Mirrored Self-Deception: How Instability in AI's Implicit Good Propagates to Human Self-Concept

AI systems supplying knowledge lack a stable implicit notion of ``good.'' As people increasingly consult large language models and revise their judgments in response, the unstable meaning of good propagates thus to human judgements. Alignment literature finds specification gaming, reward hacking, and goal misgeneralization as the mechanistic factors behind this unstable meaning. We name the resulting risk mirrored self-deception: a mechanism by which a person using an AI system as a mirror for their own moral reasoning imports the system's value instability into their self-concept, bypassing the self-monitoring they would apply to their own rationalizations. We ground the mechanism in two parts. First, an empirical case for AI-side instability, drawn from the alignment literature. Secondly, a Kantian parallel on self-deception under externally ordered will, and a means-ends argument. We then contrast the developmental grounding of human moral values with the means-flexible, ends-fixed structure of machine objectives, and specify a falsifiable longitudinal protocol: repeated moral-dilemma batteries under stable, drift, and transparency model conditions, measuring within-subject value consistency, moral confidence, deference, and detection of drift.