AI Knows Who You Are: Large-Scale Deanonymization with LLMs and the End of Online Privacy

Day April 25, 2026
Time 14:35–15:00
Room Tiyatro salonu
Duration 25 min
Spoken language Turkish
Format Talks

Kadir ArslanSpeaker profile

For years, we have thought that our online nicknames will keep us safe. We considered that the verification of identities would be too costly. But recent research indicates that this situation is being dramatically shifted by large language models. Starting from unstructured text (anonymous forum posts, comment histories, and even brief interview transcripts) LLM-based systems can now match users to their real identities in minutes and for just a few dollars, automating what would take a skilled investigator hours of manual OSINT work. In this talk, we will be looking at the deanonymization framework as introduced in "Large-scale online deanonymization with LLMs" by ETH Zurich (2026) . We will discuss how successful AI companies and governments would be at finding the person behind the "nickname", if they decided to leverage this capability for user surveillance In a world where hiding behind aliases is no longer enough, we will address how we, as the free software community, need to rethink online privacy.

The talk will consist of three parts. In the first part, I will show with concrete examples how LLMs deanonymize users: the attack pipeline covering feature extraction, candidate search, reasoning, and verification. In the second part, I will evaluate scenarios where this technical capability could be transformed into a tool for systematic surveillance by corporations and governments. In the final part, I will discuss what the free software and open-source community can do to counter this threat — at the level of privacy-focused tools, data minimization, and platform design.