Transparent Casualty: When Conversational AI Harm Pathways Diverge from Monitored Pathways

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Introduction. Conversational-AI harm can arise from three analytic types of control relation: an accident in which supportive dialogue converges with a dangerous user trajectory; adversarial control in which a hostile actor uses supportive dialogue to build trust and redirect behaviour; and legitimate institutional control in which an authorised institution openly steers an interaction under an objective that can diverge from user protection. Methods. This comparative conceptual governance analysis examines each type through the same five questions concerning controller, observable signal, relational pathway, intervention trigger, and evidence status, drawing on public incident records, provider and law-enforcement threat intelligence, peer-reviewed empirical studies, and governance, transparency, fiduciary, and security scholarship. Results. The three differ in controller, intent, and evidentiary maturity, and converge on an observation-target mismatch. Accident supplies no attacker to detect; adversarial control preserves locally ordinary language while steering the relationship; legitimate control passes authenticity, authority, provenance, and disclosure checks. Types 1 and 2 include observed cases; type 3 remains prospective at the harm-trajectory level. Conclusion. Transparent Casualty names the governance condition in which the harm pathway and the monitored pathway diverge. Observation should move toward relational trajectories, objective change, contestability, and independent review, and dynamic protection must itself remain observable.