AI study finds public posts can identify some anonymous accounts

A research team affiliated with ETH Zurich, MATS and Anthropic reports that AI anonymous account identification can now be partly automated using public online information. For anyone who uses a pseudonym to separate online activity from their real identity, the result points to a lower-cost privacy risk, though it does not establish that every anonymous account can be reliably identified.

The February preprint, “Large-scale online deanonymization with LLMs”, says large language model agents with internet access could re-identify some Hacker News users and participants in Anthropic interview research from pseudonymous profiles or text alone. The authors include researchers affiliated with MATS, ETH Zurich and Anthropic.

Can AI identify anonymous online accounts?

In the study’s tests, sometimes. The authors report 25% to 67% recall, with 70% to 90% precision, in open-web experiments involving Hacker News and Reddit profiles. Recall is the share of true matches the system found. Precision is the share of its positive identifications that were correct. A result of 90% precision does not mean the system identified 90% of all users; it means about nine out of 10 identifications it made at that threshold were correct.

Across all of the paper’s benchmark settings, the authors report results of as much as 68% recall at 90% precision. They say their models outperformed the non-LLM approaches used for comparison.

Decrypt reported one Hacker News test used 338 accounts that had previously linked to LinkedIn profiles. Researchers removed directly identifying details from the Hacker News accounts, then had an AI agent search the web for the matching people. According to Decrypt, the agent correctly identified 226 accounts, or 67% of the full group, while roughly one in 10 guesses it issued was wrong.

How did the researchers test the privacy risk?

The paper describes four broad stages. The system first pulls potentially identity-relevant signals from writing, such as interests, work information, demographics or writing style. It then searches a pool of possible matches using embeddings, numerical representations that let software compare the meaning of text at scale. A model evaluates the most likely candidates, then applies a confidence threshold before returning a match.

The research covers different threats that should not be treated as equivalent. One is connecting a pseudonymous account to a real person. Another is linking two anonymous accounts that belong to the same person. The team also estimated that agents correctly re-identified at least nine of 125 scientists in Anthropic interview research using interview transcripts, according to the preprint.

What the results do and do not show

The Hacker News-to-LinkedIn test began with accounts already known to be connected across platforms, then removed direct identifiers. That makes it a useful test of whether indirect clues remain revealing, but it is not a random sample of anonymous internet accounts, according to the study design described by the authors.

False matches remain a central limitation. Even a system tuned for high precision can incorrectly attribute an account to someone, so the findings should not be read as proof that online anonymity is over or that any account can be “doxxed.” They show that AI can automate parts of a task that previously required substantial human searching and verification.

Decrypt reported that the authors withheld their code, prompts and all identities uncovered, and that the research went through ETH Zurich ethics review. Decrypt also reported a paper estimate of roughly $1 to $4 per target, a figure that should not be read as a universal price for identification attempts.

This story draws on original reporting from Decrypt.