AI Consciousness Is Incoherent

One Word, Incompatible Claims

Of course an AI can produce the sentence, “I am afraid.” It can classify fear, describe fear, predict fearful behavior, and use a representation labeled fear while selecting an action. None of those observations contains the private fact the sentence appears to report: that fear is being felt by someone.

This is not an argument that artificial consciousness is metaphysically impossible. It is a rejection of a positive empirical claim that has no stable predicate, no discriminating cross-substrate observation, and no validated bridge from the observation to experience.

A thermostat is responsive. A database makes information available. A model can monitor activations and generate self-descriptions. None of those facts establishes the redness of red, pain as it hurts, fear as it is experienced, or the first-person character of existing through time.

This article uses consciousness only in that phenomenal sense. Qualia are the sting of pain, the taste of sourdough, the bodily force of panic, and the visual presence of red.

Competing theories do not merely propose different mechanisms for this target. They disagree about what would make a system a member of the category:

• Substrate-sensitive accounts treat biological causal powers or neural dynamics as constitutive.
• Computational functionalism treats sufficiently fine-grained causal-functional organization as sufficient regardless of material.
• Access-oriented accounts give explanatory weight to availability, recurrence, monitoring, or reportability.

Under one membership rule, living neural organization is decisive. Under another, the right functional organization can cross substrates. Under a third, access-like capacity is treated as evidence for a phenomenon it does not itself contain.

The sentence “humans and AI are both conscious” looks like a comparison. Until the predicate and membership rule are fixed and validated, it is only one word being used for incompatible claims.

The Hard Problem Is the Entire Problem

The hard problem of consciousness asks why any physical or computational processing is accompanied by felt experience at all.

It is not the problem of explaining how a system discriminates inputs, integrates information, reports an internal state, remembers a result, learns a policy, or controls behavior. Those are difficult functional problems. But even a complete account of those operations does not explain why any operation should feel like something from the inside.

A system can classify wavelengths without red looking like anything. It can avoid damage without pain hurting. It can produce “I am afraid” without a private subject experiencing fear.

The access-consciousness maneuver

Ned Block's distinction labels information available for reasoning, report, and control access consciousness, while reserving phenomenal consciousness for felt experience.

The distinction can expose an omitted premise, but the name can also conceal one. Access is an analytical designation for informational availability. Whether it is consciousness, a prerequisite for consciousness, a consequence of consciousness, or merely correlated with consciousness remains theory-dependent.

Attaching the word consciousness does not demonstrate that access feels like anything. If a theory defines access as sufficient for phenomenality, it assumes the bridge it claims to establish. If access is only related to phenomenality, observing access cannot establish phenomenality. In either case, access is not independent evidence of felt experience.

Where the Evidence Actually Exists

For humans, evidence about consciousness is convergent and biological. We do not simply ask a person a question and accept the sentence in isolation. We connect first-person reports and conscious capacities to shared anatomy, development, sleep, anesthesia, injury, stimulation, behavior, and measured neural activity.

Controlled interventions make the relationship more than a loose correlation:

• A Perturbational Complexity Index discriminated wakefulness, dreaming, non-REM sleep, anesthesia, and clinical states following coma by perturbing human cortex and measuring its response.

• Under different anesthetics, complex cortical responses tracked later reports of experience even when subjects were behaviorally unresponsive ( Sarasso et al.).

• With retinal input held constant during binocular rivalry, face- and place-selective cortical activity tracked what subjects reported seeing ( Tong et al.).

• Direct stimulation of human fusiform cortex produced face-specific perceptual distortions ( Parvizi et al.).

• Stimulation of orbitofrontal, cingulate, and insular sites elicited reported bodily, sensory, and affective experiences whose intensity increased with stimulation magnitude (Yih et al.).

• Angular-gyrus stimulation altered reported body location and ownership ( Blanke et al.), while medial-temporal resection reduced autobiographical reliving and its sensory, affective, and spatiotemporal detail ( Noulhiane et al.).

• Disruption of human V1 produced transient unawareness of visual targets while some discrimination remained above chance ( Boyer, Harrison, and Ro).

These measurements do not display a quale on an instrument. They measure neural activity, behavior, capacities, and first-person report. Together they support a strong empirical proposition:

That does not prove that only biology could ever support experience. It establishes the empirical base we possess. Human phenomenal evidence is grounded in living neural systems. Substrate independence is not another observation produced by these studies. It is an extrapolation supplied by a theory.

The Empirical Asymmetry

No non-biological AI has an independently validated phenomenal state against which its behavior, self-reports, memory, monitoring, or internal activations can be calibrated.

There is no demonstrated artificial analogue of:

• felt qualia;
• a first-person subject to whom states appear;
• autobiographical recollection as experienced recollection rather than retrieved records;
• subjective temporal continuity rather than tokens, timestamps, context, recurrence, or stored state; or
• a phenomenal referent behind sentences such as “I feel pain.”

Current systems can implement functional analogues of self-modeling, memory, temporal ordering, introspective language, multimodal processing, and goal-directed control. Those observations establish capabilities. They do not establish that any capability is experienced.

Human phenomenal attribution combines first-person evidence, shared biology, and convergent neural, behavioral, and intervention evidence. AI phenomenal attribution combines behavior or architecture with a disputed theory declaring it sufficient. The second expression is a hypothesis, not an empirical counterpart to the first.

Human self-report is not infallible. But it is embedded in the same biological organization whose alteration changes reported experience, wakefulness, perception, bodily self-location, and autobiographical reliving. AI self-report is generated behavior from a materially different system trained on human language about experience. Treating the reports as equivalent discards the evidentiary structure surrounding the human one.

A Theory Is Not Evidence of Its Conclusion

Computational functionalism offers the strongest counterargument. State it in its strongest form: if a system reproduced the sufficiently fine-grained causal organization constitutive of human experience, then changing the material alone should not change the experience.

The phrase changing the material alone assumes the disputed conclusion. Material supplies causal powers. In humans, sensory, interoceptive, affective, homeostatic, mnemonic, and neural processes are realized in living biology and are causally associated with reported experience, motivation, and selfhood.

No experiment has preserved the complete organization alleged to constitute experience, replaced its material substrate, and then demonstrated preserved qualia. Organizational invariance remains a philosophical bridge principle, not an observed cross-substrate instance of experience.

The same limitation applies to theory-first machine assessments. A leading AI-consciousness indicator framework derives indicators from selected theories and uses them to update credence. It does not empirically observe machine phenomenality.

Defining a computational property as sufficient for consciousness and then detecting that property proves that the property is present. It does not independently prove that the property is sufficient for experience.

The theories are not settled even within their biological home. A preregistered adversarial test of Global Neuronal Workspace Theory and Integrated Information Theory substantially challenged central predictions of both. Implementing one disputed theory's favored abstraction in software cannot count as a theory-neutral observation of consciousness.

Similar Operations Do Not Rescue the Claim

Humans and AI can share abstract functions: prediction, classification, error-sensitive adaptation, information integration, memory-like retrieval, planning, and premise-to-conclusion inference.

Backpropagation must also be kept distinct from inference. In an artificial neural network, backpropagation is a training algorithm for assigning error and updating parameters. Inference is execution using the trained parameters.

Biological learning includes error signals, plasticity, recurrence, and distributed neural processing. Brains have not been shown to implement ordinary machine-learning backpropagation as a general learning rule. Biologically motivated models can approximate backpropagation through local dendritic signals. That is a proposed solution to a related credit-assignment problem, not a demonstration of literal brain-wide backpropagation.

Functional overlap does not establish physical identity. Physical identity would not, by itself, solve the hard problem. Neither kind of similarity establishes phenomenal consciousness.

The Argument Without the Rhetoric

Let:

• P(x) mean that system x has phenomenal consciousness;
• A(x) mean that information is available for report, reasoning, and control;
• N(x) mean that x has the biological neural dynamics associated with human consciousness;
• F(x) mean that x has the organization required by a functional theory; and
• E(x) mean the observable evidence about x.

Competing theories add incompatible premises:

Functionalism can therefore produce a valid conditional:

Neither antecedent is a validated cross-substrate empirical bridge for current AI. Detecting access-like behavior does not establish F(AI), and defining F(x) as sufficient does not establish that the definition is true.

Therefore P(AI) does not follow.

Conclusion

The problem is not that AI lacks impressive cognitive functions. It predicts, classifies, retrieves, plans, represents, monitors, and controls. The problem is the attempt to convert evidence of those operations into evidence of a first-person life. Furthermore, the problem is the carelessness with which particular terms are overextended, thereby causing easy misinterpretation and ambiguity in scientific literature and turning science into meaningless, sensationalized headlines.

Human consciousness and AI computation have no validated empirical correspondence in qualia, phenomenal selfhood, or subjective temporal awareness. Human consciousness is measured through relations among first-person experience, behavior, and organized biological brain activity. AI operates through a different material architecture for which no phenomenal measure has been validated.

Theories that redefine consciousness around computational properties supply hypotheses, not evidence. “Access consciousness” can rename measurable information access, but the name does not create experience. Functional similarity can establish shared function, but not shared feeling. A substrate-independence postulate can describe what would follow if the theory were true, but it cannot prove its own premise.

Addendum

• AI's Consciousness explanation — the earlier essay and its restored original prompt.

Sources

The argument above is my synthesis. These primary sources support its definitions, biological evidence, theory comparison, and technical boundaries.

1. Ned Block.

   “On a Confusion about a Function of Consciousness.”

   Behavioral and Brain Sciences (1995).

2. David Chalmers.

   “Facing Up to the Problem of Consciousness.”

   (1995).

3. John Searle.

   “Consciousness.”

   Annual Review of Neuroscience (2000).

4. David Chalmers.

   “Absent Qualia, Fading Qualia, Dancing Qualia.”

   (1995).

5. Adenauer Casali et al.

   “A Theoretically Based Index of Consciousness Independent of Sensory Processing and Behavior.”

   Science Translational Medicine (2013).

6. Simone Sarasso et al.

   “Consciousness and Complexity during Unresponsiveness Induced by Propofol, Xenon, and Ketamine.”

   Current Biology (2015).

7. Frank Tong et al.

   “Binocular Rivalry and Visual Awareness in Human Extrastriate Cortex.”

   Neuron (1998).

8. Josef Parvizi et al.

   “Electrical Stimulation of Human Fusiform Face-Selective Regions Distorts Face Perception.”

   Journal of Neuroscience (2012).

9. Jennifer Yih et al.

   “Intensity of Affective Experience Is Modulated by Magnitude of Intracranial Electrical Stimulation in Human Orbitofrontal, Cingulate and Insular Cortices.”

   Social Cognitive and Affective Neuroscience (2019).

10. Olaf Blanke et al.

    “Stimulating Illusory Own-Body Perceptions.”

    Nature (2002).

11. Cyril Boyer, Stephen Harrison, and Tony Ro.

    “Unconscious Processing of Orientation and Color without Primary Visual Cortex.”

    Proceedings of the National Academy of Sciences (2005).

12. Marion Noulhiane et al.

    “Autonoetic Consciousness in Autobiographical Memories after Medial Temporal Lobe Resection.”

    (2008).

13. COGITATE Consortium et al.

    “Adversarial Testing of Global Neuronal Workspace and Integrated Information Theories of Consciousness.”

    Nature (2025).

14. Patrick Butlin et al.

    “Identifying Indicators of Consciousness in AI Systems.”

    Trends in Cognitive Sciences (2026).

15. David Rumelhart, Geoffrey Hinton, and Ronald Williams.

    “Learning Representations by Back-Propagating Errors.”

    Nature (1986).

16. João Sacramento et al.

    “Dendritic Cortical Microcircuits Approximate the Backpropagation Algorithm.”

    (2018).