LLMs can unmask pseudonymous users at scale with surprising accuracy

Researchers have discovered that artificial intelligence can effectively analyze burner accounts across social media platforms, enabling the identification of users who prefer to remain anonymous. This alarming revelation raises significant concerns regarding online privacy, as detailed in a recent study. The research highlights experiments that successfully correlated specific individuals with their pseudonymous accounts and posts, achieving a deanonymization success rate that surpasses traditional methods. Notably, the recall rate—indicating how many users were accurately identified—reached an impressive 68%. Additionally, the precision rate, which measures the accuracy of these identifications, soared to 90%. These findings challenge the assumption that pseudonymity serves as a sufficient privacy measure for users engaging in sensitive discussions online. As the researchers pointed out, the capability of AI to swiftly and inexpensively uncover the identities behind obscured accounts poses serious risks, including doxxing and stalking. Furthermore, it enables the creation of extensive marketing profiles based on personal data, such as location and occupation. The researchers emphasized the profound implications of their findings for online privacy. Many users have long operated under the belief that pseudonymity offers adequate protection, assuming that targeted deanonymization requires considerable effort. However, the advent of large language models (LLMs) significantly undermines this belief. To conduct their study, the researchers gathered multiple datasets from public social media platforms while ensuring user privacy was preserved. One dataset included posts from Hacker News and LinkedIn, which were linked through cross-platform references found in user profiles. After removing identifying information from the posts, they applied a large language model to analyze the data. Another dataset was sourced from a Netflix release, containing micro-identities that revealed individual preferences and transaction records, demonstrating the potential for identifying users and their political affiliations. These insights underscore the urgent need for enhanced privacy measures in the digital landscape, as current assumptions about anonymity are increasingly being called into question.