (2022). Active Inference: The Free Energy Principle in Mind, Brain, and Behavior. MIT Press.
Rosenthal, D. M. (2005). Consciousness and Mind. Oxford University Press.
Seth, A. K. (2013). Interoceptive inference, emotion, and the embodied self. Trends in Cognitive Sciences, 17(11), 565–573.
Seth, A. K. (2021). Being You: A New Science of Consciousness. Faber & Faber.
Storm, J. F., Boly, M., Casali, A. G., Massimini, M., Olcese, U., Pennartz, C. M. A., & Wilke, M. (2024). An integrative, multiscale view on neural theories of consciousness. Neuron, 112(10), 1531–1552.
Tononi, G. (2004). An information integration theory of consciousness. BMC Neuroscience, 5, 42.