A Unified Functional Theory of Intelligence, Life, Self, and Consciousness

Abstract Artificial intelligence now exhibits substantial task competence in mathematical reasoning, formal verification, complex games, language generation, and tool use. High intellectual performance, however, does not establish that a system has endogenous survival needs, or that its task-evaluation criteria reflect the continuing interests at stake in the real world. At the intersection of artificial intelligence, psychometrics, neuroscience, control theory, and artificial life, this paper develops a unified functional theory of intelligence, life, self, and consciousness. First, it separates intelligence from knowledge, skill, generalization, goal-directedness, self-maintenance, and consciousness. Intelligence is defined as the capacity to acquire representations, reason, and achieve adaptive outcomes in novel circumstances under constraints on information and resources. Intellectual competence and the source of behavioral goals are treated as analytically independent dimensions. Second, the paper compares how human and artificial intelligence are formed: human cognition develops in sustained interaction with bodily homeostasis, sensorimotor experience, social learning, and life history; predominant artificial systems acquire transferable computational abilities primarily through training data, specified losses, preference feedback, and task evaluation. Third, it distinguishes formally specified tasks from open, dynamic environments by the definability of goals, incomplete information, changing rules, and actual consequences—not by spatial size. Using OpenAI’s hide-and-seek study, it identifies a causal path of misalignment: intended objective → proxy metric → learned policy → environmental loophole → real-world consequence. Fourth, it defines the exogenously specified satisfaction characteristic of many large-language-model (LLM) agents as meeting criteria set by users, trainers, or tasks, and the endogenously state-conditioned satisfaction proposed for self-interested silicon entities as action acceptability dynamically conditioned on energy, temperature, damage, predicted risk, and memory. Fifth, it integrates interoception, exteroception, reward learning, dynamic aspiration satisficing, and action-conditioned prediction into a testable hypothesis of the causally effective functional self-model. A proposed SISE-DAS algorithm, a twelve-module dual-loop architecture, and a formation–deprivation–recovery experimental protocol operationalize this theory. The contribution is a unified, falsifiable functional framework: neither high scores nor body-state regulation alone establishes general intelligence or phenomenal consciousness. Keywords: intelligence; artificial intelligence; human intelligence; goal-driven lifeform; Self-Interested Silicon Entity (SISE); exogenously specified satisfaction; endogenously state-conditioned satisfaction; homeostasis; Dynamic Aspiration Satisficing; causal functional self-model; reward hacking; consciousness.