AI Models Identify 68% of Anonymous Internet Users with 90% Accuracy
Researchers at ETH Zurich found that AI language models including Gemini and ChatGPT can identify 68% of anonymous internet users with 90% accuracy by analyzing public posts from forums like Hacker News and Reddit. The finding raises concerns that online anonymity is no longer sustainable and that such AI capabilities could be used for surveillance of personal beliefs and opinions.
What happened
Researchers at ETH Zurich analyzed thousands of posts from anonymous forums including Hacker News and Reddit and asked multiple AI language models to identify the authors. Gemini and ChatGPT identified 68% of anonymous users with 90% accuracy in minutes, compared to nearly 0% for the best non-language-model method.
Reported impact
Daniel Paleka, a researcher at ETH Zurich and co-author of the research, stated that people often express opinions on pseudonymous accounts assuming they will remain private, and having a research or surveillance mechanism using language models that can reveal beliefs, political opinions, insecurities, or other information collected from an anonymous Reddit account could significantly deprive ordinary people of power.
Index analysis
The research used a limited database for ethical reasons and because researchers needed to know the real person behind forum comments. They selected user profiles from Hacker News with connected LinkedIn accounts, anonymized them, and asked AI to identify biographical or personal details such as location, profession, hobbies, demographics, or values.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.