Understanding what the 90% figure really means
The study Large-scale online deanonymization with LLMs, by Simon Lermen, Daniel Paleka, Joshua Swanson, Michael Aerni, Nicholas Carlini and Florian Tramèr, was presented at USENIX Security 2026. Their Hacker News experiment selected 338 people who had publicly linked LinkedIn profiles. This design provided a way to measure performance against known answers while limiting unnecessary privacy harm.
The Wall Street Journal reported that the system guessed 252 identities and was correct 90% of the time. The public paper describes 226 correct identifications, 25 incorrect ones and 86 abstentions. That represents approximately 90% precision among attempted identifications and 67% recall across all 338 targets. Neither figure establishes a universal probability that any particular anonymous person will be exposed.
The authors acknowledge that the evaluation profiles may be easier to identify than genuinely pseudonymous accounts because their original identities were publicly linked. Results also depend on available information, the search environment and the candidate pool. The distinction between a controlled experiment and the varied conditions of everyday online life is essential to responsible reporting.
I don’t think it’s dead.

















