The Digester

A new study shows large language models can often deanonymise users by linking small public clues across sites, prompting calls for tighter data controls and revised anonymisation practices.

LLMs matched anonymous accounts to real identities in most test scenarios by correlating small public details across platforms.
The automation and low cost of these attacks make targeted surveillance and highly personalised scams more feasible.
Researchers Simon Lermen and Daniel Paleka recommend platforms restrict bulk data access, rate limit downloads, and detect automated scraping to reduce risk.
Experts warn LLMs can produce false links and that non-social public records like hospital or admissions data may no longer be adequately anonymised.

Study: AI can identify anonymous social media accounts

Sources