LLMs can unmask anonymous social media users, study warns

This study is a wake‑up call that LLMs aren’t just chatbots; they’re powerful inference engines that can stitch together our online breadcrumbs in ways humans never could. By showing that a model can take an ostensibly anonymous account, scrape its content and link it to a real identity on another platform, the researchers are demonstrating a new class of privacy attack that is basically “open source intelligence on steroids.” Once this technique exists in public, every repressive regime, data broker and scammer will be tempted to use it.

For the race to AGI, it highlights that capability gains don’t neatly separate from misuse. The same general‑purpose reasoning and pattern‑matching that makes AGI plausible also makes it trivial to pierce anonymity at scale. That will drive demand for AI‑resistant privacy tools, differential privacy by default, and perhaps even regulatory bans on certain forms of automated correlation. It may also pressure major labs to incorporate privacy‑preserving training regimes, not just content filters.

Strategically, this pushes platforms and policymakers toward a harder stance on data minimisation: if any public trace can be fused into a dossier by an LLM, then controlling what’s published and how long it persists becomes a national-security issue, not just a UX choice.