Large language models enable mass deanonymization of online users
Research demonstrated that large language models can identify anonymous online users with high accuracy by analyzing writing patterns and cross-referencing information across platforms. The study raised concerns that such capability, if misused, could enable government surveillance of critics, targeted commercial exploitation, and sophisticated social engineering attacks.
What happened
Research showed that large language models can identify anonymous users across platforms by analyzing writing samples. The system extracted identity features, searched for candidate matches, and used reasoning to verify matches while reducing false positives.
In experiments with Hacker News and LinkedIn profiles, the AI maintained 99% precision while raising recall from 0.1% to 45.1%. In Reddit movie discussion forums, accuracy reached 48.1% to 90% when users had discussed multiple films, compared to 3.1% for single-film users.
In a survey by Anthropic about how people use AI, researchers identified 7% of 125 participants based only on general information they provided.
Reported impact
The research indicated that online anonymity protections are no longer effective. If the capability were misused, governments could target online critics, companies could build highly precise advertising profiles, and attackers could launch personalized social engineering scams.
Traditional deanonymization required professional investigators to spend hours collecting structured data, whereas large language models require significantly less time and can operate at much larger scale.
Index analysis
Researchers noted that users who post more frequently are more easily identified, and that even general information can expose identity. Large language models can browse the web like humans and use reasoning to piece together identity signals automatically.
The study indicated that old methods of protecting anonymity online are no longer effective as language models can automatically fetch web pages and verify extracted identity signals.
Governance controls
Researchers recommended that platforms limit API access rates and detect automated scraping behavior. Large language model providers should establish safeguards to refuse deanonymization requests.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.