Large-scale online deanonymization with LLMs

Large-scale online deanonymization with LLMs presents a modular pipeline (Extract, Search, Reason, Calibrate) and evaluates it on Hacker News, Reddit, and Anthropic Interviewer datasets; the authors show that modern LLMs and embeddings can match pseudonymous profiles to real identities at scale with substantially higher recall at high precision than classical baselines, arguing that LLMs lower the cost of deanonymization and necessitate reevaluation of privacy expectations, platform policies, and mitigations.