AI can unmask anonymous online users in minutes: Study
A new study from ETH Zurich and Anthropic found that AI models (LLMs) can quickly figure out who's behind pseudonymous accounts online.
What used to take hours of detective work can now be done in minutes, and for just a few bucks per profile.
This puts a big question mark on how safe pseudonyms really are.
AI can identify users with over 90% precision
Researchers built an AI system that scans posts, picks up clues about identity, and matches them to real profiles.
When tested on sites like Hacker News, Reddit, and even interview transcripts, the AI achieved about 67% recall at 90% precision on Hacker News, reached up to 68% recall at 90% precision in some Reddit experiments, and produced lower identification rates and different precision levels on interview transcripts: way better than older methods.
Research team is not sharing code or data yet
Daniel Paleka said, "If your operational security requires that no one ever spend hours or days investigating who you are, this security model is now broken."
The team stressed they're handling things responsibly: they got ethics approval and aren't sharing their code or data yet.