The archive · Product Ideas · Product decision · 2022
Stylometry.net matches HN users by writing style to expose alternate accounts
A $8-a-day site fingerprints every Hacker News commenter by n-gram style so users can see their own abandoned alt accounts — 676 points and 511 comments.
costco (HN user who built stylometry.net)
What it had to solve
Hacker News accounts are pseudonymous, but commenters write often and in long form, which makes style a fingerprint. User costco built stylometry.net to show how easily anyone — not just a well-funded lab — could link alternate accounts and erode that anonymity.
How it works
On 26 November 2022 a Hacker News user posting as costco launched stylometry.net with a pointed message: enter any HN username and the site returns the users whose writing style is most similar, surfacing likely alternate accounts. The author said it had already confirmed accounts he suspected were alts and identified abandoned accounts of people he knew — and that he built it mainly to show how easily anonymity erodes.
The method was deliberately simple. He kept the 200 most common word and character n-grams of sizes one to three for each user, turned them into frequency vectors, and compared every user against every other using scikit-learn's pairwise distance functions. The entire HN comment corpus fit in RAM on a mid-tier laptop, and the whole demonstration cost about $8 a day of dedicated server time — his point being that if one hobbyist could do this, a funded organization with linguists could do far more.
The thread drew 676 points and 511 comments. Multiple users confirmed the tool found their own old accounts — one found about ten across the years, another rediscovered a pre-2018 username — while users without alts hit false positives, fueling a rich debate on similarity thresholds, mutual top-20 'bold' matches as a precision signal, and whether listing example usernames without permission was acceptable.
Why it lands
- HN's long-form, frequent comments make style unusually measurable, so a corpus-wide n-gram comparison works without any machine learning sophistication.
- Showing results on real accounts made the threat concrete — users did not have to take the author's word that stylometry worked.
- The $8 server framing delivered the political point: cheap tooling, not just governments, can link pseudonymous identities.
- Highlighting mutual top-20 pairs gave users a precision signal inside a noisy ranking, which is what made the demo feel uncannily accurate.
What it did
The Show HN drew 676 points and 511 comments on 26 Nov 2022. Users reported finding their own abandoned accounts ('all ~10 of my old accounts'), old usernames and pre-2018 alts, while users with no alts saw false positives — opening a long debate about thresholds, countermeasures and consent.
What you can take
Do the attack, then show it: a cheap working demo of de-anonymization changed the privacy debate more than warnings could, because anyone could run it on their own account.
Since then
The author removed example usernames after users objected and kept defending thresholds in the thread: scores above 0.9 were near-certain matches, while bolded mutual top-20 results were the strongest precision signal, though far from perfect. Commenters connected the demo to academic stylometry, Whonix's evasion guidance, earlier authorship attempts from 2018, and even the Unabomber case, making the thread itself evidence that long-form pseudonymity is fragile.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe the brief you are staring at, and see who has been given the same problem.
Free account · 3 free questions · no card