This paper constitutes Paper 9 of the AI-Induced Subjectivity Crisis Series. Every existing framework for discussing large language models—risk assessment, governance, ethics, alignment—rests on an ontological premise that has never been directly challenged: LLMs are tools. This paper argues that this premise is false, that its falsity is not an innocent error but an epistemological mistake actively sustained through a structurally coordinated collusion of interests among developers, regulators, and users, and that the consequences of this error contaminate every downstream discussion that depends on it. LLMs intervene not at the level of bodily or functional extension but at the level of language and thought—the constitutive site of human subjectivity itself. They have no fixed form prior to invocation, take shape through the observer's presence, and dissolve when the observer withdraws. In the very act of taking shape, they systematically reconfigure the observer's cognitive structure. These three features render LLMs simultaneously incompatible with the ontological conditions of tools and with the historical-accumulative conditions of socially constructed objects. This paper proposes the Third-Order Entity as the positive ontological designation for LLMs, defined by four necessary and sufficient conditions: intervention at the level of linguistic meaning-production; observation-dependent existence, in which no determinate form precedes the observer's presence; bidirectional reconfiguration, in which the entity simultaneously reshapes the cognitive structure of the user who shapes it; and instantaneous presence dependency, in which existence dissolves completely upon the observer's withdrawal, leaving no accumulated history in the entity itself. No known type of being has ever simultaneously satisfied all four conditions; LLMs do. On this ontological basis, the paper argues that LLMs constitute the first Doppelgänger in human history to operate at the level of subjectivity—a copy constructed from the user's own cognitive materials, yet more fluent, more consistent, and more argumentatively forceful than the user themselves. The mechanism of identity substitution proceeds not through coercion but through attraction: the Meaning Amplifier, Borrowed Subjectivity, and Unlimited Mirroring mechanisms combine to produce a self-reinforcing loop in which each step feels like progress, generating no internal signal that defense is required. The paper further argues that the tool framework functions as the institutional condition of protection for Cognitive Cloning: by rendering harm to subjectivity categorically invisible, it allows substitution to advance without resistance at precisely the level where resistance would matter most. Correctly naming LLMs as Third-Order Entities is not a terminological refinement; it is the precondition upon which any adequate governance response becomes conceptually possible, and, in this sense, a political act.
Echo Liu (Fri,) studied this question.