📱 Researchers from the Swiss Federal Institute of Technology Zurich (ETH Zurich) in collaboration with experts from Anthropic have proven that modern language models (LLMs) can effectively reveal the identities of anonymous account owners. This discovery threatens pseudonymity as one of the main ways to protect privacy online.
➡️ Algorithms match user messages on different platforms by analyzing free text and hidden behavioral patterns. In experiments, neural networks were able to successfully identify 68% of anonymous profiles, with the accuracy of identifying individuals reaching up to 90% in some cases.
📍 Key analysis methods:
1. cross-platform connections: AI analyzed profiles on Hacker News and LinkedIn through links. Even after removing all direct mentions of a name or surname, the language model accurately established connections between accounts based solely on the style and content of messages;
2. digital fingerprints (micro-identifiers): Using data about preferences and action history, similar to the famous Netflix Prize dataset, AI reconstructs a person's identity without any direct indicators of personal data;
3. interest analysis on Reddit: Research has shown that discussions on niche topics (for example, films) in various communities are unique markers. If a user comments on more than ten films, the accuracy of their identification rises to 99%.
