An old internet joke went, “On the internet, nobody knows you are a dog.” Large language models are ruining the punchline. In a recent study, AI worked out who was behind pseudonymous accounts on the tech forum Hacker News using only public information. When it ventured a guess, in 252 of 338 cases, it was right 90% of the time. Jackie Snow asked co-author Florian Tramèr, an assistant professor of computer science at ETH Zurich, what that means.
How the crumbs connect
No single clue does the job. Tramèr’s example: one post says the writer studied at MIT in 1997, another mentions a dog named Bluey. Each detail feels harmless. Put them together, and it’s plainly the same person.
The test used Hacker News users who had posted LinkedIn profiles in their bios; researchers stripped out identifying details before the AI went to work. Job history was often the giveaway: how long someone worked somewhere, or where they studied. By our arithmetic, 90% of 252 guesses is about 227 correct matches, roughly two in three of the whole group.
On Reddit (RDDT), a shared taste in obscure films was sometimes enough to link two pseudonymous accounts. Writing style, despite its reputation, is mostly a supporting signal, Tramèr says. Governments still have better data access, but for public information, the technology is essentially there for anyone.
A retiree’s public trail
Scraping limits are easy to get around, since most are built to stop AI bots rather than people. Truly private accounts are different, though Tramèr says the real risk there is a friend sharing access with an AI agent. People also tend to underestimate how much of their information is public.
Now picture a retiree’s trail — our illustration, not the study’s. A retirement-party photo names the old employer. A birthday post names a grandchild. A travel post says when the house is empty, and the dog’s name sits in a caption. Stitched together cheaply, it’s a script for a caller who seems to know you.
Is anonymity dead? “I don’t think it’s dead,” Tramèr says. But unmasking is now cheap and available to almost anyone, which makes stalking easier and blackmail at scale a real risk. His advice: share less, especially unusual or specific details. Posting fake details is plausible but hard to pull off, because these models are good at catching contradictions.