Artificial intelligence, human cognition, and conscious supremacy

Reviews, theory, and governance2024Frontiers (Psychology)Approved editorial review

Original title: Artificial Intelligence, Human Cognition, and Conscious Supremacy

Authors: Ken Mogi

Keywords: consciousness, artificial intelligence, cognitive science, conscious supremacy, computational significance, philosophy of mind, neural correlates

Source: Open primary source (opens in a new tab)

1
Authors
15
Findings
34
Limitations
18
Evidence

Editorial summary

English

This Hypothesis and Theory article asks a conceptual question: if presumably non-conscious AI systems already perform linguistic and cognitive tasks once associated with consciousness, is there any computation that human consciousness can execute within practical biological limits but a non-conscious system cannot? Ken Mogi calls that possible advantage “conscious supremacy,” by analogy with quantum supremacy. The paper contains no experiment, systematic review, executable computational model, or mathematical proof. It is an eight-page theoretical essay linking selected literature on consciousness, cognition, LLMs, quantum computing, and AI alignment.

The proposed definition has two related forms. First, conscious supremacy denotes computational domains that conscious processes can perform but systems lacking consciousness cannot execute in practical time. Second, among computations in the brain’s neural networks, it denotes computations accompanied by consciousness that are performed efficiently and integratively and that unconscious processing could not finish in biologically meaningful time, even if possible in principle. The paper also defines a set X of computations “unique to humans” as the complement, within human computation, of the union of tasks performed by successive AI systems; X would shrink as more systems appear. The text itself acknowledges that X does not map cleanly onto consciousness: it may include unconscious human computation and mixed conscious-unconscious processing, and no clear line currently separates what AI can and cannot do.

Candidate domains include flexible attention modulation, acquiring skills and deciding in novel contexts, multimodal integration, ad hoc judgment, situated choice, embodied cognition, and metacognition. These are contrasted with well-learned skills that humans may execute automatically and unconsciously. LLMs are presented as an example of language processing that could occur without consciousness. No model is evaluated, however, and no task is specified that measurably separates a conscious from a non-conscious system. ChatGPT, GPT-4, AlphaZero, and autonomous-driving systems appear only as examples from the literature. DALL-E is used solely to create the yellow-Volkswagen illustration in Figure 2; the image is not evidence.

The quantum analogy is extended to visual binding: integrating color and form into a conscious percept is compared with combinatorial explosion and, illustratively, with factorization accelerated by Shor’s algorithm. The author also coins “conscious error correction” to suggest that conscious experience could rectify neural noise, by analogy with quantum error correction. The paper explicitly labels this relationship speculative and supplies no mechanism, equations, predictions, or data. It also states Shor’s complexity as polynomial “in N”; the standard statement is polynomial in the bit length, log N. More fundamentally, quantum advantage has formalized problems, machines, and resource bounds, whereas conscious supremacy has no candidate task, metric, resource budget, or observable criterion for consciousness.

For alignment, the paper proposes a division of labor: reserve hypothetically consciousness-specific computation for humans and use AI to augment rather than replace it. It interprets RLHF as an interaction between conscious human evaluation and automated optimization, and argues that artificial consciousness would not necessarily improve alignment. These are conceptual recommendations, not safety results. The author explicitly notes that limiting AI to non-conscious operations does not guarantee robust alignment and that developers, stakeholders, and other ecosystem parameters also matter.

The most faithful conclusion is provisional and negative: the article offers vocabulary for searching for a practical advantage of conscious processing, but it has not identified any computation unique to consciousness. It also does not solve how to establish that a comparison system lacks consciousness, which is required by the definition. Treating current AI as non-conscious under a null hypothesis may be a methodological stance, but it is not a finding of the paper. Claims about LLMs, autonomous driving, metacognition, binding, evolution, and alignment are arguments based on selected references without a search protocol or systematic evidence assessment. The contribution is therefore a philosophical-computational hypothesis intended to generate questions, not evidence of human conscious superiority, LLM unconsciousness, or a demonstrated boundary between human and artificial computation.

Español

Este artículo de tipo “Hypothesis and Theory” plantea una pregunta conceptual: si sistemas de IA presumiblemente no conscientes ya realizan tareas lingüísticas y cognitivas antes asociadas a la consciencia, ¿queda algún tipo de cómputo que la consciencia humana pueda ejecutar dentro de límites biológicos prácticos y un sistema no consciente no? Ken Mogi denomina a esa posible ventaja “supremacía consciente”, por analogía con la supremacía cuántica. No presenta un experimento, una revisión sistemática, un modelo computacional ejecutable ni una demostración matemática. Es un ensayo teórico de ocho páginas que organiza bibliografía de consciencia, cognición, LLM, computación cuántica y alineación de IA.

La definición propuesta tiene dos formulaciones relacionadas. Primero, serían dominios computacionales que procesos conscientes pueden realizar, pero sistemas sin consciencia no pueden ejecutar en un tiempo práctico. Segundo, dentro de las redes neuronales del cerebro serían los cómputos acompañados por consciencia que se realizan de modo eficiente e integrado y que un procesamiento inconsciente no podría completar en un tiempo biológicamente significativo, aunque sí en principio. El artículo también define un conjunto X de tareas “únicas de los humanos” como el complemento, dentro del espacio de cómputos humanos, de la unión de tareas realizadas por sucesivos sistemas de IA. A medida que aparecen más sistemas, X se reduciría. El propio texto reconoce que X no equivale limpiamente a consciencia: puede incluir cómputos humanos inconscientes y combinaciones de procesamiento consciente e inconsciente, y no existe una línea clara entre lo que la IA puede o no puede hacer.

Como candidatos, el autor enumera la modulación flexible de la atención, la adquisición de habilidades y decisiones en contextos nuevos, la integración de información multimodal, juicios ad hoc, elección situada, cognición corporizada y metacognición. Contrasta estas funciones con habilidades ya adquiridas, cuya ejecución humana puede automatizarse inconscientemente. Los LLM se presentan como ejemplo de procesamiento lingüístico que podría ocurrir sin consciencia. Sin embargo, no se evalúa ningún modelo ni se especifica una tarea que separe de forma medible un sistema consciente de uno no consciente. ChatGPT, GPT-4, AlphaZero y sistemas de conducción aparecen como ejemplos tomados de la literatura. DALL-E solo se usa para crear la ilustración de un Volkswagen amarillo en la Figura 2; esa imagen no constituye evidencia.

La analogía cuántica se extiende al problema de binding visual: integrar color y forma en un percepto consciente se compara con una explosión combinatoria y, de manera ilustrativa, con la factorización acelerada por el algoritmo de Shor. El autor acuña además “conscious error correction” para sugerir que la experiencia consciente podría rectificar el ruido neuronal, en analogía con la corrección de errores cuántica. El texto califica expresamente esta relación como especulativa y no ofrece mecanismo, ecuaciones, predicciones o datos. Hay además una imprecisión técnica: describe Shor como polinómico “en N”; la complejidad estándar se expresa como polinómica en la longitud en bits, log N. En cualquier caso, la ventaja cuántica dispone de problemas, máquinas y recursos formalizados, mientras la supremacía consciente carece todavía de tarea candidata, métrica, presupuesto de recursos y criterio observable de consciencia.

En alineación, el artículo propone una división del trabajo: reservar para humanos los hipotéticos cómputos exclusivamente conscientes y usar la IA para aumentarlos en lugar de sustituirlos. Interpreta RLHF como un ejemplo de interacción entre evaluación humana consciente y optimización automática, y sugiere que crear consciencia artificial no sería necesariamente una buena estrategia de alineación. Son recomendaciones conceptuales, no resultados de seguridad. El propio autor advierte que limitar la IA a operaciones no conscientes no garantiza una alineación robusta y que también importan desarrolladores, stakeholders y otros parámetros del ecosistema.

La conclusión más fiel es negativa y provisional: el artículo propone un vocabulario para buscar una posible ventaja práctica del procesamiento consciente, pero todavía no identifica ninguna computación exclusiva de la consciencia. Tampoco resuelve cómo determinar que un sistema carece de consciencia, condición necesaria para comparar ambos lados de la definición. Adoptar como hipótesis nula que la IA actual no es consciente puede ser una postura metodológica razonable, pero no es un hallazgo del paper. Las afirmaciones sobre LLM, conducción autónoma, metacognición, binding, evolución y alineación son argumentos apoyados en referencias seleccionadas, sin protocolo de búsqueda ni evaluación sistemática de evidencia. Por tanto, la contribución es una hipótesis filosófico-computacional generadora de preguntas, no evidencia de consciencia humana superior, de inconsciencia de los LLM ni de una frontera demostrada entre cómputo humano y artificial.

Research question

Can a practical computational advantage of conscious processes over non-conscious systems be defined and discovered, and what would that possible "conscious supremacy" imply for the division of labor and alignment between humans and AI?

Method

Narrative theoretical essay. The author combines selected literature on consciousness, conscious and unconscious processes, LLMs, AGI, quantum computing, visual binding, metacognition, and alignment; proposes a definition by practical limits, a residual set X with respect to AI capabilities, and analogies with quantum supremacy and error correction.

Sample: There are no participants, model runs, observations, benchmarks, or defined documentary sample. The reference corpus is not selected through a search strategy, inclusion criteria, or published flow diagram.

Findings

  • The article proposes the term conscious supremacy by analogy with quantum supremacy.
  • The proposal refers to a practical advantage of time or resources, not to incomputability in principle.
  • It adopts as a null hypothesis that current AI systems are not conscious.
  • It uses the advance of AI to progressively reduce the set of tasks considered exclusively human.
  • It recognizes that exclusively human tasks and conscious tasks are not equivalent sets.
  • It suggests flexible attention, novelty, decision, multimodal integration, embodiment, and metacognition as candidates.
  • It contrasts the conscious acquisition of a skill with its subsequent more automatic execution.
  • It considers that the linguistic processing of LLMs could be carried out without consciousness.
  • It proposes an analogy between visual binding and combinatorial explosion problems.
  • It introduces conscious error correction as a hypothetical mechanism.
  • It proposes reserving for humans the hypothetical exclusively conscious computations.
  • It presents RLHF as an example of interaction between human evaluation and automatic computation.
  • It argues that creating artificial consciousness would not guarantee alignment.
  • It recognizes that limiting AI to non-conscious operations also does not guarantee alignment.
  • It explicitly concludes that it has not yet identified computations exclusive to consciousness.

Limitations

  • It is a hypothesis article, not an empirical study.
  • There is no systematic review or bibliographic search protocol.
  • There are no inclusion, exclusion, or reference quality evaluation criteria.
  • No LLM, AI system, or human participant is tested.
  • No executable benchmark of conscious supremacy is proposed.
  • No concrete candidate task is defined.
  • No performance metric is specified.
  • No quantitative limits of time, energy, memory, or biological resources are set.
  • It is not defined how to verify that the comparator system lacks consciousness.
  • The non-conscious condition is difficult to observe without an accepted theory or test of consciousness.
  • The definition attributes efficiency and integration to conscious computation in advance.
  • The set X captures tasks not performed by AI, not tasks caused by consciousness.
  • X may include unconscious human processes, as the article itself recognizes.
  • X depends on the contingent set of AI systems considered.
  • The boundary shifts with new models and does not identify a stable property.
  • The analogy with quantum supremacy lacks a comparable formalism.
  • The comparison between binding and factorization is illustrative, not a computational reduction.
  • Conscious error correction has no mechanism, model, or testable prediction.
  • The complexity of Shor is imprecisely expressed as polynomial in N rather than in log N.
  • Claims about LLM capabilities depend on discussed literature and general examples.
  • It is not specified which version of ChatGPT or GPT-4 supports the claims.
  • Passing a behavioral test does not establish consciousness, but the argument borders on that inference.
  • Theory of mind in LLMs is recognized as controversial.
  • The impossibility of fully autonomous driving without conscious intervention is an unproven extrapolation.
  • There is no evidence that consciousness is necessary for novel decisions or multimodal integration.
  • There is no causal comparison between conscious and unconscious processing.
  • Consciousness is not separated from attention, reportability, metacognition, intelligence, or embodiment.
  • The evolutionary hypothesis about adaptive value is not modeled or tested.
  • The alignment implications are not evaluated with risk or safety analysis.
  • The human-AI division of labor offers no operative criteria for assigning tasks.
  • RLHF does not demonstrate a division between conscious and unconscious computation.
  • The DALL-E image in Figure 2 is decorative and does not validate the analogy.
  • There is no software, data, substantive supplementary material, or reproduction artifacts.
  • The contribution statement lists data curation and software even though the article presents no dataset or software.

What the study does not establish

  • It does not demonstrate that conscious supremacy exists.
  • It does not identify any computation exclusive to consciousness.
  • It does not demonstrate that LLMs are unconscious.
  • It does not demonstrate that LLMs are conscious.
  • It does not demonstrate that intelligence and consciousness are dissociated in current AI.
  • It does not demonstrate that novel tasks require consciousness.
  • It does not demonstrate that visual binding obtains an advantage analogous to the quantum one.
  • It does not demonstrate the existence of conscious error correction.
  • It does not demonstrate that qualia have an error correction function.
  • It does not provide a test of consciousness for humans, animals, or machines.
  • It does not establish that passing Turing, Winograd, or theory of mind is sufficient for consciousness.
  • It does not establish a computational limit of classical or unconscious systems.
  • It does not demonstrate that humans outperform all AI on a defined task.
  • It does not demonstrate that keeping tasks in human hands improves safety.
  • It does not demonstrate that artificial consciousness harms or benefits alignment.
  • It does not demonstrate that RLHF is a robust alignment mechanism.
  • It does not solve the hard problem or the real problem of consciousness.
  • It does not provide sufficient falsifiable predictions to validate the hypothesis.

Traceability

Scope: Full text

Version: Frontiers in Psychology 15:1364714, Hypothesis and Theory, version of record published 13 May 2024, DOI 10.3389/fpsyg.2024.1364714

Consulted source: https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2024.1364714/pdf

Review: Codex full-text, bilingual-fidelity, visual, hypothesis-theory, consciousness-definition, computational-complexity, quantum-analogy, LLM-claim, falsifiability, alignment-implication and evidence-level audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • No AI model is experimentally evaluated
  • ChatGPT and large language models discussed generically
  • GPT-4 discussed through cited literature
  • AlphaZero and other AI systems used as conceptual examples
  • DALL-E used only to generate the illustrative image in Figure 2

Instruments and metrics

  • Narrative theoretical argument
  • Conscious-supremacy definition based on practical time and resource limits
  • Set-complement sketch X relative to accumulated AI capabilities
  • Analogy with quantum supremacy
  • Analogy between visual binding and factorization
  • Hypothetical conscious error correction
  • Human-AI division-of-labor argument for alignment

Data used

  • No empirical dataset
  • No model outputs or benchmark results
  • Selected published literature
  • One DALL-E-generated illustrative image

Evidence and location

  • Publication and article type: Frontiers in Psychology 15:1364714, Hypothesis and Theory, published 13 May 2024, DOI 10.3389/fpsyg.2024.1364714
  • Complete source: .cache/editorial-sources/article-110/source.pdf; 8 pages; sha256 853aacd21c6cb8cc133523a20d649ce23086aaf94e85550e588839e3af972280
  • Author and current editorial status: Frontiers article page and Crossmark entry; Ken Mogi; checked 15 Jul 2026
  • Official abstract: Full text p. 1
  • Null hypothesis of non-conscious AI: Full text pp. 2 and 6
  • Definition of conscious supremacy: Full text p. 3
  • Residual set X: Full text p. 3, Figure 1
  • Recognition of conceptual nature: Full text p. 3
  • Candidates for conscious computation: Full text pp. 4–5, Sections 2–3
  • Driving, situated choice, and language: Full text p. 5
  • Binding-factorization analogy: Full text p. 5, Figure 2
  • Use of DALL-E: Full text p. 5, Figure 2 caption and disclosed prompt
  • Speculative conscious error correction: Full text p. 5, Section 4
  • Division of labor and alignment: Full text pp. 5–6, Section 5
  • Alignment not guaranteed: Full text p. 6
  • Absence of identified exclusive computation: Full text p. 6, Discussion: has not yet specifically identified computations unique to consciousness
  • Data and materials: Full text p. 6, Data availability: contributions are contained in the article
  • Visual inspection: All 8 PDF pages rendered and visually inspected, including Figures 1–2 and all six sections; checked 15 Jul 2026