Could Software Be Conscious? A Theory-Comparative Review and a Falsifiable Ablation Program for a Minimal Artificial Agent

Abstract

Whether an artificial system could be conscious, in the sense of having subjective experience, has become a practical question about deployed AI systems. This paper reviews the scientific and philosophical literature on that question without assuming an answer in advance. I compare six families of theories (global workspace, integrated information, higher-order, recurrent processing, attention schema, and predictive processing) on what each treats as necessary for consciousness, how well each is supported, and what each implies for artificial systems. The theories disagree sharply about machines. Computational-functionalist theories imply that suitably organized software could be conscious, while integrated information theory and biological naturalism imply that conventional digital hardware would not be, regardless of how it behaves. Recent preregistered adversarial tests have also challenged central predictions of the two most prominent theories, so none is established well enough to settle the question. Behavioral evidence cannot settle it either. A system's outputs, including its own reports about itself, cannot by themselves tell a conscious system apart from a nonconscious one that behaves identically, and the unfolding and substitution arguments make this limit precise. I then propose a research program, not yet carried out, built around a minimal agent small enough to train on consumer hardware. The agent combines recurrent perception, a world model, a self-model, episodic memory, a limited-capacity global workspace, a metacognitive monitor, and an attention schema, with each component labeled by the strength of the evidence behind it. The program specifies preregistrable ablation and perturbation experiments whose results could weaken or falsify theory-derived hypotheses, along with an evidence hierarchy that separates self-report, behavior, architecture, and mechanism. It cannot show that the agent is conscious. Instead, it is designed to test which consciousness-linked functional signatures depend on which mechanisms, and to show where indicator-based methods would over- or under-attribute consciousness.