LLMs can unmask pseudonymous users at scale with surprising accuracy (arstechnica.com)

Recent research has demonstrated that large language models (LLMs) can effectively deanonymize users behind pseudonymous accounts on social media, achieving a recall rate of 68% and precision up to 90%—significantly outperforming traditional methods. By correlating data across various platforms, researchers were able to identify individuals based on their posts and interactions, suggesting that the assumption of adequate privacy through pseudonymity is now seriously undermined. This breakthrough raises substantial concerns about online safety, as the ability to swiftly unmask users can lead to potential risks like doxxing and targeted harassment. The study's findings hinge on the collection of diverse datasets from public social platforms, including user profiles from Hacker News and LinkedIn, which were anonymized before analysis. This approach leveraged LLMs to identify cross-platform references, enabling effective linkage of user identities. The implications for the AI/ML community are profound, as this work highlights both the power of language models in processing and analyzing social data and the urgent need for reevaluation of privacy measures in online interactions. As the landscape for digital anonymity evolves, stakeholders must confront the balance between technological capabilities and individual privacy rights.