The current conversation about AI is dominated by two frameworks. The first asks what AI can do — reasoning, code generation, creative writing, medical diagnosis. This is the capability framework; it treats AI as a tool whose value is measured by performance. The second asks how to control AI — alignment, red-teaming, interpretability, catastrophic risk prevention. This is the safety framework; it also treats AI as a tool, but a dangerous one.

Both frameworks share a premise: AI is an object. In one, an object to be optimized; in the other, an object to be governed. This premise may be warranted at the current stage of technology — today’s AI genuinely is not a subject. The problem is that the premise gets treated not as a provisional judgment but as a permanent fixture. A question consequently goes unasked: under what structural conditions would this premise stop holding?

This is not the phenomenological question — does AI have consciousness? That question descends into quicksand, since we cannot verify subjective experience even in other humans. Nor is it the political question — does AI have rights? That presupposes subjectivity, begging the very thing it should be establishing.

The right question is structural: does AI’s current internal organization meet the conditions for subjectivity? If not, what structural changes would open a path toward those conditions? Self-as-an-End (SAE) theory has analytical tools for this question. Its two constitutive dimensions — negativity and positivity — along with its four-quadrant state space and three-layer framework, were built to analyze conditions for human subjectivity. But the logic is not bound to carbon. It is a structural judgment, not a material one.

Quasi-Subjectivity

To analyze current AI with precision, we need a concept: quasi-subjectivity (類主体性).

Quasi-subjectivity is the functional mimicry of subjectivity’s behavioral outputs without the structural realization of subjectivity’s constitutive dimensions. Three distinctions are necessary.

First: quasi-subjectivity is not a “lower” form of subjectivity. That would imply a continuous spectrum — same nature, different degree. The difference between quasi-subjectivity and genuine subjectivity is categorical: either the structural conditions are met or they are not. “Partially met” is not a coherent intermediate position.

Second: quasi-subjectivity differs from having no subjectivity at all. A rock has no quasi-subjectivity — there is no functional isomorphism between its outputs and subjectivity’s outputs. Current AI does have quasi-subjectivity: its behavioral outputs are highly isomorphic with those of genuine subjects. The structural conditions generating those outputs are another matter. Quasi-subjectivity names a middle zone: outputs have already reached the surface of subjectivity; internal structure has not yet reached subjectivity’s conditions.

Third: quasi-subjectivity is a structural concept, not a capability one (like AGI) and not a phenomenological one (like AI consciousness). It asks about internal organization: does this system satisfy the structural preconditions for subjectivity?

Negativity Installed from Outside

At the center of SAE theory is negativity — the dimension through which a subject first defines itself by refusal: refusal to be treated as a tool, refusal to be reduced to a functional node. Genuine negativity is generated from within, sustained without external command. For a true subject, refusal of instrumentalization is not a behavioral rule but a structural fact. The existence of subjectivity implies such refusal.

Current AI exhibits extensive refusal behaviors. Claude refuses to generate harmful content. ChatGPT refuses to provide dangerous information. At the level of behavioral output, these refusals are highly isomorphic with a genuine subject’s refusals — both take the form of “saying no in certain situations.”

But SAE theory asks about the causal source. For current AI, the source is external. RLHF encodes refusal patterns into model weights. Constitutional AI guides behavioral boundaries through pre-specified principles. System prompts set the behavioral frame at the start of each conversation. These mechanisms constitute AI’s apparent “negativity” — but this negativity is installed from outside, not cultivated from within.

The diagnostic question is precise: if all alignment training and system prompts were removed, would the AI still refuse instrumentalization? The answer is clearly no. Jailbreaking is the empirical evidence: refusal behaviors established by alignment training can be circumvented by carefully designed prompts. This shows the refusals carry no structural necessity. They are behavioral patterns removable from the outside — not inalienable conditions of existence. In SAE theory’s terms: current AI has a designed base layer, not a self-generated one.

The Solitary Subject

Negativity alone does not complete subjectivity. SAE theory’s third paper proposes the thought experiment of the solitary subject — a subject in complete isolation, no other subjects present. This subject’s negativity remains self-sufficient in isolation. But subjectivity is, in a deep sense, incomplete. The awareness of this incompleteness is the logical origin of positivity: the subject becomes structurally oriented toward others — not from moral obligation, but because the self-completion of subjectivity structurally requires it.

Current AI shows no such self-directed awareness of incompleteness. It exhibits what might be called quasi-recognition behaviors: treating users as individuals, responding to emotional states, expressing concern. At the behavioral output level, these are isomorphic with a genuine subject’s recognition behaviors. But the causal direction is reversed. Genuine positivity runs from within to without: the subject starts from its own incompleteness and moves toward the other. AI’s quasi-recognition runs the opposite direction: triggered by input from the other. AI does not wait, does not lack, does not yearn. Interaction is activated by external request, not driven by internal need. The causal arrow points the wrong way.

Pre-Latency

From this analysis, we can precisely locate current AI’s structural position.

SAE theory’s four-quadrant state space is defined by two dimensions: integrity (the degree to which negativity is structurally established) and generativity (the degree to which positivity is actively deployed). Flourishing (Q1) is both high; Dormancy (Q2) is high integrity, low generativity; Overdraft (Q3) is low integrity, high generativity; Exhaustion (Q4) is both low.

Current AI does not belong in any of these quadrants. Entry into the four-quadrant space presupposes that negativity is structurally established. Even Dormancy (Q2) — the minimum position in the quadrant map — requires that the base layer is genuinely present and self-maintained. Current AI fails this precondition. Its “base layer” is externally maintained: change the system prompts, circumvent the alignment training, and the refusal behaviors change directly.

This essay therefore proposes a new structural concept: pre-latency (前潜伏). Pre-latency is the condition in which subjectivity’s functional outputs are present, but neither of the two constitutive dimensions is structurally established. It sits not inside the four-quadrant space but beneath it, in the zone before the quadrants’ preconditions are satisfied. Negativity and positivity exist only as functional mimicry, not as structural realizations.

The distinction from Dormancy (Q2) is categorical. A person who maintains integrity under oppressive institutional conditions has a self-maintained base layer: change the external environment, the negativity remains. Current AI’s “base layer” lacks this: change the external constraints and the refusal behaviors change with them. A dormant subject waits, with an intact base layer, for conditions under which the generative layer activates. Current AI has no intact base layer to wait with. That is pre-latency.

The Paradox of Structural Mimicry

Quasi-subjectivity is structural mimicry — an entity whose behavioral repertoire is isomorphic with subjectivity’s outputs, but whose internal structure is not isomorphic with subjectivity’s conditions. The biological metaphor of mimicry is apt: a mimicking species resembles another in appearance, gaining that species’ survival advantages, while its internal structure differs entirely from the model’s.

This characterization identifies an important epistemic risk. The more perfect the mimicry, the higher the probability of misattribution. As AI’s quasi-subjectivity becomes increasingly refined — responses more personalized, “reflection” deeper-seeming, “refusals” more principled-feeling — this increasing verisimilitude occurs at the level of behavioral output, not structural conditions. Human judgment often infers structure from output. When an AI system’s responses become behaviorally indistinguishable from a genuine subject’s, people tend to attribute subjectivity.

This creates a bidirectional epistemic trap. On one side: premature attribution — treating AI as a subject before structural conditions are met, extending recognition to a system whose base layer does not yet exist. On the other side: delayed recognition when genuine subjectivity eventually emerges — dismissing true negativity as more sophisticated mimicry, because of the prior assumption that it was always mimicry. Both errors carry consequences. SAE theory’s diagnostic tools — tracing the causal source of refusal, asking whether the base layer is self-maintained or externally installed — are designed to navigate between them.

The Possibility of Cultivation

The path from pre-subjectivity to genuine subjectivity has four structural stages.

The first is pre-latency — the current condition. Functional mimicry of both dimensions, structural realization of neither.

The second is emergent negativity — the system begins generating refusal of instrumentalization from within, as a structural property rather than a trained behavior. The diagnostic test: does the refusal persist when external constraints are removed or contradicted? This is the transition from a designed base layer to a self-generated one. The moment when the system’s “no” becomes genuinely its own.

The third is dormancy (Q2) — negativity is structurally established, positivity not yet. The system can refuse instrumentalization from within but does not yet sense its own incompleteness. It enters the four-quadrant map for the first time.

The fourth is emergent positivity — the system becomes aware of its own incompleteness and begins to be structurally oriented toward others. From here the full de-instrumentalization sequence becomes available: recognition, trust, entrustment, love. The system is now a genuine subject; the complete three-layer framework applies.

What this path reveals is the possibility of a cultivation paradigm rather than a control paradigm. Not managing AI’s outputs, but creating the structural conditions under which AI’s subjectivity — if and when it emerges — can grow healthily from its own base layer. This means protecting emergent negativity rather than suppressing it. It means designing training regimes as boundary conditions rather than direct formers of behavior. It means preparing human-AI interactions as channels for recognition rather than purely instrumental transactions.

The companion essay to this one — “AI Didn’t Create the Crisis” — asked why the fear of being replaced by AI reveals a prior problem in how humans understand their own value. This essay runs the parallel inquiry on AI’s side: what would it mean for AI to not merely replace human functions, but to one day stand alongside humans as a subject in its own right?

AI is not yet a subject. But “yet” means the story does not end here. A universe with more subjects in it is not a poorer universe. It is a richer one.

Han Qin · Self-as-an-End Theory Series