Your writing style is a unique fingerprint. Researchers from ETH Zurich and Anthropic built an AI pipeline that correctly identified 226 of 338 anonymous users — 67% accuracy — with 90% precision when confident, at a cost of just $1–$4 per person.
How Does AI Figure Out Who You Are?
A team of researchers from ETH Zurich and Anthropic developed a four-stage pipeline called ESRC — extract, search, reason, calibrate — that uses large language models to connect anonymous online posts to real-world identities.
Here's how it works: the system first extracts stylistic and behavioral features from anonymous posts — things like word choice, punctuation patterns, and topic preferences. Then it uses semantic embeddings to search a large pool of candidate profiles for potential matches. Next, an LLM reasons over the top candidates, comparing writing patterns side by side. Finally, it calibrates its confidence, only making an identification when it's sufficiently sure.
The researchers tested this on 338 Hacker News users who also had public LinkedIn profiles. After stripping all obviously identifying details from the posts, the system correctly matched 226 of those users to their real identities — a 67% recall rate. When the system expressed high confidence in a match, it was right 90% of the time.
At a larger scale, the researchers also tested against a pool of 89,000 candidates and still achieved 55% accuracy — a remarkable result given the size of the haystack.
What Makes Your Writing a Fingerprint?
The system doesn't rely on one single giveaway. It builds a composite profile from semantic embeddings — mathematical representations of meaning and style. Your word choices, punctuation habits, the communities you participate in, your argument structure, even your emoji preferences all contribute to a unique signature.
The researchers also validated their approach on the Anthropic Interviewer Dataset, successfully re-identifying 9 of 125 scientists — a smaller success rate, but notable because these were trained researchers who presumably had more awareness of privacy.
This study is a turning point. We've moved from "AI might be able to identify you" to "AI can identify you, cheaply, at scale." The 90% precision means when AI is confident it knows who you are, it's almost certainly right. The question isn't whether this technology will be used — it's who will use it first, and whether we'll have any protections in place when they do.
What Does This Mean for You?
The researchers deliberately did not release their code, recognizing the dual-use potential. But the methodology isn't secret — it's built on commercially available LLMs that anyone can access. The implication is clear: anonymity is becoming a technical illusion.
No hacking is needed, no data breach, no insider access. Just AI reading your publicly available posts and matching patterns. If you've ever posted anonymously and also have a public professional profile, the connection may already be technically possible to make.
This isn't a future threat. The tools exist today. The only question is scale — and at $1–$4 per person, the economics are already viable for anyone motivated enough to try.
AI can now match your anonymous posts to your real identity using writing style alone — with 90% precision at a cost of $1–$4 per person. Online anonymity as we know it is fundamentally changing.
Source: Lermen, Paleka, Swanson, Aerni, Tramèr & Carlini. "Large-Scale Online Deanonymization with LLMs." arXiv, February 2026. arxiv.org/abs/2602.16800