Artificial intelligence, human cognition, and conscious supremacy

The paper examines the computational significance of consciousness in light of recent advances in large language models (LLMs) that perform many tasks traditionally associated with human cognition. It questions whether consciousness is necessary for cognitive processes such as perception, language, creativity, and decision-making, given that AI systems may pass tests linked to consciousness (e.g., theory of mind, Turing test, Winograd schema) without being conscious. It reviews foundational perspectives, including neural correlates of consciousness, global workspace theory, integrated information theory, and the free-energy principle. The author proposes the concept of "conscious supremacy"—domains of computation that conscious processes can perform efficiently and integratively within biological time and resource limits, whereas non-conscious systems could only perform them in principle but not practically. This framework aims to separate computations requiring consciousness from those executable unconsciously, using examples in perception, language, and driving.

The paper surveys multiple theoretical approaches to consciousness and cognition: Velmans’ analyses of consciousness in information processing and speech production; neural correlates of consciousness (Crick & Koch) and their progress and limits; global neuronal workspace theory (Dehaene et al.; Mashour et al.) and its multimodal integration; integrated information theory (Tononi et al.); and the free-energy principle (Friston), including Wiese & Friston’s proposal of computational correlates of consciousness and emphasis on trajectories over states. It discusses debates on quantum theories of consciousness (Penrose, Hameroff; Orch OR) and critiques regarding decoherence at biological temperatures (Tegmark). It references AI benchmarks and abilities (Turing test, theory of mind tasks, Winograd schema challenge; Deep Blue, AlphaZero; GPT-4), problems like hallucination and confidence calibration (Ji et al.; Yeung & Summerfield), metacognition (Nelson; higher-order theories), instrumental convergence (Bostrom), vetoing actions (Libet’s free won’t), and issues in self-driving and moral decision-making (Badue et al.; Kosuru & Venkitaraman; Awad et al.; Shladover). The lexical hypothesis (Crowne) is cited to explain broad representational capacity in language aligning with LLM performance.

This is a hypothesis and theory paper employing conceptual and comparative analysis rather than empirical experimentation. The author: (1) defines and motivates the concept of conscious supremacy by analogy to quantum supremacy, focusing on practical computational limits rather than computability in principle; (2) analyzes domains of human cognition typically associated with conscious processing (e.g., flexible attention, novel-context decision-making, multimodal integration, embodied cognition) versus skills executed largely unconsciously after acquisition; (3) compares capabilities and limitations of contemporary AI (especially LLMs and deep reinforcement learning systems) to human conscious cognition, using thought experiments and real-world tasks (language use in situated contexts, driving with trolley-like dilemmas); (4) introduces a set-theoretic framing of uniquely human computations as the complement of the union of AI-executable computations, refining the residual space as AI advances; (5) develops analogies to quantum computing (e.g., Shor’s factoring vs. combinatorial binding in visual feature integration) and proposes a tentative mechanism of conscious error correction (CEC), conceptually paralleling quantum error correction (QEC).

• Conscious supremacy is introduced as a tentative framework: certain computations are efficiently and integratively executable by conscious processes within biological resource and time constraints, while non-conscious systems could only perform them in principle but not feasibly. • Conscious cognitive domains potentially unique or central to consciousness include: flexible attention modulation; robust handling of novel contexts; choice and decision-making under uncertainty and time pressure; integrated multimodal cognition across widespread neural networks; and embodied cognition coupling perception, action, and social context. • Human-acquired skills often operate unconsciously; in contrast, tasks requiring flexible, ad hoc judgments likely involve conscious processing, explaining why some AI successes in language and games need not imply consciousness. • There may be dissociations between intelligence and consciousness: consciousness without intelligence in some biological states, and AI intelligence without consciousness. • An analogy is drawn between the visual binding problem’s combinatorial explosion and factoring in Shor’s algorithm, motivating exploration of efficient integration mechanisms in conscious perception. • A speculative conscious error correction (CEC) mechanism is proposed, addressing how noisy neural firings yield stable, Platonic-like qualia; relationships to QEC are conjectural. • For AI alignment and safety, a division of labor is suggested: leave computations belonging to conscious supremacy to humans while designing AI to augment non-conscious computations; techniques like RLHF can instantiate such collaboration.

The paper discusses how a computational framing helps address the question of what consciousness contributes beyond what advanced AI can do unconsciously. By positing conscious supremacy and exploring candidate domains (novel-context decisions, multimodal integration, embodied choices), it highlights evolutionary advantages and practical relevance. It argues that starting from the null hypothesis that current AIs lack consciousness clarifies which computations are uniquely and efficiently performed by consciousness. The proposed division of labor informs AI alignment: aligning AI behavior with human values and safety constraints is better achieved by augmenting human conscious computation rather than attempting to replace it. Analogies to quantum supremacy and error correction are used to conceptualize mechanisms that might underlie efficient conscious processing, though empirical confirmation remains open.

The paper introduces the concept of conscious supremacy to focus inquiry on computations potentially unique to conscious processes and practically infeasible for non-conscious systems within biological constraints. It synthesizes theoretical perspectives and AI capabilities to outline candidate domains and suggests that aligning AI with humans should leverage a division of labor that preserves human-led conscious computations. Future research directions include: empirically identifying and operationalizing specific computations unique to consciousness; testing mechanisms for multimodal binding and error correction in conscious perception; clarifying relationships (if any) between quantum processes and consciousness; and developing alignment strategies that systematically partition tasks between AI (non-conscious augmentation) and humans (conscious decision-making).

The arguments are speculative and conceptual. The paper does not provide empirical demonstrations or a precise list of computations uniquely requiring consciousness, nor validated mechanisms for conscious supremacy or conscious error correction. The analogy to quantum computing is heuristic and unproven; the feasibility of simulating conscious computation on classical systems remains unresolved. Conclusions depend on evolving AI capabilities and lack experimentally grounded criteria for consciousness in artificial systems.