PSEUDONYMITY on the internet may be less durable than many users realise, with researchers showing that large language models can deanonymise burner accounts at scale. In experiments, recall reached as high as 68 percent and precision up to 90 percent when linking posts and identities across platforms, a finding that could upend conventional privacy assumptions.
The study used datasets from public sites such as Hacker News and LinkedIn, Netflix micro-identities, and Reddit histories, applying an LLM to anonymised text before attempting to identify the speaker. One experiment found that a simple questionnaire given to participants allowed 7 percent to be positively identified at a 90 percent precision level.
The researchers also reported that, across cinema-related discussions in 2024, the probability of identification rose with the number of shared movies, admitting notable increases in recall at higher precision thresholds. They suggest mitigations including rate limits on data access and model guardrails, while noting the broader implications for privacy, online influence, and targeted social engineering. 3 March 2026. Dan Goodin.