Live data from Hacker News

Show HN: Using stylometry to find HN users with alternate accounts

stylometry.net

271–280 of 562 posts

Re: Show HN: Using stylometry to find HN users with alternate accounts

#271
post #57

ive had maybe a hundred throwaway accounts on HN over the past ten years. generally, i make an account, say something that is apparently wildly offensive to someone else, get flagged and down-voted and then muted or hell-banned. then i make another account because i never did anything wrong and start the process over again. ive emailed the admins, tried to reason with the admins, it never does any good. the power is…

You really don't need advanced AI to do it. Just a bunch of scrapers and some run of the mill statistics. And guess what, it's been done by many companies already. They just don't care to create such a site.

Re: Show HN: Using stylometry to find HN users with alternate accounts

#273
Over in the D language forums, we welcome people who post under a pseudonym, and our policy is we won't allow attempts to unmask them.

This is to protect high profile users who are secretly enjoying programming in D rather than the language they are supposed to use.

And, of course, to protect users who feel they might be discriminated against if their background was known.

Re: Show HN: Using stylometry to find HN users with alternate accounts

#274
post #106
post #90

nice one. are you using gpt3 under the hood?

I'm not that smart - my site is basically just doing some calculations on word frequencies. You can read https://academic.oup.com/dsh/article-abstract/17/3/267/92927... and https://www.tandfonline.com/doi/abs/10.1080/09296174.2011.53... and https://news.ycombinator.com/item?id=33755898 for more information.

I am curious whether it could pick GPT3 out of the crowd.

Re: Show HN: Using stylometry to find HN users with alternate accounts

#275
post #158

This is somewhat similar to how they ended up catching the Unabomber. The FBI were literally at a dead end. They ended up posting one of his letters/manifestos in the paper, somebody recognised a turn of phrase the unabomber used that was unusual and reported it as possibly being their brother, FBI investigated the lead and it lead them straight to him. Excerpts from wiki: > Before the publication of Industrial Socie…

So is the lesson you should have GPT rewrite your manifesto so as to obscure your personal idioms?

Or something purpose-built like Anonymouth (https://github.com/psal/anonymouth), although it seems to be both unique and dead.

Also interesting:

> Ross Ulbricht aka Dread Pirate Roberts, the mastermind behind the infamous Silk Road site which served as a black market for drugs, weapons and fake documents was also well aware of the potential danger of stylometry being used against him. At the time of his arrest in a San Francisco public library, the FBI captured images of his laptop screen as evidence. Guess what what he had bookmarked — “Science of Stylometry.”

https://medium.com/svilenk/the-case-for-anonymity-12db114f0c...

Re: Show HN: Using stylometry to find HN users with alternate accounts

#278

This is a super interesting tool for self reflection. Looking at the top 10 similar accounts to mine, it gives me an arms-length view of how other people probably interpret my tone. I appear to be a well-educated, over-confident know-it-all.

My #3 match is cstross, and now I’m convinced that my life-long secret dream of being a successful sci-fi novelist is basically a matter of typing. (Ideas? Character development? Ruthless editing? Developing an audience? Having a publisher? What do I need of those when the Computer told me I’m practically a genius…)

Re: Show HN: Using stylometry to find HN users with alternate accounts

#279

Over in the D language forums, we welcome people who post under a pseudonym, and our policy is we won't allow attempts to unmask them. This is to protect high profile users who are secretly enjoying programming in D rather than the language they are supposed to use. And, of course, to protect users who feel they might be discriminated against if their background was known.

It's very important for those people to be aware of these style analysis attacks! Glad this post is raising awareness.

Re: Show HN: Using stylometry to find HN users with alternate accounts

#280
post #106
post #90

nice one. are you using gpt3 under the hood?

I'm not that smart - my site is basically just doing some calculations on word frequencies. You can read https://academic.oup.com/dsh/article-abstract/17/3/267/92927... and https://www.tandfonline.com/doi/abs/10.1080/09296174.2011.53... and https://news.ycombinator.com/item?id=33755898 for more information.

As you mention on the site, you don't do punctuation. But I'm guessing there are some pretty good fingerprints like:

two spaces after a period

Whether someone uses an em-dash/single hyphen/double hyphens (which may correspond to house style they're used to)

Whether they use semi-colons

(Presumably harder) but consistent substitutions like loose for lose, break for brake, etc.

Use of accents

Post reply on HN