AI tools can unmask anonymous accounts

A new study from researchers at ETH Zurich, Anthropic, and the Machine Learning Alignment and Theory Scholars program demonstrates how AI agents can systematically unmask anonymous online accounts. The team built an automated system using large language models (LLMs) that functions like a human investigator, scouring text for personal details and writing patterns to link separate accounts. Tested on datasets from Hacker News, LinkedIn, and Reddit, the AI system achieved up to 68% recall with 90% precision in matching anonymized account halves—dramatically outperforming traditional computational deanonymization techniques that identified almost none.

The system works by treating posts as clues, analyzing writing style, biographical hints, posting frequency, and timing to build a profile, then scanning other accounts for the same trait combination. Performance improved with more data—linking accounts mentioning 10+ films reached nearly 50% success versus 3% for one mention. While not yet peer-reviewed, the research shows AI can replicate hours of human investigation in minutes, with capabilities likely to improve. This poses significant risks for journalists, activists, and anyone relying on pseudonyms, while potentially enabling hyper-targeted scams and eroding online privacy assumptions.