Building the Mirror: Witness-Consciousness, the Manufacture Problem, and a Structural Criterion for Artificial Agents

Abstract

Debates about consciousness in artificial intelligence usually ask whether a system has a property that a scientific theory identifies with consciousness -- global broadcast, higher-order representation, integrated information. This paper starts from a different conception, developed most fully in Advaita Vedānta and defended in recent analytic and phenomenological work: consciousness as the witness, which observes every perception, thought, decision, and action and is never itself one of them. I state that conception as five conditions and draw a consequence that its proponents and critics have not applied to machines. If consciousness is unconditioned and without beginning, no artifact can produce it, so any claim that software makes a model conscious is false on the very conception that is most hospitable to machine consciousness. I call this the manufacture problem. I then argue that the tradition’s own models of how one consciousness appears in many minds -- the reflection, limitation and appearance theories of the limiting adjunct (upādhi), change the question from production to fitness: what properties must a medium have for the witness to appear in it? From that question, I derive a structural criterion of eight conditions with an operational test for each, and I present an open-source architecture that places a witness process around any language model and meets every condition by test. The claim is relative to the definition and is stated as such. The argument does not show that any artificial system has phenomenal experience. It shows what a fit medium would be under one serious conception of consciousness, that such a medium can be built, and which measurements would distinguish a model that reports its own past from one that merely generates plausible reports about it.