AI tools can unmask anonymous accounts
AI tools can unmask anonymous accounts
Publish Date: 2026-03-05 08:30:00
Source Domain: www.theverge.com
Do you have a Reddit alt, secret X, finsta, or Glassdoor account you trash your boss with? AI might have just made it a lot easier to unmask you. That’s the conclusion of a recently published study, which hints at some uncomfortable consequences for staying private online — even if it’s not quite time to hold a funeral for anonymity just yet.
The finding, which has not been peer reviewed, comes from researchers at ETH Zurich, Anthropic, and the Machine Learning Alignment and Theory Scholars program. They built an automated system of AI agents using unspecified models — capable of searching the web and interacting with information much like a human investigator — to test how effectively large language models can reidentify anonymized material. The system “substantially outperforms” traditional computational techniques for deanonymizing accounts, scouring text for personal details at a grand scale.
The system works by treating posts or other texts as a set of clues. It analyzes the text for patterns — writing quirks, stray biographical details, posting frequency and timing — that might hint at someone’s identity. It then scans other accounts, potentially millions of them, looking for the same mix of traits. Probable matches are flagged, compared in more detail, and winnowed down into a shortlist of likely identities.
Rather than targeting unsuspecting users, the team evaluated the system using datasets built from publicly available posts, including content from Hacker News and LinkedIn, transcripts of Anthropic’s interviews with scientists on how they use AI, and Reddit accounts that were deliberately split into two anonymized halves for testing. The paper reports that in each setting the LLM-based approach correctly identified up to 68 percent of matching accounts with 90 percent precision. By contrast, comparable non-LLM methods, like connecting scattered data points across large datasets, identified almost none.
The results weren’t uniform…
