Artificial intelligence, human cognition, and conscious supremacy

The paper addresses the question: what is the computational significance of consciousness given recent advances in AI, especially LLMs that perform many human-like tasks without presumed consciousness? The author frames consciousness in terms of phenomenal properties (qualia, intentionality, self-awareness) and reviews how tasks such as theory of mind, the Turing test, and the Winograd schema relate to consciousness and language. Despite language typically being processed during conscious states in humans, LLMs appear to achieve strong linguistic performance without consciousness, prompting inquiry into which computations, if any, require consciousness. The introduction situates this inquiry within established consciousness theories (NCC, Global Neuronal Workspace, Integrated Information Theory, Free-Energy Principle) and distinguishes computational correlates of consciousness. It proposes shifting focus from the hard problem to computational roles, offering preliminary criteria for conscious vs. unconscious computation and introducing the notion of "conscious supremacy"—domains where conscious processing achieves practically efficient computation that unconscious systems cannot within biologically relevant time and resource limits.

The paper surveys multiple theoretical frameworks and findings relevant to consciousness and AI: (1) Classical and contemporary theories of consciousness and their correlates, including Neural Correlates of Consciousness (Crick & Koch), Global Workspace/GNW (Baars; Dehaene et al.), Integrated Information Theory (Tononi et al.), and the Free-Energy Principle and Computational Correlates of Consciousness (Friston; Wiese & Friston). (2) Debates on quantum models of consciousness (Penrose; Hameroff & Penrose; critiques by Tegmark), noting controversy and lack of consensus. (3) Language and consciousness scholarship (Velmans; role of attention, speech production complexity, and language’s dual-aspect modeling of experience). (4) AI capabilities and limits: transformer-based LLMs and emergent abilities, hallucination and confidence/metacognition issues, representation debates (Marr; constructivist views), behaviorist and instrumental convergence perspectives, and metacognition’s role in monitoring and control (Nelson; Yeung & Summerfield). (5) AI benchmarks (Turing test, theory of mind, Winograd schema), game-playing systems (Deep Blue, AlphaZero), and AGI task proposals. (6) Visual perception and binding problem literature (Treisman; Feldman; Zeki; Logothetis) as analogies to computational complexity (e.g., Shor’s algorithm). The review motivates a conceptual space where some computations may be uniquely efficient under conscious processing compared to unconscious/AI systems.

Conceptual and theoretical analysis combined with a narrative review. The author: (1) Defines working distinctions between conscious and unconscious computation (objective neural integration and subjective phenomenology); (2) Introduces the concept of conscious supremacy by analogy to quantum supremacy, focusing on practical (biological time/resource) rather than theoretical computability limits; (3) Uses set-theoretic framing to delimit uniquely human/conscious computations as the complement of growing AI-capable sets; (4) Surveys literature across consciousness science, AI, decision-making, and visual perception to support hypotheses; (5) Employs analogical reasoning (e.g., quantum error correction vs. proposed conscious error correction) and domain case discussions (language, driving, embodied interaction) to identify candidate areas requiring conscious processing. No empirical experiments or quantitative analyses are conducted; arguments are based on theory synthesis and conceptual models.

- Proposal and definition of conscious supremacy: domains of computation executed efficiently by conscious processes but infeasible for non-conscious systems within practical biological time and resources, despite being computable in principle. - Identification of candidate cognitive domains potentially unique or especially advantaged under consciousness: flexible attention modulation; robust handling of novel contexts; ad hoc choice and decision-making (including metacognitive veto or “free won’t”); integrated multimodal cognition across distributed brain regions; and embodied cognition in situated environments. - Language is characterized as a sequence of micro-decisions; while LLMs can reproduce knowledge and pass limited tests, they may lack situated, embodied language use and metacognitive confidence calibration, and exhibit hallucinations. - In perception, the binding problem and combinatorial explosion suggest a need for mechanisms that integrate distributed features into coherent percepts; an analogy is drawn to quantum factoring (Shor’s algorithm). The paper hypothesizes a conscious error correction (CEC) process, analogous in spirit to quantum error correction, to reconcile noisy neural activity with stable qualia, though details remain speculative. - Practical implications for AI: certain tasks like fully autonomous driving in open, ambiguous, morally fraught contexts may require conscious-level integration and decision-making, potentially limiting full automation without human oversight. - Alignment strategy: divide labor by letting AI handle computations not requiring consciousness, and reserve conscious-supremacy tasks for humans; RLHF is interpreted as operationalizing such division by incorporating human conscious judgments. - The paper recommends starting from the null hypothesis that current AI lacks consciousness, enabling clearer delineation of what consciousness uniquely computes.

The paper argues that recent AI advances intensify the need to clarify what computations consciousness enables that current non-conscious systems lack, especially within practical constraints. By defining conscious supremacy, the author offers a framework to separate tasks well-suited to unconscious/AI computation from those that may require conscious, globally integrated processing and metacognition. This framework addresses the research question by suggesting that consciousness contributes unique efficiency in integrating multimodal information, handling novelty, making value-laden choices, and operating in embodied, situated contexts—areas where AI systems struggle with uncertainty, hallucination, and lack of self-monitoring. Analogies to quantum computing highlight the possibility that consciousness confers practical computational advantages without implying non-computability. For AI alignment, the proposed division of labor—augmenting human conscious computation rather than replacing it—could improve safety and performance, mirroring the interplay of conscious and unconscious processes in the human brain.

The paper introduces and motivates the concept of conscious supremacy as a way to focus on the computational roles of consciousness distinct from the hard problem. It suggests that consciousness may uniquely enable efficient, integrated processing for tasks involving flexible attention, novel-context decision-making, multimodal integration, and embodied interaction. The author recommends treating current AI as non-conscious by default, using human conscious oversight to address limitations such as hallucinations and value-sensitive judgments, and adopting alignment strategies that partition tasks according to whether conscious processing is advantageous. Future work should specify and empirically test concrete instances of conscious-supremacy computations, clarify mechanisms such as the hypothesized conscious error correction, and develop benchmarks that distinguish conscious from non-conscious computational performance in realistic settings.

The arguments are speculative and conceptual, lacking empirical demonstrations of specific computations uniquely requiring consciousness. No concrete algorithmic or neuroscientific mechanism for conscious supremacy or conscious error correction is established. The delineation of tasks requiring consciousness versus those feasible for AI is not formally specified and may shift as AI capabilities evolve. The paper acknowledges it does not yet identify definitive computations unique to consciousness, limiting generalizability and testability at present.