LLMs and online anonymity: a precarious balance

A recent study has highlighted how large language models (LLMs) can be used to identify users hiding behind pseudonymous accounts on social media. The research, published on arXiv, shows alarming results regarding the ability of these models to correlate specific individuals with accounts or posts on different platforms.

The results indicate a significantly higher success rate compared to traditional deanonymization methods, which rely on manual analysis of structured data. In some cases, the recall (the percentage of users correctly identified) reached 68%, while the precision (the accuracy of the identifications) reached up to 90%.

Implications for privacy and security

The ability to unmask pseudonymous users on a large scale has significant implications for online privacy. Anonymity, albeit imperfect, allows many people to participate in sensitive public discussions without revealing their identity. The ability to quickly and cheaply identify these users paves the way for practices such as doxxing, stalking, and the creation of detailed marketing profiles.

In an era where privacy is increasingly at risk, the ability of LLMs to compromise anonymity represents a significant challenge for online security and the protection of personal data.