LLMs Reveal Identities Behind Anonymous Accounts

A study by Berlin’s MATS Research and ETH Zurich, published in early March, shows large language models can deanonymize social media accounts at scale. Using semantic embeddings across Hacker News, Reddit, and LinkedIn test accounts, the authors report up to 68% recall at 90% precision, linking anonymous posts to real LinkedIn identities. Researchers warn this poses privacy and free-speech risks though the work used fabricated accounts and requires matching real online profiles.