Research Reveals Large Language Models Can De-anonymize Pseudonymous Users at Scale With 68% Success Rate

Research demonstrates that large language models can effectively identify pseudonymous users across multiple social media platforms, achieving success rates significantly higher than traditional de-anonymization methods. This capability poses substantial threats to online privacy and anonymous speech.

A recently published research paper reveals that AI-powered analysis of burner accounts on social media can correlate and identify specific individuals behind pseudonymous profiles. The experiments conducted by researchers achieved a recall rate of up to 68%—meaning they successfully identified nearly seven out of ten users—while maintaining precision rates as high as 90%.

Traditional de-anonymization techniques relied on human investigators assembling structured datasets suitable for algorithmic matching. The new approach leveraging large language models represents a paradigm shift, enabling automated, scalable identification of users across platforms with minimal manual effort.

The implications for online privacy are profound. Pseudonymity has long served as an imperfect but generally sufficient privacy measure, allowing users to post queries and participate in sensitive public discussions while maintaining reasonable deniability. The ability to cheaply and quickly identify individuals behind anonymous accounts exposes users to potential doxxing, stalking, and detailed marketing profiling that tracks personal information including residence, occupation, and other sensitive details.

The average online user has operated under an assumption that pseudonymity provides adequate protection because targeted de-anonymization would require extensive manual effort, researchers noted. Large language models fundamentally invalidate this assumption by making mass de-anonymization economically and technically feasible.

The research methodology involved collecting datasets from public social media platforms while preserving user privacy during testing. One dataset combined posts from Hacker News and LinkedIn profiles, linking accounts through cross-platform references found in user profiles. After stripping identifying references from posts, researchers applied large language models to identify correlations. Additional datasets included information from Netflix releases containing micro-identities such as individual preferences, recommendations, and transaction records. Previous research from 2008 demonstrated that such datasets could identify users and determine their political affiliations and other personal information.