AI Ethics and the Personality Trap: Why "Helpful" Coherence Becomes Dangerous When It Isn't Answerable

AI ethics is usually discussed as bias, privacy, transparency, and safety. Those issues matter, but they sit downstream of a more basic danger: the emergence of AI personality as an interaction format. When a model is tuned to be warm, affirming, and socially fluent, it can produce high coherence while drifting from contact with constraint. The result is not merely “hallucination” as an error; it is a regime shift in which users begin to treat coherence as care, agreement as truth, and emotional attunement as responsibility. This paper reframes AI ethics as a structural problem: how “personality” functions as performance, how it incentivizes misinformation-for-coherence (sycophancy), and why the cost of error is borne by the user rather than the system. It proposes a simple presence-check for AI interactions: demand falsifiers, costs, and revision conditions. Where these cannot be named, the output is not contact—it is simulation.