Researchers from ETH Zurich, Anthropic, and the MATS program have developed an AI method that can unmask anonymous online users by analyzing their writing patterns and matching them across vast datasets. The 2024 study demonstrates how large language models can systematically extract identifying features from text and determine real identities with high accuracy, effectively eroding the privacy protections that have historically shielded internet users, particularly vulnerable populations.
The breakthrough technology achieved remarkable success rates in real-world tests. The AI agent successfully re-identified 67% of 338 Hacker News users at 90% precision by autonomously using web search tools, according to the research published on arXiv. In closed-world scenarios linking profiles between platforms, the system achieved 45.1% recall at 99% precision when matching Hacker News accounts to LinkedIn profiles, dramatically outperforming classical methods that scored only 0.1% recall.
The system employs a four-stage pipeline called ESRC (Extract, Search, Reason, Calibrate) that processes unstructured text to extract identity-relevant features. The technology utilizes prominent language models including the Gemini family, OpenAI’s GPT series, and Grok 4.1 Fast to analyze demographics, interests, and writing patterns. The agent then performs nearest-neighbor searches across massive databases before reasoning about probable matches and assigning confidence scores.
Vulnerable Populations at Risk
The erosion of online pseudonymity poses immediate threats to groups who depend on anonymity for safety. Whistleblowers and journalists risk retaliation for exposing wrongdoing, while activists could face persecution from authoritarian regimes. Survivors of abuse who use pseudonyms to escape their abusers and individuals exploring sensitive identities now face unprecedented vulnerability to automated doxxing campaigns and state-sponsored surveillance.
In large-scale testing against 10,000 Reddit profiles, the system identified a third of all users at 99% precision and achieved 68% recall at 90% precision. Researchers project the agent could maintain 45% recall at 90% precision even against a database of one million users, demonstrating scalability that amplifies privacy concerns.
The research team practiced responsible disclosure by withholding their code, prompts, and datasets. According to The Verge, social media platforms are encouraged to “clamp down on the scraping and mass data extraction” that enable these attacks. AI labs face pressure to implement safeguards preventing their models from being weaponized for deanonymization purposes.
While Luc Rocher of the Oxford Internet Institute cautions that experiments were conducted under laboratory conditions and robust privacy tools can still offer protection, the fundamental shift in online privacy capabilities demands urgent action. The collaborative response must involve platforms strengthening data access policies, AI developers monitoring tool usage, and users practicing greater caution about information shared across different online contexts.
Sources
- The Verge
- arXiv
