Researchers show AI can link many pseudonymous accounts to real people

In short: Researchers say a system using large language models can identify people behind pseudonymous online accounts faster and cheaper than many expect.

What happened

Researchers from ETH Zurich, the MATS program, and Anthropic described a pipeline that tries to “deanonymize” online accounts. Deanonymize means figuring out who is behind an account that uses a handle instead of a real name, like a Reddit alt, a secret X account, or a finsta.

The paper is called Large-scale online deanonymization with LLMs. It is unpublished and not peer-reviewed, meaning other independent experts have not formally checked the work yet.

The system follows four steps called ESRC, short for Extract, Search, Reason, and Calibrate. In simple terms, it reads someone’s posts and bio, pulls out clues like job, interests, and life details, searches public sites for possible matches, then weighs the evidence and ranks the best guesses (like a fast assistant that gathers clues and sorts suspects).

The researchers report several results. In one test, it re-identified at least 9 out of 125 people from redacted interview transcripts connected to Anthropic. In another “closed-world” setup, it linked split user profiles across platforms like Hacker News and Reddit with 67% accuracy.

They also say it was cheap, about $1.41 to $5.64 per target using standard paid AI and search services. The researchers say no Anthropic Claude models were used, although an Anthropic researcher advised on the work.

Why it matters

Many people rely on pseudonyms to separate parts of their lives, like talking about health issues, workplace problems, or politics. This research suggests that if your writing and personal details are scattered across the internet, AI can piece them together at scale. The authors note the approach depends on access to public data and may work poorly on people who share fewer identifying details.

Source: The Verge AI