As courts confront the question of whether interactions with commercial AI systems constitute "third-party disclosure," an emerging doctrinal drift threatens to create a wealth-stratified privilege regime. Self-hosted, firewall-isolated large language models (LLMs) allow well-resourced actors to preserve confidentiality, while ordinary users—who rely on commercial cloud-based AI services—may find their communications treated as disclosures to a third party. This paper argues that the resulting inequity is not a necessary consequence of settled doctrine but the product of a classification error: confusing the identity of the tool with the behavior of the system. Drawing on relational-misclassification concepts from human-layer security frameworks, this paper analyzes how the gap between the user's mental model ("I am using a tool") and the law's emerging interpretation ("You communicated with a third party") creates structural vulnerability. The paper proposes a functional classification model grounded in system behavior and reasonable expectations rather than tool identity. Privilege should not depend on the ability to afford private infrastructure.
Narnaiezzsshaa Truong (Thu,) studied this question.