← Back to feed

A pseudonym used to be practical protection. Linking an anonymous forum handle to a real person required hours of manual trawling. New research eliminates that bottleneck.

A paper titled “Large-scale online deanonymization with LLMs” (arXiv 2602.16800) shows that a three-stage LLM pipeline can identify anonymous users at scale with 68% recall and 90% precision — meaning that when the system flags a match, it is correct nine times out of ten. Prior automated methods hover near 0% on the same task.

The pipeline

The system works in three stages:

1. An LLM reads raw posts and extracts structured identity clues — location signals, technical vocabulary, recurring interests, writing quirks, references to life events.
2. Those clues are used to retrieve a ranked candidate pool from a target platform.
3. A reasoning step compares the top candidates against the source profile and rejects weak matches.

The reasoning layer is what makes it work. Earlier systems either retrieved too broadly or matched too naively. The LLM’s ability to weight inconsistent evidence — someone mentions Melbourne in one post but UTC+5 in another — and still arrive at the correct identity is the core capability.

Test cases

The researchers ran three experiments:

• Hacker News → LinkedIn: Match anonymous technical forum users to their professional profiles.
• Reddit movie community → Reddit sports community: Link the same user across unrelated communities with different pseudonyms.
• Reddit users across time: Match a user’s posts from different time periods against each other.

All three demonstrated significant accuracy gains over baselines. The cross-platform case — linking a person’s anonymous handle to their real-name LinkedIn — is the commercially and legally consequential one.

Why this matters now

The study’s framing is that pseudonymity was never cryptographic privacy — it was just friction. The identity hints have always been present in public writing. What changed is that LLMs can now extract and reason over those hints at scale and at low cost.

Practically: any sufficiently prolific online poster — technical forums, subreddits, Discord servers, GitHub issues — is now linkable to other accounts by an automated system that anyone with API access can run. The system performs better on users who post frequently and across varied topics, which describes most people who have built a reputation on a platform.

The implications span security (tracking activists or journalists through forum posts), commercial intelligence (linking competitor employees to anonymous job forums), and legal discovery (mapping pseudonymous accounts to litigation targets).

Current defences — writing style changes, geographic obfuscation, account rotation — reduce but don’t eliminate the attack surface, according to the paper. The more substantive long-term defence is architectural: platforms that never surface raw user text to third parties, or that apply local differential privacy to posts before they leave the platform.

That is not the design of any major platform today.