February 25, 2026
AI knows what you did last post
Large-scale online deanonymization with LLMs (using HN posts)
Hackers panic as AI starts doxxing their secret online lives
TLDR: Researchers showed that modern AI can match your “anonymous” posts to your real identity across sites like Hacker News, Reddit, and LinkedIn with worrying accuracy. The community reacted with a mix of fear, jokes about hiding in l33tsp34k, and debate over whether to tighten platforms or empower users to protect themselves.
On Hacker News, the nerds’ favorite hideout, a new study just dropped a bomb: large language models – the same tech behind chatbots – can now play super‑sleuth and reclaim your real identity from your so‑called “anonymous” posts. The paper claims AI can link your casual comments to your LinkedIn or Reddit accounts with scary accuracy, and the community instantly went into a mix of panic, denial, and gallows humor.
One user basically asked the question on everyone’s mind: “Did the AI actually find my real LinkedIn… or some poor stranger instead?” Others weren’t impressed with the usual “lock everything down more” solution. Commenter mhitza pushed back, saying tightening platforms just cages users, and instead begged for tools that warn you when you’re oversharing personal details like your job, location, or politics.
The thread quickly turned weird and wonderful. Someone joked that old‑school l33tsp34k – writing like a 2003 hacker kid – might be the original privacy shield, then imagined whole communities speaking through a single AI “brand voice” so nobody’s individual style can be traced. Meanwhile, another commenter dryly pointed out that most LinkedIn profiles are boring lists, not long confessions, subtly poking holes in the hype. And when one critic got a bit too dismissive, the moderator dang stepped in with a firm “be thoughtful or be quiet,” turning the privacy scare into full‑on comment‑section courtroom drama.
Key Points
- •LLMs can re-identify pseudonymous users at scale using unstructured text and cross-platform content.
- •An LLM-driven pipeline performs feature extraction, embedding-based candidate search, and reasoning to verify matches.
- •Evaluations used three ground-truth datasets: HN-to-LinkedIn links, Reddit cross-community matches, and time-split Reddit profiles.
- •LLM methods achieved up to 68% recall at 90% precision, far outperforming classical non-LLM baselines (near 0%).
- •Findings indicate practical obscurity for pseudonymous users is weakened, suggesting privacy threat models need updating.