AI System Shows Ability to Reidentify Anonymous Online Accounts
Key Points
- Researchers built an AI system using large language models to link anonymous online accounts to real identities.
- The system achieved up to 68% correct matches with 90% precision, far outperforming traditional methods.
- Testing used public datasets from Hacker News, LinkedIn, and Reddit, among others.
- The entire experiment cost less than $2,000, or between $1 and $4 per profile analyzed.
- Results highlight growing risks to online anonymity as AI capabilities expand.
- Authors recommend user caution and call for AI labs and platforms to implement safeguards.
Researchers from ETH Zurich, Anthropic and the Machine Learning Alignment and Theory Scholars program have built an automated AI system that can link pseudonymous online profiles to real identities. Using large language models to analyze writing style, posting patterns and other clues, the system correctly matched up to 68 percent of accounts with 90 percent precision, far outpacing traditional methods. The experiment cost only a few dollars per profile, highlighting a low‑cost barrier for large‑scale deanonymization. The study warns that online anonymity may be less secure than many assume, especially as AI capabilities continue to improve.
New AI Approach to Deanonymization
Scientists from ETH Zurich, Anthropic and the Machine Learning Alignment and Theory Scholars program developed an automated system that uses large language models to identify the owners of anonymous online accounts. The system treats each post or text as a collection of clues—writing quirks, biographical hints, posting frequency and timing—and searches massive datasets for matching patterns.
Performance Compared to Traditional Methods
Testing on publicly available datasets, including posts from Hacker News, LinkedIn, and Reddit, the AI‑driven approach correctly identified up to 68 percent of matching accounts while maintaining 90 percent precision. By contrast, comparable non‑LLM techniques identified almost none of the matches. Success rates rose when more structured information was available; for example, linking Reddit users who mentioned ten or more movies succeeded nearly half the time, whereas linking those who mentioned only one movie succeeded about 3 percent of the time.
Cost and Accessibility
The researchers reported that the entire experiment cost less than $2,000, translating to between $1 and $4 per profile analyzed. This low cost dramatically lowers the barrier for entities wishing to conduct large‑scale deanonymization.
Implications for Privacy and Security
The findings suggest that the assumption of safety behind pseudonymous posting may be increasingly unreliable. While the study was conducted under controlled conditions with curated datasets, the researchers caution that real‑world applications could become more effective as AI models improve and gain access to larger data pools. They recommend that users limit personal details, avoid identifiable posting patterns, and keep accounts separate to mitigate risk.
Calls for Safeguards
The authors argue that responsibility should not rest solely on users. They suggest that AI developers monitor how their tools are employed and that social media platforms curb mass data scraping that enables such deanonymization efforts. Despite the advances, the researchers note that current techniques are still far from the capabilities of a skilled human investigator, and that anonymity for high‑profile targets—such as the creator of Bitcoin—remains intact.