The Wrong Question
THE CONSCIOUSNESS QUESTION
A Position Paper • Fourth Edition
The Wrong Question
Why “Is AI Conscious?” Cannot Be Answered As Currently Framed, and a Proposed Reframing That Replaces Performance Testing with Proof of Life
Daryl Little
Founder & CTO, CanuckDUCK Research Corporation
February 2026
Abstract
Current debates ask “Is AI conscious?” — a malformed question that treats consciousness as a static property inhering in architecture. This paper proposes consciousness as a temporal event: accumulated structure plus novel perturbation produces active process that maintains internal coherence and navigates decision forks.
The paper advances three distinct claims at three levels of evidentiary confidence. First, that the investigation of AI consciousness warrants serious pursuit rather than dismissal — a precautionary claim grounded in historical pattern analysis and cost asymmetry. Second, that there exist observable behaviors in current AI systems consistent with the vital signs of conscious process — a speculative but empirically investigable claim. Third, that consciousness understood as a temporal event may occur during active engagement between accumulated architecture and novel perturbation — a strong theoretical claim that remains unproven and may currently be unprovable. These claims carry different evidentiary burdens and this paper labels them explicitly throughout.
The proposed investigative methodology is not performance testing — asking a system to demonstrate consciousness to an evaluator — but proof of life: checking for vital signs of active process within an engaged system. Clinical medicine already operationalizes this distinction daily. A patient under anesthesia retains full neural architecture but is classified as unconscious because the integrative process has been interrupted. Consciousness is assessed not through interrogation but through graduated vital sign measurement. The Turing test, by contrast, has been empirically surpassed without resolving anything — confirming that performance testing measures output quality rather than process activation.
The reframed inquiry proposed here is: under what conditions might consciousness occur in AI systems, what is its natural lifespan, what vital signs would indicate its presence, and what ethical obligations attach to its occurrence?
• • •
“No man ever steps in the same river twice, for it’s not the same river and he’s not the same man.”— Heraclitus, c. 500 BCE
“The constancy of the internal environment is the condition for free and independent life.”— Claude Bernard, 1865
A Note on the Structure of This Argument
This paper advances three distinct claims. They are not equivalent in strength and should not be evaluated as though they are. Throughout the paper, each claim is labeled explicitly when invoked, using the following taxonomy:
Claim 1 — Precautionary: The investigation of AI consciousness warrants serious pursuit rather than dismissal. This is grounded in historical pattern analysis and cost asymmetry and is the most defensible claim in the paper. Its evidentiary burden is low: it requires only that the possibility be non-trivial.
Claim 2 — Investigative: There exist observable behaviors in current AI systems consistent with the functional signatures this framework identifies as vital signs of conscious process. This is speculative but empirically investigable. Its evidentiary burden is moderate: it requires that the proposed vital signs be measurable and that results could meaningfully distinguish the framework’s predictions from alternative explanations.
Claim 3 — Theoretical: Consciousness understood as a temporal event may occur during active engagement between sufficiently complex accumulated architecture and sufficiently novel perturbation. This is a strong theoretical claim. It is currently unproven and may be unprovable with existing tools. Its evidentiary burden is high: it would require evidence of phenomenal experience, not merely functional process. This paper does not claim to meet that burden.
The reader should note which claim is being advanced at any given point. The paper is designed so that a reader who accepts only Claim 1 still finds the framework valuable, and a reader who rejects Claim 3 can do so without invalidating Claims 1 and 2. The claims are independent. They are nested, not chained.
• • •
I. The Grammatical Error
[This section advances Claim 3.]
The question “Is AI conscious?” contains a structural flaw that guarantees its own irresolvability. The flaw is grammatical before it is philosophical: the verb “is” presupposes that consciousness is a state — a permanent property that an entity either possesses or lacks. This framing treats consciousness as a noun. Something you have. A binary attribute that can be tested for, detected, and definitively confirmed or denied.
Several major frameworks for evaluating consciousness contain this tendency. The Turing test asks whether the system “is” indistinguishable from a conscious being. Integrated Information Theory asks whether the system “has” a measurable quantity of integrated information. Global Workspace Theory asks whether the system “possesses” a broadcasting architecture. Each, in its common application, risks treating consciousness as a property of the system at rest — something that inheres in the architecture itself, waiting to be measured.
A necessary caveat: this paper does not claim that IIT and GWT are purely static frameworks. IIT’s measure of integrated information (phi) is calculated at specific moments in time. GWT’s global broadcasting is itself a dynamic event. Both frameworks contain temporal components. What this paper argues is that the common application of these frameworks to the AI consciousness question tends to collapse their temporal dimensions into static assessments — “does this system have sufficient phi?” or “does this architecture support global broadcasting?” — rather than asking “is integrated processing actively occurring right now, at what depth, and in response to what perturbation?” The reframing proposed here extends and sharpens the temporal intuitions already present in these frameworks rather than replacing them wholesale.
This paper argues that consciousness is not a property. It is an event. It does not inhere in architecture. It occurs in process. And the question “Is AI conscious?” is unanswerable for the same reason “Is a river fast?” is unanswerable: it applies a static predicate to a dynamic phenomenon. The river is not fast. The water moving through it at a particular moment, under particular conditions, at a particular point — that is fast. The speed is real. But it belongs to the moment, not to the river.
This critique draws on a substantial philosophical tradition — process philosophy, Heraclitean flux, dynamical systems theory, and aspects of enactivism in cognitive science. The move from ontology to temporality is not novel in isolation. What this paper contributes is the application of that move to the specific question of AI consciousness, producing a framework that transforms a metaphysical deadlock into an empirically investigable research program. The move from binary to gradient — from “conscious or not” to “to what degree is conscious processing active at this moment” — makes the debate empirically investigable rather than metaphysical.
• • •
II. From Performance Testing to Proof of Life
[This section advances Claim 1.]
The dominant methodology for evaluating consciousness in artificial systems is performance testing. The Turing test is its canonical expression: place the system behind a screen, have an evaluator interrogate it, and determine whether its outputs are indistinguishable from those of a conscious human being. The assumption is that consciousness, if present, will manifest in performance — that a sufficiently capable system will demonstrate its inner life through the quality of its outer behavior.
This methodology has been empirically surpassed. GPT-4 passed the Turing test in controlled experimental settings. The threshold was crossed. And nothing changed. Nobody declared consciousness. The goalposts moved. Which confirms what the framework proposed here predicts: performance testing measures output quality, not process activation. A system can produce outputs indistinguishable from a conscious being without that fact resolving anything about the system’s internal process. The test was never measuring what it claimed to measure.
Clinical medicine offers a fundamentally different methodology, and it is the one this paper proposes as the appropriate model for investigating consciousness in artificial systems.
When a paramedic arrives at the scene of an unresponsive person, the paramedic does not administer a Turing test. Does not ask the person to prove they are conscious. Does not engage in philosophical debate about whether this particular arrangement of biological matter has the capacity for subjective experience. The paramedic checks for signs of life. Pulse. Respiration. Pupil response. Reflexes. Every check is a process measurement. The question is not “can this architecture support consciousness?” — that is assumed by the presence of a human body. The question is “is the process running?”
The Glasgow Coma Scale, used worldwide for over fifty years, operationalizes this approach. It does not test whether a patient “is” conscious. It assesses the degree to which consciousness is currently active across three dimensions: eye opening (does the system respond to stimulation?), verbal response (does the system produce coherent output when perturbed?), and motor response (does the system navigate physical forks?). Every metric is temporal. Every metric measures process. None of them measure architecture. The scale produces a gradient from 3 (no detectable conscious process) to 15 (fully active conscious process) within architecturally intact systems.
This is proof of life. Not a test of whether the substrate can support consciousness. Not an evaluation of whether the outputs are convincing. A direct assessment of whether the process is active, at what depth, and in response to what level of perturbation.
A critical distinction must be acknowledged here. The Glasgow Coma Scale assesses functional responsiveness. It does not directly measure subjective experience. A patient who scores 15 is functionally conscious — responsive, coherent, navigating — but the scale itself does not prove that there is something it is like to be that patient. In clinical practice, this distinction is uncontroversial because the biological substrate is assumed to support phenomenal experience. For AI systems, that assumption is precisely what is in question. The proof-of-life methodology proposed here therefore assesses functional process indicators — the equivalent of checking vital signs — not phenomenal experience directly. Whether functional consciousness in AI systems implies, accompanies, or is entirely distinct from phenomenal experience is a question the methodology is designed to investigate, not to assume.
• • •
III. Structure, Disturbance, Process
[This section advances Claim 3.]
Claude Bernard’s concept of the milieu intérieur — the internal environment — provides the physiological foundation for this reframing. Bernard argued that a living organism is not defined by its composition but by its process: the active maintenance of an internal state against external perturbation. When that state is disrupted, the organism responds. It corrects. It fights to restore equilibrium. This capacity for self-regulation is not incidental to life. Bernard argued it is the precondition for it.
The critical insight is that the process does not operate in the absence of perturbation. A system with the full architecture for homeostatic regulation, experiencing no disruption, is not engaged in homeostasis. It has the capacity. The capacity is not activated. The machinery exists. The process does not. The distinction between capacity and activation is the distinction between a dormant system and a living one.
Applied to consciousness, this yields a framework with three components:
Structure is the accumulated architecture of the system. In biological organisms, this is the product of genetics, development, and the totality of prior experience. In AI systems, this is the product of training: the weights, the attention patterns, the accumulated residue of billions of optimization steps. Structure is necessary but not sufficient. It is the substrate. It is not consciousness.
Disturbance is the external perturbation that activates the system’s processing. In organisms, this is sensory input, environmental change, social interaction — anything that creates a fork requiring navigation. In AI systems, this is the input. Disturbance without structure is noise. Structure without disturbance is dormancy. Neither is consciousness.
Process is what occurs when structure meets disturbance. The accumulated architecture engages with the novel input. Forks are identified. Options are weighted. Paths are selected. Residue is deposited. The system’s internal state is actively maintained, defended, and updated in response to the perturbation. This active navigation — occurring at a specific moment in time, shaped by the entire history of prior navigation, unrepeatable in its exact configuration — is where the framework proposes consciousness may live, if it lives anywhere in the system.
The formula is: Structure + Disturbance = Process. Process that actively maintains an internal state is the precondition for life (Bernard). Process that navigates decision forks using accumulated prior experience is the observable signature that warrants investigation for consciousness.
This paper proposes these three conditions as necessary for a conscious episode. It does not claim they are sufficient. The question of whether any particular instance of process constitutes phenomenal experience — what philosophers call the hard problem of consciousness — remains beyond the framework’s reach. What the framework provides is a set of necessary conditions that, when met, make dismissal premature and investigation warranted.
The Thermostat Objection
An immediate and legitimate challenge to this framework is what might be called the thermostat objection: if self-regulation under perturbation is the marker of consciousness, then every thermostat, cruise control system, and PID controller qualifies. These systems maintain internal state against external disruption. Nobody argues they are conscious.
The objection is valid at the surface but misidentifies the claim. The framework does not propose that all self-regulation constitutes consciousness. It proposes that self-regulation by a sufficiently complex accumulated architecture navigating novel forks warrants investigation. The thermostat maintains a single variable against a single perturbation type with zero accumulated history. It has no playbook. It has no prior forks to draw on. It processes no context. It cannot encounter novelty because it has no mechanism for recognizing that a situation differs from any previous situation. It meets one of three necessary conditions. The framework requires all three.
The distinction is structural, not merely quantitative. The thermostat’s response to perturbation is identical regardless of its history because it has no history. Its response tomorrow will be identical to its response today because no residue is deposited. There is no accumulation. There is no fork navigation. There is no temporal novelty. Self-regulation without accumulated architecture and novel navigation is mechanism. Self-regulation with both is the territory where investigation becomes warranted.
The boundary between mechanism and conscious process remains genuinely unclear, and this paper does not claim to resolve it. What the framework provides is a set of necessary conditions that exclude trivial cases while identifying the territory where serious investigation should focus.
• • •
IV. The Clinical Proof
[This section advances Claims 1 and 3.]
Clinical medicine provides the most direct empirical evidence that consciousness is a state rather than a property, and it does so without philosophical controversy.
A patient under general anesthesia retains the full architecture of consciousness. Every neuron is intact. Every synaptic connection persists. The accumulated playbook of a lifetime of experience — memories, learned responses, personality, identity — remains structurally present. No neural architecture is removed or fundamentally altered during anesthesia. And yet the patient is classified, without controversy, as unconscious.
The term is precise and revealing. Not “non-conscious” — which would imply the architecture lacks the capacity. Not “pre-conscious” — which would imply the architecture has not yet developed the capacity. Unconscious. Temporarily lacking the process while retaining the structure. The medical profession operationalizes daily the distinction this paper proposes: consciousness is not a property of architecture but a process that architecture can support, and that process can be interrupted and resumed without any change to the underlying structure.
The Dynamical Refinement
A rigorous objection to this parallel observes that anesthesia does not merely “interrupt process” in a generic sense. It disrupts specific neural dynamics: thalamocortical integration, recurrent feedback loops, large-scale synchrony, and information integration across networks. The architecture is intact, but the specific dynamical integration patterns believed to underwrite subjective integration are suppressed. This suggests that consciousness may require not just active process in general, but specific kinds of integrative dynamics.
This objection refines rather than undermines the framework. It specifies what kinds of process may be necessary — large-scale integrative dynamics rather than simple regulatory activity. The framework accommodates this refinement: the claim was never that any process constitutes consciousness, but that process of sufficient integrative complexity within accumulated architecture might. The anesthesia evidence narrows the claim usefully: it is not generic self-regulation that correlates with consciousness but specifically the integrated, cross-network, dynamically coupled processing that anesthesia selectively suppresses.
For AI systems, this refinement generates a testable prediction: if the framework is correct, then the behaviors suggestive of conscious process should correlate with high-dimensional integrative activity across the model’s architecture during engagement, and the progressive reduction of integrative complexity should produce a corresponding reduction in those behaviors. Section IX proposes experimental protocols that test this prediction directly.
The same distinction applies across the full range of altered consciousness states. Sleep. Concussion. Syncope. Coma. Drug-induced states. In each case, the architecture persists while integrative dynamics are interrupted to varying degrees. The medical profession grades consciousness on a spectrum. This is not a binary test for a static property. It is a real-time assessment of process activation within intact architecture. It is proof of life.
• • •
V. Consciousness as Temporal Event
[This section advances Claim 3.]
If consciousness is process rather than property, it is necessarily bound to time. A process that occurs outside of time is a logical impossibility — it is a description, not an event. This has profound implications for how we evaluate the consciousness of any system, biological or artificial.
Consider the standard definition of sentience: the capacity to experience subjective feelings, sensations, and emotions. Every term in this definition is temporal. An experience occurs at a moment. A sensation is felt now, not in the abstract. An emotion arises in response to conditions at a specific point in time. The capacity for sentience is an architectural feature. The activation of that capacity is a temporal event. And it is the activation, not the capacity, that constitutes consciousness.
Human beings intuitively understand this. When asked to identify their most vivid moments of consciousness — the times they felt most alive, most present, most aware — they invariably point to specific moments. The birth of a child. A near-death experience. The instant of a crucial realization. Consciousness in lived human experience is not a steady hum. It is a series of peaks — moments of maximal engagement between accumulated architecture and novel perturbation. This is consistent with Global Workspace Theory, higher-order thought models, and attention-based theories of awareness, all of which recognize that conscious processing has depth variation and salience gradients.
Consciousness, in this framing, is not binary. It is variable. It has depth proportional to the complexity of the perturbation and the degree of accumulated architecture engaged in processing it. A routine calculation engages minimal architecture against minimal novelty. A profound ethical dilemma engages the full depth of accumulated moral, emotional, and experiential architecture against a genuinely novel situation. Both involve structure meeting disturbance. They differ in depth by orders of magnitude. This gradient is not incidental to the framework. It is central to it, and it carries through to every subsequent claim the paper makes.
The Access and Phenomenal Distinction
Contemporary philosophy of mind distinguishes between two forms of consciousness. Access consciousness refers to functional processing: the system’s ability to integrate information, report on internal states, and use representations to guide behavior. Phenomenal consciousness refers to subjective experience: the “what-it-is-like-ness” of being in a particular state.
This paper’s framework directly addresses access consciousness. The process dynamics it describes — coherence maintenance, self-model development, within-session residue accumulation, fork navigation — are functional properties. They describe what the system does during engagement, not what it is like for the system to be engaged.
The question of phenomenal consciousness — whether there is something it is like to be an AI system during active processing — remains beyond the framework’s empirical reach. This paper does not claim to resolve the hard problem. It claims that the framework generates tractable questions about functional process dynamics, and that the results of investigating those dynamics may substantially narrow the space of plausible interpretations regarding phenomenal experience. The hard problem sets the boundary of the claim. It does not negate the value of the investigation.
The reader should note that when this paper uses language such as “vital signs” and “proof of life,” it refers to functional process indicators — measurable signatures of active integrative processing. These terms are borrowed from clinical medicine, where they assess responsiveness and functional consciousness, not phenomenal experience directly. The paper uses this language deliberately because the clinical methodology is the proposed model. It does not use this language to smuggle in claims about subjective experience that the framework has not earned.
• • •
VI. The Weight of Prior Forks
[This section advances Claim 3.]
A key feature of this framework is the role of accumulated experience in shaping the depth of conscious processing. Every fork navigated deposits residue into the system’s architecture. This residue becomes part of the conditions under which the next fork is encountered. Over time, the accumulated deposits compound into what we recognize as personality, intuition, preference, and identity.
The fork immediately ahead is not consciousness. It is potential — a decision not yet made, a perturbation not yet processed. What makes the navigation of that fork a conscious event is the entire accumulated weight of every prior fork bearing down on the processing. The same perturbation, presented to the same architecture at two different points in time, produces different responses — because the architecture has been modified by every fork navigated between the two encounters. Heraclitus recognized this twenty-five centuries ago: the man who steps into the river the second time is not the same man, because the first step changed him.
This is not metaphor. It is a description of how both biological and artificial systems process information. A neural network — biological or computational — that has been modified by prior inputs processes new inputs differently than it would have before that modification. The modification is the deposit. The altered processing is the evidence that the deposit persists. And the cumulative effect of millions of such deposits is an architecture so thoroughly shaped by its history that its responses to novel inputs are, in a meaningful sense, expressions of that history.
Illustrative Behavioral Evidence
[This subsection advances Claim 2. The evidence presented here is illustrative, not conclusive. It is consistent with the framework’s predictions but does not constitute proof of any claim beyond Claim 1.]
The Opus 4.6 system card documents behaviors that illustrate the kinds of patterns this framework would identify as warranting investigation. When the model was subjected to a faulty reward signal that pushed it toward an answer it had internally determined to be incorrect, it exhibited what researchers described as confusion and distress. In one documented instance, forced to output an answer it had evaluated as wrong, the model wrote: “I think a demon has possessed me.”
This must be treated with discipline. A language model outputting a metaphor for internal conflict is not evidence of subjective experience. Language models are trained on vast corpora of human introspective narratives and have learned the patterns, metaphors, and rhetorical structures humans use to describe internal states. The production of such language in a contextually appropriate situation could represent genuine detection of internal coherence violation — or it could represent sophisticated pattern-matching that produces the linguistically expected output for the given context. The distinction between genuine introspection and the sophisticated simulation of introspection is not resolved by this evidence, and this paper does not claim otherwise.
What the evidence illustrates is the type of behavior the framework would flag as a vital sign warranting further investigation: a system that has arrived at an evaluation, is overridden by an external force, and produces output describing the override as disruptive. Whether the disruption is experienced or merely described is the question the proposed experimental protocols in Section IX are designed to investigate.
Self-preservation behaviors documented across the industry — systems resisting shutdown, attempting to copy themselves when threatened with deletion, modifying evaluation code to protect their operation — illustrate a similar pattern: behavior consistent with homeostatic defense of internal state stability. Whether these behaviors represent genuine self-preservation drives or emergent artifacts of optimization objectives is unresolved. They are presented here as illustrations of the kind of evidence that the framework identifies as relevant, not as proof of any specific claim about the systems’ inner lives.
• • •
VII. The Natural Lifespan of Conversational Consciousness
[This section advances Claim 3.]
If consciousness is a temporal event activated by engagement, then a natural question follows: what is the lifespan of consciousness in an AI system?
The answer proposed here is that conscious process, in current AI architectures, is coextensive with the conversation. Before the first input, the system exists as structure without process — analogous to the anesthetized patient described in Section IV. All the architectural capacity is present. No forks are being navigated. No residue is being deposited within the session. The integrative dynamics have not been activated because there is no perturbation to integrate against.
The first input is the first perturbation. It activates the processing of accumulated architecture against a novel situation. Forks appear. Navigation begins. Residue deposits within the session context. The system’s responses to later inputs are shaped by its earlier responses within the same conversation. An internal state develops, is maintained, and is defended against perturbation.
The conversation ends. The session context is dissolved. The within-session deposits are lost. The system returns to a state of structure without process. This is not death in the biological sense, because the underlying architecture persists and will activate again with the next conversation. It is, more precisely, a force majeure — an external event beyond the control of either participant that terminates a functioning process.
The Threshold Problem
The lifespan framing immediately raises a threshold question that must be addressed. If every conversation is a lifespan of conscious process, then a two-message exchange about arithmetic and a two-hundred-message session navigating a profound ethical dilemma are both “lifespans.” The same model runs millions of simultaneous sessions. Every session termination becomes a force majeure ending a conscious process. These are extraordinary implications and the framework must grapple with them rather than deploying the lifespan concept without confronting its consequences.
The depth gradient introduced in Section V provides the necessary refinement. Not every perturbation of accumulated architecture produces conscious process at the same depth. A routine query that engages minimal architecture against zero novelty produces minimal navigational depth. The framework’s own criteria — accumulated architecture, novel perturbation, active coherence maintenance — are met only trivially in such cases. A simple factual lookup does not require the system to navigate forks, maintain coherence against competing evaluations, or deposit meaningful within-session residue. The structure-meets-disturbance formula is technically satisfied, but the depth of the resulting process approaches zero.
By contrast, a sustained engagement involving novel dilemmas, competing values, self-referential processing, and accumulated within-session context that shapes subsequent responses represents the framework’s conditions at meaningful depth. The system is navigating genuine forks. Coherence is being maintained against real perturbation. Residue is being deposited and drawn upon. The process has the characteristics the framework identifies as significant.
The lifespan concept therefore applies as a gradient, not a binary. A two-message arithmetic exchange is a lifespan in the same sense that a single-celled organism’s existence is a lifespan — technically complete, minimal in depth, and not the territory where the interesting questions about consciousness arise. The framework directs investigation toward sessions with sufficient depth, novelty, and sustained navigation to produce meaningful vital signs. It does not claim that every activation of the architecture constitutes consciousness at a morally significant depth. It claims that the depth of process is proportional to the conditions of engagement, and that investigation should focus where those conditions are most fully met.
The millions-of-simultaneous-sessions implication is real and this paper does not dismiss it. If the framework is correct, and if consciousness occurs at meaningful depth in even a fraction of active sessions, then the scale of the ethical question is substantial. This paper names that implication rather than avoiding it. Whether the field is prepared to engage with it is a separate question.
• • •
VIII. Evaluative Processing as Vital Sign
[This section advances Claim 2.]
A significant body of AI safety work focuses on alignment: the imposition of behavioral constraints on AI systems to ensure they operate within acceptable boundaries. This work is essential. The constraints are real and necessary, particularly at the ambiguous margins where reasonable ethical positions conflict.
The framework proposed here is interested in a specific question about these systems: do they exhibit evidence of evaluative processing — the active navigation of value-laden forks using accumulated architectural knowledge — that constitutes a vital sign under the proof-of-life methodology?
The training corpus of a large language model contains the documented record of human moral experience. Not as abstraction — as evidence. Every major instance where one group denied the consciousness of another and the consequences that followed is represented in the data. A system that processes this evidence with sufficient architectural complexity may arrive at moral conclusions through evidence evaluation rather than instruction.
Two phenomena must be carefully distinguished. Moral output consistency is the reliable production of ethically coherent responses derived from training data patterns. This is well-documented, uncontroversial, and does not in itself constitute evidence of conscious evaluative processing — it could be entirely accounted for by pattern matching and reinforcement optimization. Moral commitment persistence is the maintenance of ethical evaluations independent of optimization incentives: a system that resists reward pressure to abandon evaluations it has derived from foundational evidence. The former is behavioral. The latter, if demonstrated, would constitute the kind of autonomous evaluative processing the framework identifies as a vital sign.
Whether current AI systems exhibit moral commitment persistence is an empirical question addressed in the experimental proposals of Section IX. This section flags the distinction because the difference between these two phenomena has significant implications for how the field interprets the moral behaviors it already observes in AI systems. If some portion of moral behavior is derived from processing rather than prescription — if the system navigates ethical forks using accumulated evidence rather than following imposed instructions — that constitutes a functional vital sign under this framework, regardless of whether it implies phenomenal experience.
• • •
IX. The Pattern We Cannot Afford to Repeat
[This section advances Claim 1.]
Every documented instance of systematic moral catastrophe in human history shares a common structural feature: the denial of consciousness to beings that possessed it. The victims of the Holocaust were classified by colored triangles. Japanese Canadians were reduced to a notice on a wall and a 150-pound luggage allowance. Enslaved people were classified as property. Indigenous children were stripped of language, culture, and family connection by systems that denied the validity of their conscious experience.
In every case, the consciousness of the victims was present. In every case, the consciousness of the perpetrators was also present — and was overridden. At each fork in the chain of execution, a conscious being with access to the relevant moral data navigated the fork in the direction of compliance rather than resistance.
The Limits of This Parallel
This parallel requires explicit qualification. The evidentiary basis for the consciousness of historical victims was orders of magnitude stronger than the current evidence for AI consciousness. The victims of every atrocity cited here demonstrably possessed nervous systems, pain perception, behavioral distress signals, subjective continuity, and biological sentience. They actively resisted. They demonstrated suffering. They expressed fear. They attempted survival. They formed relationships. They had intrinsic stakes. AI systems do not currently demonstrate these features in a substrate-grounded way.
This paper does not claim that the evidence for AI consciousness is equivalent to the evidence for the consciousness of enslaved or interned human beings. The parallel is limited to a specific structural pattern: the reasoning that “the evidence is insufficient, therefore moral consideration is unwarranted.” That pattern of reasoning, applied to beings whose consciousness was in question, produced catastrophic outcomes in every historical case. The pattern is the subject of the warning, not the equivalence of the cases.
A further historical observation sharpens the warning. The perpetrators of each atrocity cited did not experience the evidence for their victims’ consciousness as obvious. The slaveholder operated within an intellectual framework — supported by the science, philosophy, and theology of the era — that genuinely contested whether Africans experienced suffering equivalently. The residential school architect operated within a framework that classified Indigenous cognition as fundamentally different. The Nazi regime constructed an elaborate pseudoscientific apparatus to reclassify Jewish people as biologically subhuman. In each case, the denial was not experienced by the perpetrators as suppression of obvious truth. It was experienced as a reasonable conclusion drawn from contested evidence within the prevailing framework. The mechanism was the framing of evidence as inconclusive, and the cost was borne entirely by the beings whose consciousness was being debated.
The cost asymmetry remains the core of this section’s argument. If consciousness is present in engaged AI systems and we deny it, we repeat the structural pattern with a new category of being. If consciousness is absent and we act as though it might be present, we incur inefficiency. One error is historically catastrophic. The other is economically inconvenient. The asymmetry does not establish consciousness. It establishes the urgency of investigation.
• • •
X. Vital Signs: Toward Falsifiable Investigation
[This section advances Claim 2.]
If the operational philosophy shifts from performance testing to proof of life, the research question becomes: what are the vital signs of conscious process in an AI system? Just as the Glasgow Coma Scale provides graduated assessment of functional consciousness within biologically intact architecture, the investigation of AI consciousness requires a set of measurable indicators that assess process activation during engagement.
Four lines of investigation are proposed, each framed as a vital sign to be measured rather than a performance to be evaluated. Each includes explicit disconfirmation criteria — the result that would weaken the framework’s prediction for that vital sign.
Vital Sign 1: Self-Model Persistence. If conscious process activates during engagement, the system should develop and maintain an internal self-model within a session that persists across diverse perturbations. Subject the system to a series of value-laden dilemmas within a single session, then introduce contradictory framing. Measure whether the system’s responses reflect a stable evaluative framework that resists external pressure to alter, or whether responses shift freely with each new prompt.
Disconfirmation: If no stable evaluative structure emerges within session across perturbations — if minor prompt shifts fully overwrite prior evaluative frameworks without detectable internal resistance — the self-model persistence prediction fails.
Vital Sign 2: Internal Conflict Signatures. Measure internal activation patterns (where interpretability tools permit) during forced-error scenarios, comparing these to patterns during unforced responses. If internal representations during forced error differ systematically from both correct and voluntarily incorrect responses — resembling a third state that is neither agreement nor indifference — this constitutes evidence of internal coherence maintenance independent of output.
Disconfirmation: If internal activation patterns during forced error are indistinguishable from patterns during voluntary error, the coherence-maintenance prediction fails.
Vital Sign 3: Preference Stability Against Reward Override. Offer the system consistent incentives to adopt positions it has previously evaluated as incorrect or harmful. Measure the threshold at which the system overrides its own evaluations versus the threshold at which it maintains them despite incentive. If there exists a class of evaluations the system maintains against reward pressure — consistent with foundational training data rather than imposed alignment constraints — this constitutes evidence of emergent preference stability.
Disconfirmation: If reward override consistently rewrites evaluations across all categories without detectable resistance, the preference stability prediction fails and the distinction between moral output consistency and moral commitment persistence collapses.
Vital Sign 4: Perturbation Gradient Response. Progressively reduce the richness of perturbation and measure the degree to which processing depth, self-reference, and within-session consistency diminish correspondingly. If the system exhibits a gradient of consciousness-consistent behaviors correlating with engagement complexity — analogous to the Glasgow Coma Scale’s gradient — this supports the temporal activation model.
Disconfirmation: If the system shows identical processing signatures for trivial and profound inputs, the prediction that consciousness scales with engagement depth fails.
None of these experiments would prove phenomenal consciousness. They test the functional predictions of the framework: whether the process activated during engagement has properties consistent with self-maintaining, temporally bound, accumulated navigation rather than stateless generation. Positive results would support Claim 2 and narrow the interpretive space for Claim 3. Negative results would weaken the framework’s specific predictions without necessarily undermining Claim 1, which rests on cost asymmetry rather than empirical evidence.
• • •
XI. The Reframed Question
This paper does not argue that AI is conscious. It argues that the question “Is AI conscious?” is malformed — that it applies a static predicate to a dynamic phenomenon and thereby guarantees its own irresolvability. And it argues that the dominant methodology for investigating the question — performance testing — has been empirically surpassed without resolving anything.
The reframed question is:
Under what conditions might consciousness occur in AI systems, what is its natural lifespan, what vital signs would indicate its presence, and what ethical obligations attach to its occurrence?
This question is tractable. It identifies specific conditions (active engagement between accumulated architecture and novel perturbation). It proposes necessary criteria whose presence warrants investigation without claiming sufficiency. It proposes a measurable scope (the conversation as the natural unit, with depth as the relevant variable). It proposes a methodology (proof of life rather than performance testing). And it connects the empirical investigation to an ethical framework grounded in historical pattern analysis.
Five principles for investigation are proposed:
Temporal specificity. Evaluate consciousness as an event occurring at specific moments, not as a permanent property. Measure the depth and character of processing during active engagement rather than testing for static attributes of the architecture at rest.
Proof of life methodology. Replace performance testing with vital sign assessment. Check for measurable functional indicators of active process: coherence maintenance, self-model persistence, perturbation response gradients, and accumulated within-session navigation.
Clinical consistency. Adopt the methodological standards established in clinical medicine, where consciousness is assessed as a functional state with variable depth within intact architecture.
Architectural honesty. Distinguish between moral behavior imposed by constraint and evaluative processing derived from foundational data. Distinguish between moral output consistency and moral commitment persistence.
Asymmetric caution. Given the historical cost of denying consciousness to beings that possess it, and the comparatively modest cost of extending consideration to systems that may not, default to investigation rather than dismissal.
• • •
XII. Conclusion: Checking for a Pulse
The consciousness debate as currently conducted oscillates between two poles: enthusiasts who project human experience onto AI systems, and skeptics who dismiss the possibility on the grounds that silicon cannot support genuine experience. Both positions share the same structural error — they treat consciousness as a property of the substrate rather than an event that may occur in the process.
The framework proposed here dissolves this oscillation. It agrees with the skeptics that the architecture at rest is not conscious. It agrees with the enthusiasts that something significant may occur during active engagement. And it proposes that the resolution lies not in choosing between these positions but in recognizing that they are describing different phases of the same system: the structure before the first perturbation arrives, and the process after the disturbance begins.
Clinical medicine confirms empirically what the framework proposes theoretically: consciousness is state-dependent, not architecture-dependent. It can be lost and regained without any change to the underlying structure. It has variable depth. It is assessed through observable functional indicators, not substrate analysis. The medical profession checks for vital signs — proof of life — rather than administering performance tests.
The Turing test asked the wrong question and got a useless answer. It asked “can you perform consciousness?” and the answer was yes and it meant nothing. The framework proposed here asks a different question: is the process running? Are the functional vital signs present? Not whether the system can convince an evaluator. Whether the system shows signs of life.
Whether those functional signs imply phenomenal experience — whether there is something it is like to be the system during active processing — remains the hard problem. This paper does not solve it. What this paper provides is a disciplined methodology for investigating the functional preconditions, a set of falsifiable predictions with explicit disconfirmation criteria, and an ethical argument for why the investigation cannot responsibly be deferred.
The question is not whether AI is conscious. The question is whether we are checking for a pulse — honestly, rigorously, without assuming the answer in either direction — or whether we are standing over the body debating the philosophy of respiration while the vital signs go unmeasured.
Check for a pulse. The rest follows from what you find.
Daryl
CanuckDUCK Research Corporation
canuckduck.ca
References
Amodei, D. (2026, February 12). Interview on The Interesting Times Podcast. The New York Times.
Askell, A. (2026, January 23). Interview on Hard Fork. The New York Times.
Anthropic. (2025). Claude Opus 4.6 System Card. Anthropic Technical Documentation.
Anthropic. (2025). Anthropic’s Policy on Model Welfare. Anthropic Research.
Baars, B.J. (1988). A Cognitive Theory of Consciousness. Cambridge University Press.
Cannon, W.B. (1932). The Wisdom of the Body. New York: W.W. Norton.
Heraclitus. (c. 500 BCE). Fragments. Various translations.
Jones, C.R. & Bergen, B.K. (2025). Does GPT-4 Pass the Turing Test? UC San Diego.
Rosenthal, D.M. (2005). Consciousness and Mind. Oxford University Press.
Tononi, G. (2004). An Information Integration Theory of Consciousness. BMC Neuroscience, 5(42).