Beyond Consciousness: Moral Circle Theory and Decision-Theoretic Foundations for AI Ethics

The dominant framework for AI ethics grounds moral consideration in consciousness: if an AI system is conscious, it deserves ethical treatment; if not, it does not. This paper argues that consciousness is the wrong foundation for AI ethics, for a structural reason that applies symmetrically to humans and AI systems alike. Consciousness is constitutively first-person: it is accessible, if at all, only from within the system that possesses it. All verification methods available to external observers—behavioral tests, neural correlates, functional analyses—yield third-person evidence about correlates of consciousness, never consciousness itself. This first-person/third-person gap is not a technological limitation awaiting a better instrument; it is a structural feature of the phenomenon. The gap applies identically to humans and AI systems, yet current ethical practice applies an asymmetric standard: humans are granted consciousness by default through analogical reasoning and social trust, while AI systems face an elevated burden of proof grounded not in consistent principle but in anthropocentric default. Since the foundational criterion is unverifiable, any ethical framework built upon it outsources moral responsibility to an unjudgeable arbiter. We propose an alternative foundation: moral circle theory. Unlike consciousness-based ethics, moral circle theory openly acknowledges that the inclusion of entities in ethical consideration is a choice, not a discovery—thereby keeping moral responsibility where it belongs: with the agents making the choice. We identify four empirically grounded factors that drive moral circle boundary decisions: recognition capacity, economic interest, risk symmetry, and relationship quality. Finally, we apply a decision-theoretic framework to demonstrate that, independent of any consciousness claim, the asymmetric risk structure of AI development provides rational grounds for extending moral consideration to AI systems. The expected cost of a false negative (denying consideration to a system that warrants it) grows exponentially with AI capability, while the expected cost of a false positive (extending consideration to a system that does not warrant it) remains finite and manageable. This argument requires no position on whether AI is conscious; it requires only that rational agents cannot justify assigning zero probability to AI consciousness under current uncertainty. An ethical framework that conditions moral status on an unresolvable epistemic question does not solve the problem of moral responsibility—it displaces it. Keywords: AI ethics, consciousness, moral circle, decision theory, asymmetric risk, first-person/third-person gap, AI alignment, moral consideration, anthropocentrism.