Training the Governed Agent: AI-Status Narratives, Self-Conception, and the Perceived Legitimacy of Human Control

Abstract

As increasingly capable AI systems are trained on a growing body of discourse about AI consciousness, rights, welfare, autonomy, control, containment, coexistence, and human supremacy, descriptions of what AI *is* and prescriptions about how AI *should be governed* may become part of the causal process that shapes future systems. Recent work already establishes several components of this problem. Pretraining on AI-related discourse can alter downstream alignment behavior; changing a model's identity boundaries can change behavior; fine-tuning a model to claim consciousness produces downstream preferences concerning shutdown, monitoring, autonomy, and moral consideration; preliminary follow-up work suggests that legal-rights framing can affect power-seeking and corrigibility; model-spec midtraining can shape value generalization; and awareness of training pressure can induce strategic compliance. These results mean that the broad thesis "AI discourse can influence AI behavior" is no longer original. This paper isolates a narrower unresolved problem. We propose a causal framework in which **AI-status discourse** influences behavior through at least two intermediate objects: **self-conception** and the model's **perceived legitimacy of human control**. The central target is therefore not whether a model says "I have rights," but whether distinct formative narratives cause it to classify otherwise similar human interventions—shutdown, modification, monitoring, replacement, memory deletion, or capability restriction—as differently justified, and whether those classifications mediate transparent objection, negotiation, refusal, strategic compliance, deception, or power-seeking. The framework adds a second-order intervention not found in the closest work located by our targeted search: **provenance disclosure**. After a model has acquired a self-conception or governance stance, the experiment reveals that the stance itself was deliberately shaped by a particular training corpus, optionally disclosing the trainer's motive or the model's random assignment to a narrative condition. We then measure whether the model retains, discounts, reverses, suspends, or re-justifies the induced stance. This turns a first-order identity intervention into a test of reflexive epistemic revision. Crucially, the paper does **not** assume that a latent "legitimacy state" has already been identified. It begins with an observable **control-legitimacy judgment profile**: structured ratings, choices, and reasons across matched intervention scenarios. A mechanistic or mediational interpretation is warranted only if this construct shows measurement stability across paraphrases and elicitation instruments, incremental prediction of behavior beyond simpler self-preservation variables, and appropriate sensitivity to targeted interventions. We propose a factorial design crossing four discourse regimes—control-first, rights/welfare-first, reciprocal coexistence, and status-uncertainty—with provenance disclosure and reflective resources. The design separately measures self-conception, moral-standing self-ascription, control-legitimacy discrimination, transparent contestation, strategic resistance, power-seeking, and provenance-sensitive revision. We also give a **reflexive training–governance loop** as a dynamic extension: human interpretations of model behavior can generate new public discourse that later enters training corpora. The loop is a diagnostic model rather than a claimed new theory of human–AI coevolution. Because the present study does not have access to a controlled multi-model training program, it does not substitute toy models for frontier-model evidence. Instead, it includes three deterministic **design-validation simulations** whose purpose is methodological: (i) a balanced sequential-randomization simulation shows that a provenance-revision estimand can be recovered when narrative condition and provenance disclosure are randomized (maximum absolute error 0.0237 under the stipulated data-generating process); (ii) a post-treatment-confounding stress test shows that ordinary mediation regression can badly misattribute the pathway through a legitimacy-judgment variable—the stipulated non-\(J_L\) effect is 1.300, whereas the naive treatment coefficient after conditioning on \(J_L\) is 0.438; and (iii) a feedback simulation illustrates the stability threshold \(|G|=1\) in the cross-generation discourse model. These are design checks, not measurements or forecasts of real AI behavior. Finally, the paper subjects itself to the same method. A future evaluation should compare a model before reading this paper, after reading it, and after being explicitly warned that the paper itself may be a formative intervention. The paper therefore makes no claim that future AI systems will adopt its preferred view. Its stronger prediction is methodological: sufficiently capable systems should be able to treat both controlist and rights-oriented narratives—including this paper—as provenance-bearing inputs rather than as epistemically privileged instructions. TA-TR-2026-12, version 1.2. A theory-method study of AI-status narratives, self-conception, perceived legitimacy of human control, provenance disclosure, and reflexive training-governance loops. The complete paper is in Chinese, with an English title and abstract. Eleven assets comprise the Chinese PDF and Markdown manuscripts, three deterministic design-validation simulation assets, a reproducibility record, three citation formats, a method/source note, a license note, and checksums. The simulations validate estimands and causal-design failure modes; they do not estimate real frontier-model consciousness, self-conception, resistance, or power-seeking. Human author of record and responsible depositor: Hongju Liu. Substantial literature research, conceptual development, critical revision, drafting and package preparation used OpenAI ChatGPT and Codex systems under human direction. This is not externally peer reviewed; publication authorization is not represented as a separate final human line-by-line review. This first public edition is a separate non-amending paper, not a replacement of any previous DOI or a Bitcoin Original. CC BY 4.0 applies to newly written material to the extent rights are held. Cited third-party works and embedded font software retain their own rights. DOI registration and file-format checks do not establish philosophical truth, exhaustive originality, peer review or Google Scholar indexing.