March 3, 2026
Mask off, drama on
LLMs can unmask pseudonymous users at scale with surprising accuracy
Your secret alt might not be secret—readers split between panic and “prove it”
TLDR: A new study claims AI can link pseudonymous accounts with striking accuracy, putting online privacy at risk. Commenters split between alarm, legal worries, and swaggering skepticism—some propose hiding in AI-generated noise, while others demand proof and better platform policies before declaring anonymity dead.
Researchers say smart chatbots (large language models) can link burner accounts across sites and guess who’s behind them, with recall up to 68% and precision up to 90%—that’s how many they find, and how often the guess is right. Pseudonymity—using a fake name—has been a safety net for people in sensitive threads. If these bots can connect your posts, your alt might be less safe. Cue the comments: a gloomy chorus declares the web is getting boring, while the skeptics roll their eyes and demand proof. One reader even threw down a bounty: “LLMs aren’t magic,” plus 500,000 JPY if anyone can dox them via posts alone. Privacy diehards warn this could turbocharge doxxing and spark GDPR headaches, with people pointing at Hacker News’ deletion policy and asking if platforms are ready for the fallout. Old-schoolers chimed in with memories of writing-style tools that already linked accounts—so is this new, or just bigger and faster? Then came the chaos camp: “flood the web with LLM slop and hide in the noise,” turning panic into a meme. The biggest clash: OPSEC vs. AI reasoning—is this only dangerous when you slip, or can models truly work from scrubbed text to your real identity, as the paper claims?
Key Points
- •LLMs achieved up to 68% recall and 90% precision in deanonymizing pseudonymous users across platforms.
- •Datasets included linked Hacker News–LinkedIn posts, Netflix micro-identities, and split Reddit histories.
- •Researchers removed identifying references and used LLMs to infer identities from free text.
- •AI agents can browse and reason across web content, reducing the need for structured datasets in re-identification.
- •An Anthropic questionnaire dataset yielded positive identification of 7% of a 125-person sample.