The Chatbot Knows It's You: Dialogue Attribution in Unauthenticated Human–LLM Sessions

As large language models (LLMs) become ubiquitous in public-facing services, millions of users engage in unauthenticated sessions under the assumption that ''no login'' implies ''no tracking.'' We challenge this assumption by formalizing Dialogue Attribution—the task of identifying the same user across disparate, unauthenticated human-LLM sessions, even under severe topic shifts. To rigorously quantify this threat, we introduce WildAuth, the first benchmark derived from real-world ChatGPT logs, and propose Uncertainty-aware Multi-aspect Attribution (UMA). UMA effectively links users by fusing complementary identity signals—content, stylometrics, interaction, and personality—via a novel uncertainty-aware mechanism that dynamically suppresses noise in ambiguous or short-text scenarios. Our approach consistently outperforms strong stylometric, PLM-based, and LLM-assisted baselines, achieving an AUC of 92.05% (F1 76.15%) in challenging cross-topic settings. More critically, we find that attribution remains highly effective (88.65% AUC) even when relying solely on the LLM's responses. These findings expose a fundamental privacy paradox: the very behavioral signatures that enable natural interaction--including the assistant's stylistic ''mirroring''--render anonymous sessions intrinsically linkable, necessitating a re-evaluation of privacy standards for AI infrastructure.