What datasets did researchers use to test LLM deanonymization?

Hacker News posts linked to LinkedIn, Netflix micro-identities, and split Reddit histories—all public, scrubbed of direct IDs.

How accurate are LLMs at unmasking pseudonymous users?

Up to 68% recall (catching targets) and 90% precision (correct guesses)—way beyond human methods.

Pretty much; LLMs invalidate easy pseudonymity. Obfuscate hard or go dark.

☁️ Cloud & Databases

Burner accounts? Think again. LLMs are linking your anonymous rants to your real identity faster than you can say 'doxxing.'

theAIcatchup Apr 09, 2026 3 min read

LLMs achieve 68% recall and 90% precision in deanonymizing pseudonymous users across platforms. 𝕏
This shatters the assumption that throwaway accounts protect privacy, enabling doxxing and profiling. 𝕏
Expect a privacy arms race: zk-proofs and obfuscation tools incoming to counter LLM fingerprinting. 𝕏

Published by

Community-driven. Code-first.

#LLMs #deanonymization #online privacy #pseudonymity

Get the best Open Source stories of the week in your inbox — no noise, no spam.

Originally reported by Ars Technica - Tech