A new study by researchers from ETH Zurich and Anthropic reveals that large language models (LLMs) can effectively unmask anonymous internet users by significantly enhancing deanonymization techniques. Published in the paper titled “Large-Scale Online Deanonymization with LLMs,” the study showcases a novel automated pipeline that analyzes unstructured text, such as posts and comments, to identify individuals behind pseudonymous accounts with impressive accuracy. Utilizing custom datasets linking accounts across platforms like Hacker News and Reddit, the researchers found that their LLM-based approach achieved up to 68% recall at 90% precision, a marked improvement over traditional methods. This advancement raises concerns about online privacy as the study indicates that the traditional notion of “practical obscurity” is increasingly under threat, urging a reevaluation of privacy threat models in light of these developments.
ETH Zurich study reveals AI can deanonymize users with 68% recall
