A new research paper demonstrates that large language models can identify real identities behind anonymous online usernames with high accuracy. The method, costing as little as $4 per person, analyzes posts for clues and cross-references them across the internet. Researchers from ETH Zurich, Anthropic, and MATS warn of reduced online privacy.
Published on February 26, 2026, the paper titled "Large-scale online deanonymization with LLMs" explores how advanced AI chatbots can uncover the real people behind pseudonyms on platforms like Reddit and Hacker News.
The study, conducted by researchers from ETH Zurich, Anthropic—the parent company of Claude—and the MATS research group, introduces a technique called ESRC: Extract clues, Search, Reason, and Calibrate. The AI first examines posts for personal hints, such as interests in Python game coding, Marvel movies, complaints about school in Seattle, or distinctive writing styles. It then searches sites like LinkedIn, Google, and other Reddit accounts to find matching profiles. Finally, it reasons through alignments in style, hobbies, and timing to assess confidence levels, achieving matches without human intervention.
Testing on real Hacker News users yielded a 67% success rate in linking secret usernames to real identities, with 90% accuracy when the AI made predictions. For Reddit posts from the same user across different years or groups, the success rate reached 68%. The process is inexpensive, requiring up to $4 per individual using accessible chatbots like future versions of ChatGPT or Claude.
Simon Lermen, one of the main researchers, highlighted the implications for privacy. Previously, maintaining anonymity online relied on the effort required for manual investigations, which could take hours or days. Now, this automation allows individuals, companies, or authorities to rapidly analyze thousands of accounts, potentially revealing names, schools, cities, or jobs from a few comments. The researchers describe this as the end of "practical obscurity," where obscurity was once feasible despite technical possibilities.
