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

Wow. This gives a lot of false positives, but it found all ~10 of my old accounts over the years.

The most interesting thing is that my writing style changed pretty drastically since a decade ago. Searching for my oldest account matches my earliest usernames, whereas searching this account matched the rest.

The details of the algorithm are fascinating: https://stylometry.net/about Mostly because of how simple it is. I assumed it would measure word embeddings against a trained ML model, but nothing so fancy.

The method used, i.e. to calculate the cosine of the two authors' word vectors, is poorly suited for stylometric analysis because it is based on a poster's lexicon and the word frequencies of each word, but ignoring stylistically relevant factors like word order.

Also, the cosine of the vectors of word frequencies conflates author-specific vocabulary and topics; in other words, my account is grouped (with >51% similarity, according to the demo) with someone probably because we wrote about similar things. A strong stylometric matcher ought to be robust against topic shifts (our personal writing style is what stays constant when we move from writing about one topic to writing about another topic, just like our personality is what stays constant about our behavior over time - of course styles do change, but the premise then has to be that such changes happen very slowly).

Stylometrics/authorship identification is interesting and has led to some surprising findings, e.g. in forensic linguistics (Malcolm Coulthard wrote several good books about the topic).

This paper lists some other features that could be used and compares a bunch of techniques: https://research.ijcaonline.org/volume86/number12/pxc3893384...

Ha, gruseom shows up for pg, which is dang’s old account. A worthy successor.

This is a fascinating way to find similar HN users who aren’t the same person. It’s a surprisingly great recommendation engine. “If you like pg, you might also like…”

Sure, the privacy concerns are valid, but the cat’s out of the boot. Might as well enjoy the benefits.

montrose is almost definitely pg. Someone who talks about ancient history, Occam’s razor, VCs and startups, uses the phrase “YC cos” (relatively uncommon), etc. https://news.ycombinator.com/item?id=17112567

Nicely done. One of the best hacks I’ve seen in a long time.

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 Society and Its Future, Kaczynski's brother, David, was encouraged by his wife to follow up on suspicions that Ted was the Unabomber.[91] David was dismissive at first, but he took the likelihood more seriously after reading the manifesto a week after it was published in September 1995. He searched through old family papers and found letters dating to the 1970s that Ted had sent to newspapers to protest the abuses of technology using phrasing similar to that in the manifesto.[92]

> In early 1996, an investigator working with Bisceglie contacted former FBI hostage negotiator and criminal profiler Clinton R. Van Zandt. Bisceglie asked him to compare the manifesto to typewritten copies of handwritten letters David had received from his brother. Van Zandt's initial analysis determined that there was better than a 60 percent chance that the same person had written the manifesto, which had been in public circulation for half a year. Van Zandt's second analytical team determined a higher likelihood. He recommended Bisceglie's client contact the FBI immediately.[96]

> In February 1996, Bisceglie gave a copy of the 1971 essay written by Ted Kaczynski to Molly Flynn at the FBI.[87] She forwarded the essay to the San Francisco-based task force. FBI profiler James R. Fitzgerald[98][99] recognized similarities in the writings using linguistic analysis and determined that the author of the essays and the manifesto was almost certainly the same person. Combined with facts gleaned from the bombings and Kaczynski's life, the analysis provided the basis for an affidavit signed by Terry Turchie, the head of the entire investigation, in support of the application for a search warrant.[87]

https://en.m.wikipedia.org/wiki/Ted_Kaczynski

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.

After a few tries on boring accounts, I thought to try the account of somebody who was notorious for an incident outside of HN, and had a (deservedly) bad time at HN for a couple of years before the account went dark.

And yeah, there's a bunch of high confidence (.6-.8) hits for that account, and from a quick browse of the comments of the recently active ones, they look really likely to be alts. Like, all three that I looked at had comments that made it very clear it was this person writing pseudonymously. (E.g. writing on their signature issue, and saying they couldn't go into more detail due to fear of self-doxxing; or somebody literally saying that the alt's claims reminded them of the public writings of the notorious guy years ago).

Obviously I'm not naming the account, but this functionality turned out way creepier than I thought the moment I tried it on the account of somebody who has a reason to disassociate from an existing public persona, but still wants to participate here.

It would be nice to make the names clickable.

I don't think the list of pg alternate account is accurate. I checked a few. They have many oneliners that is typical of pg, but the topics and style don't look similar.

I searched a few more and got better results. :)

I searched myself (that I know that I have no alternate accounts). I recognize a few users that are interested in similar topics, and I discuss/upvote them many times. But I didn't recognize most of the user of the list.

Cool stuff, thank you for sharing your findings!

I don't do throwaway. I either post or STFU. I also STFU on darknet. Its why I found it fun to read/lurk on things like I2P back when it was new. And I know that on a pseudonymous account it is only a matter of time until it can be linked to another pseudonymous account. It would not surprise me if stylometry was used on Dread Pirate Roberts or the people behind The Pirate Bay or the people behind Wikileaks (Assange's sockpuppet accounts). Such can also have been used to verify afterwards instead of beforehand. Though with TPB since it was on clearweb an advanced adversary could have used correlation/timing attack to figure who wrote what.

I'm having fun times recognizing other Dutch people though their usage of English language. For example, a distinctive word I see Dutch people use a lot is 'oke' instead of 'OK' or 'okay'. Its a red flag the person is native Dutch. I wonder if there are stylometry tools available for figuring if someone used physical vs touchscreen keyboard (I used Glider to write this post, spellchecker unavailable).

And yes, organizations like secret service and police should use such tools as well. It is a known tool, why not use it for good? As with any tool, it can be used for good and evil. On HN this could be useful for the mod team (AFAIK nowadays only dang) to find banned people's sockpuppets. Cross-community could also be a fun project: find a HN user's Twitter or Reddit account. And I hope this method is also used to find Russian trolls on social media.

The top hit on my list looked familiar. I looked at their recent comments and saw a discussion between that user and me. We were quoting eachother directly throughout.

I wonder if this explains our similarity. And if so, could we tweak the algo by e.g. Removing text that is prepended with ”>”

The scary thing is that once you have this data, finding HN matches for individual targeted users on other sites becomes trivial, even if those sites are harder to scrape. I bet most people here have an anonymous Reddit account, for example. If you wanted to know who was behind a particular Reddit account, you could feed it into something like this and compare the results with HN, where accounts are less likely to be anonymous. Or build a database based on blogs, Github comments, etc.

Also, since this uses only word frequency, there are probably relatively easy improvements to make that would make it even more powerful, like looking at particular runs of words that are unique. Some expressions or figurative language only show up in combinations of words, and tend to be highly style specific.

Forget the alternate accounts — if two users are close in style, there’s a decent chance they should be friends. This is an HN friendship machine.

It would be convenient if the usernames linked to the comment pages on Hacker News (to avoid having to copy/paste and URL hack, which is made even slightly more annoying because for some reason when I tap and hold the usernames to copy them your markup--I haven't looked at why yet--is causing an extra space character to get copied on the left).

This is interesting.

I'm 0.566 correlated with logfromblammo -- and while we are definitely not the same person, I could easily imagine writing a sentence such as:

"For some bizarre reason, management has not yet assigned a task to their programmer underlings to automated themselves out of existence. I can't imagine why."

which is theirs, not mine, from about a year ago. I like that.

On the other hand, I'm nearly as correlated with peterwwillis: 0.5485 -- who has no comments and no submissions.

How about this for countermeasure:

As you're typing out a comment the software gives you a list of accounts you're becoming similar to. That way you can adjust your writing as you type.

This found an alt that I created specifically to see if I could write artificially to defeat this kind of analysis. I have seen other tools like it posted to HN, but none before had found that account. I guess I need to up my game.

See also: https://serhack.me/articles/unveiling-anonymous-author-stylo...

Ahhh, anyone remembers this hacking crew who leaked BLUEETERNAL and other NSA tools and exploits? Shadowbrokers.

They were always communicating in some kind of meme-russian, and their texts were funny to read. [1]

I believe their writing mostly defeated this kind of analysis, at the cost of looking like idiots (which was probably the reason no one sent them crypto-dollars to buy that stuff exclusively).

Here's an excerpt:

"Attention government sponsors of cyber warfare and those who profit from it !!!!

How much you pay for enemies cyber weapons? Not malware you find in networks. Both sides, RAT + LP, full state sponsor tool set? We find cyber weapons made by creators of stuxnet, duqu, flame. Kaspersky calls Equation Group. We follow Equation Group traffic. We find Equation Group source range. We hack Equation Group. We find many many Equation Group cyber weapons. You see pictures. We give you some Equation Group files free, you see. This is good proof no? You enjoy!!! You break many things. You find many intrusions. You write many words. But not all, we are auction the best files."

[1] https://archive.ph/20160815133924/http://pastebin.com/NDTU5k...

Have you tried including parts of speech (for example, as bigrams and trigrams) as part of the features considered in your model? I’ve had great success with stylometry that goes beyond TF-IDF with bags of words; including grammar patterns was shockingly good.

(FWIW, it didn’t find my throwaways; my own model didn’t, either, because I knew that word choice wasn’t enough to avoid being outed by stylometry)

Edit: by bigrams and trigrams, I mean reducing word to their parts of speech labels and using THOSE as word tokens. You’ll find that native English speakers have higher weights on some phrase construction patterns than, say, folks from Romania. TF-IDF is useful for these POS-grams (just made that word up) as well.

How long until this becomes the algorithm for a dating site?

“Find hot single women who write just like you”

This is one reason why I like legal doctrines such as "beyond a reasonable doubt." Even a 0.9 match in a tool like this could be a coincidence, if there are millions of users. But that won't stop people from casually believing "aha it must be an alt account", based on some anecdata.

It's so easy for something like this to be turned into a tool for a witch hunt, targeting innocents.

Hmmm, doesn't seem to work. But you have convinced me (and many others?) to search our alts consecutively and so now do know who has alts?

I wonder what's a reasonable threshold for "probably the same person". I've never had an alt on HN, and when I searched myself, it found 3 other users above 0.6, none of whom I've ever heard of before.

Oh wow, it's really sure that I'm stavrosk, which I am:

https://stylometry.net/user?username=stavros

The next person is 30% less certain, that's huge! This would basically identify any alt I might have with near certainty.

This is an evil website. We won’t have any anonymity soon. The highest match is my years old banned account that I forgot about. Where did you get the data from?

Interesting that the Op doesn't come up in the search: https://stylometry.net/user?username=costco

My nearest match is only at 0.406. It'd be interesting to see who the most unique commenters are, but it's also quite possible it wouldn't be flattering.

I have no alternative accounts besides making a single throwaway account to post one "Ask HN" five years ago, but I have a decent number of matches above 0.5. I think this is due to the relatively uniform style of "who is hiring posts," since my matches did that in a similar way for other companies. I made many of those for about two years when I was at a start-up.

On the how to avoid section: Isn't running comments through a randomised translator a few times then back considered a countermeasure also?

Also think it's probably poor form to list users as examples without their permission.

It found my old account (ara4n; i lost the password) at 0.63. More amusingly it found my cofounder too, who hardly ever posts here (at 0.48)

> ... This site works primarily by analyzing for each user the frequencies of the most common words and phrases in the English language. Accordingly, the easiest way to avoid being identified is to simply use different words than you ordinarily would when writing. More sophisticated models than the one I made can use punctuation, comma usage, and capitalization to identify you so try alternating those as well. Services like Quillbot can help with you this but depending on your circmstances you may not want to send your writings to a third party service.

HN offers many other threads which could be tied together, including:

- time of posting

- ratio of replies to top-level comments

- comments being mainly upvoted or downvoted

- sentiment (mostly angry, dismissive, questioning, etc.)

- most common topics (keyword analysis of post being replied to)

- ratio of new posting to post replies

- first-to-comment on a post

- lone comment on a post

- etc...

It seems very likely that sooner or later every pseudonym for posting content will get discovered and linked. The lesson here is don't post anything that would cause you undue shame or harm if linked directly to your legal name.

Well now I'm self conscious about my closest match being an 0.34 when so many other people are reporting much closer matches with accounts that aren't alts. Do I write weirdly?

What does the bold signify? For example when I search for dang (https://stylometry.net/user?username=dang) the 4th most likely user is not bold whereas the 16th is?

And this is why I’m a reader and not a poster on HN :)

The second that I found out that requesting deletion of an account and its posts needed a MANUAL request to a single user (dang) I noped out so fast

But happy that the rest of you are still happy to contribute :)

I really liked the informative and straight-to-the-point about page - describing how the algorithm works in a way that is easy to understand. All the important details are summarised there. Well done!

Edit: From the "How to avoid .." page, there is the following sentence:

> Also, most authorship identification algorithms have poor accuracy when working with small amounts of words. This means the optimal strategy would be discarding an account either after every comment or after a small number of comments. Unfortunately, this is against HN rules and may result in a ban.

Can you clarify what this means and why it would result in a ban?

Just a heads up that for everyone who doesn't like to link their alt accounts, maybe not use this tool to see if it works.

Unless the author would run this against all HN user accounts, no need to flag the ones "of interest".

Have you done any data analysis on distributions of similarity? How similar you'd expect any 2 people to be given English focused around tech? Or any other interesting stats you'd like to share?

Very nice clean site, great work.

What match level would you expect to see between two randomly chosen individuals?

It's accurate enough that I had to create a new account now :)

I guess it's difficult to evade it as the word frequency certainly catches all about the countries I frequently refer, programming languages, interests etc.

Similar to how they make adversarial fashion[0][1] in order to not be tracked by face id AI, I wonder if we can make adversarial stylometry tools to run your comments through in order to anonymize it

.. [0] https://hackaday.com/2022/10/20/render-yourself-invisible-to...

.. [1] https://adversarialfashion.com/

This is absolutely bonkers. I tried it with my alt and it got my original correct! So I'm writing this comment with a fresh account which hopefully will not get correctly linked too lol

Did something similar in 2018 (still running locally) which could damask anyone

https://twitter.com/austingwalters/status/104189476543920128...

Made both Metacortex.me and insideropinion.com

The idea being you don’t actually need an active directory. It would drop in, figure out all the users (provided one account was on the AD) and would monitor everyone’s skill sets, morale, schedule, etc. Worked super well for what it was / is.

Neat work!

Out of curiosity: do you filter sentences than begin with ‘>’, indicating a block quote from another user? That might improve the accuracy a little here, if you don’t already.

Sorry dang, aka sctb: https://stylometry.net/user?username=dang

I don't have an alt but it would be cool to meet my stylometry-neighbors. I'm curious whether the writing similarity translates to oral communication too

I tried dang's old account (gruseom) expecting to see his dang account listed. Nothing. Tried dang, sctb (a previous admin) was listed as closest match.

I wouldn't rely on these results

https://stylometry.net/user?username=gruseom

https://stylometry.net/user?username=dang

writing "antirez" shows accounts with spanish names (none is mine). I guess Italian and Spanish speakers write very similarly English, but on HN there are a lot more Spanish speakers than Italian ones so that's what I get.

Tried my account thinking "I don't have any alts" but it turns out I do! In 2018 I changed my username from "cbr" to "jefftk" and it pulled that right up: https://stylometry.net/user?username=jefftk

Rebrand it as a soulmate-finder?

Well done, it found my ancient old account.

I only got 0.9999999999999992 for myself :(

Honeypot to see what accounts are tested in sequence?

;-)

Ingenious idea. At the very least, this is just about finding people who write like us, the same way we seek those with similar tastes (music...)

How long before large commercial indexers start offering an efficient (AI based ?) stylometry to agencies and states ?

wait... do you think the NSA is already doing this?

Site down? I'm keen to see if it catches my alts.

Since it looks for similar word usage, false positives seem to appear more often when specific topics are talked about, like stocks or crypto.

Does this ignore stop words? Or do all words have the same weighting? I wonder if only focusing on stop words would give a more accurate measure. Maybe we are more comfortable with certain stop words more than others?

https://en.wikipedia.org/wiki/Stop_words

"Stop words are the words in a stop list (or stoplist or negative dictionary) which are filtered out (i.e. stopped) before or after processing of natural language data (text) because they are insignificant."

1. Interesting. I was kinda expecting to be grouped with other Russian speakers, and I am (based on some nicknames). Probably the frequencies of “the” and “a” are telling. But I swear to God that I sometimes spend some extra time trying to insert as much “the” and “a” in my texts as I could.

2. There is a Russian mnemonic verse, which can’t be properly translated to English, at least it’s beyond my humble capabilities. It goes:

“Это я знаю и помню прекрасно:

Пи многие знаки мне лишни, напрасны”

The number of letters in the words give you the pi number: 3,1415… The meaning is: “I know and remember perfectly: too many signs (positions) of pi are useless and impractical”. Sometimes it’s nice to remember both things.

Nice work! Thank you, of course I plugged in the obvious HN usernames

Edit to add;

Would be nice to have the https://news.ycombinator.com/user?id=username links included.

Amazing and I thought my doxxing tool was terrifying - https://news.ycombinator.com/item?id=32278871

I am afraid to combine all these methods

Incredible! There was a very active throwaway account here a while back that I always enjoyed interacting with. I suspected the person had more than one account and this found one that is incredibly close, down to the topics.

I checked a few random user names and I am confused.

- Why is the author costco[0] not in this lookup?

[0]: https://stylometry.net/user?username=costco

I played a little bit with it and it is baffling how well it finds accounts of people that know each other in real life. So it's not only good for finding alternate accounts but could be used to find peer groups.

This doesn't seem to include text from submissions.

I ran it on Brian Armstrong's temp account from here, and it said it didn't write 10,000 characters:

https://news.ycombinator.com/item?id=3754664

EDIT: Or maybe it's something else because Brian only wrote less than 6k characters. But then why can my account be looked up?

Also, I would guess quoted replies are included, which muddies the analysis. Seems to be a very naive implementation. Much more can be done, but this was probably just a quick project.

How much should we fear de-anonymisation ?

A lot of discussion on the thread are over "how can we prevent this". I would like to know why should we not embrace this and similar technologies?

The benefits in my view are large - online behaviour tracks back to real life - and epidemiology speaking the value of millions of test subjects across every question are invaluable - from traditional medicine to "mass psychology recommendations"

I can guess some downsides (hiding from abusive exes) but am interested in studies, surveys, reports etc - any HN thoughts welcome

Amusingly can't run it on the author since not enough comments

I have only ever had a single account but it returned 19 possibles with no confidence above .54 but 11 bolded. My own account was listed at the top with a confidence of .9999.

Why are some users bold?

The bias is interesting here.

https://stylometry.net/user?username=nickstinemates

Number 2 for me is someone I worked closely with for a few years, and then putting his name into this results in all of the people we worked with for a few years. So it seems content>style, or, we are all more alike than we thought.

I'd be very curious to know if these algorithms can link very different types of text. I'm not surprised that my style is "derivable" on HN, but what if you included my slash-fic pieces, my research papers, etc, would it still "catch" me?

Also, talk about a chilling effect. I was already vaguely aware of this, and now I'm overthinking every word I'm thinking/typing.

Oh god, that thing starts with direct focus on the search field, opening it showed a bunch of old nicknames, I thought it was the result of some study.

The top hit for me, though not a very high correlation (0.3 ish), is to my surprise someone I have met. I don't appear on their top 20 though.

Can we find Satoshi with this?

I interviewed years ago with someone who let me know that they use a pseudonym as an employee and their chosen name even got posted as the author for articles they wrote for the company. They were very concerned about their privacy.

I know their blog, which is their HN username, and this tool found their other account.

Perhaps ironically, this person stood out a lot because of this and I didn't forget them.

It's funny that I only match at 0.9999999999999982 with myself while all other username I tried matched with themselves at 1.0 ^^.

Sticking myself in (I haven't ever had another account) my closest match (at 0.43) is the maintainer of an Open Source project which I have occasionally commented about. They are also British, as am I.

My guess is that as they commonly mention the project and I have on a number of occasions, that has formed the link. Plus maybe usage of common British terms, but that seems far less significant.

It's super interesting!

It would be good if there were more controls to filter the type of words and language that are used for the matching algorithm. So you could say exclude words not in the dictionary. I wander how that would effect my link with this other person.

That’s why I always use throwaway :) everywhere. Reddit. HN. Twitter. Everywhere. I’ll spam every site with my throwaways.

Long live throwaways.

> This means the optimal strategy would be discarding an account either after every comment or after a small number of comments. Unfortunately, this is against HN rules and may result in a ban.

Is this? I thought that it was ok to have throwaway accounts, as long as they're not specifically to avoid a ban or something like that.

I find this tool to be disturbing. It is reality so I accept it. But I'm going to make effort to change my style between accounts.

A question for the author (costco): You created that account in 2019 but you didn't post or submit a single thing until 4 hours ago. Why did you create an account almost 3 years ago for no purpose?

Alone out here in the 0.30s. the three times I've used a throwaway account, they've been for a single post on a single topic, so no surprise they did not get picked up by this analysis I guess.

Does a low correlation with other users imply higher susceptibility to de-anonymization if I were using alts regularly?

There's someone (michaelmior if you're around!) with a false positive 0.46 match to me.

Maybe we could be friends :)

This needs to exclude who’s hiring post because it confuses me with a few of my wonderful former colleagues!

Well the only solution is too have too many alts so that nobody can believe you can possibly have that many

Wow. This is insane, it found my old accounts. So throwaway obviously (because I'm a bit of an asshole) but this really is amazing. It also highlighted another account that's not me, but looking through their comments i don't see any resemblance to me either.

I've complained a lot about Haskell and now it thinks I like Haskell =(

Needs sentiment analysis IMO, otherwise you'll get "Here's a bunch of people who are JUST LIKE YOU", except they use a similar grammar style but hold opposite opinions on the same nouns.

I have two accounts. This one, “soneca”, that is my first one and most active by far, and another one that I use sometimes mostly for Show HN and few comments.

When I searched the other one, “soneca” was the first guess, with 0.4.

But when I searched “soneca”, the other one was not in the top 20.

Those interested in the implications of this kind of analysis might enjoy the book The Secret Life of Pronouns http://secretlifeofpronouns.com/

Thank you for this.. I thought I was being careful but evidently it's not enough. It found 13 of my previous accounts with the topmost being 0.4937 and lowest one being 0.3616 bold. All the bold ones were right, some correct matches weren’t bold.

Seems pretty spot-on to me. I tried it with two accounts I was already certain were alts - based on other factors like favorite topics and common enemies as well as style/tone - and the top hits for both were the ones I would have expected.

Very interesting, .59 is my lowest, .64 is my highest match, none of these accounts are one of my alts. Though to be fair the handful of times I've used a throwaway I used it for a single comment so I didn't give it much to go off.

Anything like this for Reddit?

Would translating to other language and back defend against this algorithm?

Trailing (and probably leading, didn't check) spaces confuse the user lookup.

I wonder how much this can be improved if metadata is taken into account as well. Especially the distribution of common post dates and times modulo a week, which also exposes in which timezone somebody probably lives.

On one hand, thank you for showing us all how easy it is to make something like this. No doubt organizations with more resources already have more sophisticated systems in the same vein.

On the other hand, can we agree that this product is unethical?

In many cases, when a person uses an alt, it is a direct and strong signal that they do not wish their other posts to be associated.

So this product is circumventing the explicit will of the person, and making it available to anyone with zero effort i.e. there is no barrier to getting this info.

I met someone about 10 years ago who said they built this at a university. And their argument also was "actually this enhances privacy because it lets you know something something something". And yet their research grants were coming from one source only.

It can be used for good, but most often it won't.

Fun exercise would be to find all accounts that suddenly stopped posting around today and correllate them with new accounts created around today.

All those scared folks who naively think that it's not too late yet. Busted.

502 Bad Gateway

The asymmetry is interesting. I have no alts but of course it nonetheless reported accounts similar to mine.

Running then the most similar person to my account did not put me in their top 20.

I’m guessing that a small corpus for a given account doesn’t produce a very good score? I’ve done throwaways a couple times in the past and this has not “outed” them.

I've only had one account here. The highest match has a 0.624 score and the lowest a 0.572. I'm not sure if that means I'm unique or common but I'd like to know.

One way to get around this legitimately would be by posting a lot of quotes/lyrics/excerpts and the like thus fooling the algorithm unless it had a way to filter them out

This has been a great way to find people whose commentary I enjoy!

We knew this was possible and was coming, and probably around a few years. Fascinating from a technology perspective, terrifying from a long-term privacy perspective.

It's moments like this I'm proud to have my insanity on full display without obscurity. Was surprised to see a bunch of ~30% matches despite not having any alts.

My runner-up has a rating of 0.42378790667730715

C'mon guys, work harder. That's not even close! :-D

Btw, I myself am only at 0.9999999999999999 so I guess I need to work harder at being myself.

I tried it on a few user-ids that I strongly suspected were owned by the same person. My hunches stand corroborated. Not sure who is corroborating whom though, me or the script.

Good job.

Oddly, I am not an exact match to myself.

> Most likely candidates:

skymarshal: 0.9999999999999997

The other few usernames I tested (pg, dang, some random ones from this thread) all matched themselves at 1.0.

I had hard time to understand some comments made by my closest match. I guess this is good reality check. I need to learn how to write more legible posts now.

It didn't find my alt, but the second match is one of my twitter mutuals - I wonder if we've inadvertently borrowed style quirks from each other.

I wasn't aware this was even a thing! Scary stuff. 2 alts are listed but not with any great accuracy, so easy to dismiss. What an interesting topic.

Does anyone here have a reasonably wide variety of similarity ratings? I'd love to see the difference between a 0.2 and a 0.8 for the same account.

Interesting; I must have a fairly unique style as there are no matches over 0.40 for me.

I’m a native English speaker as well, so I’m unsure how to feel about that.

> I made this site mostly to show how easy this is and how it can erode online privacy

looks like it can indeed

> Here are some frequent HN commenters: (EDIT: Removed due to privacy concerns)

How surprising that someone might object to being included in a demonstration of the erosion of privacy!

Is the site opt-in or opt-out?

Is it weird that my rating is very low compared to alternative options? I have no alts, but I'm curious how similar others might write to me.

What is the threshold to be reasonably confident that two accounts are from the same individual?

I ever had only one account here and the closest match is at 0.47.

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 held by power-users who flag people -- most of the power of an admin at the end of the day but without any of the accountability. as long as they are following the mainstream dogma, its all good.

anyway, this app was able to identify a lot of my accounts. but a lot of the matches werent me. bold matches were almost all me. but i know there are many more matches than those that were listed. it mainly showed my most recent accounts.

i think most people would get a sick feeling in their stomach if they tried this app. i dont think people are prepared for a world where you can type someones name into an app like this and produce everything ever recorded online that was created by that person. not only this but everything highlighted and summarized to answer any question about that person. this is what advanced ai will bring us. an information implosion where the planet-sized ocean of data that is just floating all around us suddenly and violently coalesces into the objects of our new societal calculus. violent is a good word. and this is just the change that one can see coming with ai.

pretty cool- i think there should be a term for two accounts that have each other as the top most similar account. kinda sad i dont have one :(

Make a fundraiser and start doing it for other sites.

Clicking on my top match (0.61) - I can see the similarity. I also note they quote the same way, with a > symbol. I wonder if that helps!

Inserting random Unicode blank, 1/4, 1/2, or zero space characters into your writing may help thwart it too, if you are paranoid

Very cool! And really a shame that you’re not allowed to delete an old alt account or comments on HN! It follows you forever apparently.

All false positives for me - I want to reach out to the accounts that talk similar to me and see if we make good friends

Maybe this is a good tool to find new friends. :P

How do you protect yourself from impersonators?

So what are some good tools to obfuscate style?

It found my “alternate” account. If someone puts my username in, it’s not hard to figure out which alternate is mine.

No alt, and the highest match is 0.36

And that accounts last several comments were flagged as dead.

I'm a native speaker, but my english succcccks.

Funny thing would be to find most unique user account stylistically.

Which user has lowest best match?

Mine is 0.58 so I'm really not that unique.

would probably work better with case and punctuation preserving n-grams, sentence length, paragraph length and use of whitespace stats.

also maybe a tf-idf vector of top n words per user.

also could maybe do a same phrase analysis across the corpus to find some hand picked features.

timestamps could be interesting.

or, of course, let the machine do it with comment2vec.

I was curious to use this on myself to see if anyone writes like me. Closest was a .51 confidence, so I guess not?

This is cool!

If an account returns a high score for many accounts, does that also mean they’re relatively less original in style?

It puts almost all of my old accounts decently near the top, but my original account is almost comically low.

Cool! I wonder if it could be run backwards, to identify the users on hackernews with the most unique voices.

This is creepy.

My alt accounts (not really, all below 0.5) seem to also be European or German Firefox users. Good for us ;)

Obviously the next thing to do is make this a popup on someone's account name when you hover over it.

This is super impressive!

Is there a common open source library (Python, JS, whatever) that implements something like this?

> imagine what a company with millions of dollars and a couple dozen PhD linguists could do.

Could they do much better?

How much writing do you need to analyze results? Would changing account every X sentences eliminate this?

I have no alts, but to those of you compared to me by this engine : "Hey, good lookin'!"

wow, this is way off on me, didn't find my alts and the bolded accounts on my list are from different countries, use language I'd never use (cusses) and I see I've downvoted some of them...

I'd love to have the experience and or apparent wealth my "alts" have

This is great.

One funny thing though, while your example says 1.0, for my own account it says 0.99lotsof9s4

I like the way some usernames are only 0.9999999 correlated with themselves.

Perhaps 6 or 7 digits is enough?

This found an old account that I forgot I even had but with a lot of false positives. Neat!

I have no alternate accounts, and all my matches are below 0.4 for whatever it’s worth.

Interesting, but it gave me 20 accounts, and I know that I only have this one.

Sounds like a nice tool to find friends. You locate people who might think like you.

Strip leading/trailing white space from the name if it says no match.

I would have expected to be a closer match to myself.

> uberduper: 0.9999999999999991

Well, one of the closest on my list is my twin, so there's that.

Love a little NLP project on a public dataset - thanks for sharing!

Would this work for Fernando Pessoa and all his heteronyms? :)

I’d like to request the author takes this offline please until the implications can be thought through.

This is breaking anonymity that people incorrectly thought would not be revealed.

For some it might be awkward, others it might be quite problematic.

Found my phone account; I'm quite impressed, really !

Haha, you got me and my main account. That's spooky.

Im tempted to use it to find likeminded friends :)

This could be a good idea for identifying bots.

at what threshold is it considering alt account?

Jokes on you, this is my one and only account.

Are short sentences better for anonymity?

Well, interesting. This is one of the reasons we have the GDPR. @costco, if I were to make a GDPR erasure request, would you service it?

And I'm no lawyer, but it seems like there's also an outside chance of a breach of section 171 here as well, which is a criminal offence committed by a person who reidentifies de-identified data.

Plus - the laws have extraterritoriality. Vanishingly unlikely that you'd actually be pursued for it, but it's worth bearing in mind when you munge people's personal data.

Wow... that's shockingly effective

Welp, so much commenting for me then.

What's a high correlation number?

Are you going to try it on Twitter?

Now I can find my HN doppelganger

heh, I looked up the top bold hit for my name and they really do sound a bit like me (:

writing from throwaway:

Holy shit, it works really, really good. It found all of my older accounts.

What algorithm is being used?

I changed my nickname so my employer can't find me here. I'm not amused by this.

The website is down...

Now do one for reddit

why is my username not exactly equal to 1? https://stylometry.net/user?username=julienreszka

does it use the most used words or least used?

Possibility to hide user comments in profile should be optional.

didn't find a single one of my alts. nice

Of the top ten accounts listed for my name two of them are me.

nice one. are you using gpt3 under the hood?

we leave fingerprints everywhere

ColinWright is Dang?

Woah

totally on spot

my current and my old account

w

Wow... how !

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.

What's up with cluster of users like:

j_s,password4321,carolinew,colinwright,kuharich etc.

https://stylometry.net/user?username=j_s https://stylometry.net/user?username=carolinew https://stylometry.net/user?username=colinwright https://stylometry.net/user?username=password4321

Lowest match for j_s is 0.80 and all but one is black.

Most likely candidates:

    pg: 1.0
    montrose: 0.604073065373204
    mattmaroon: 0.5900372458160795
    natsu: 0.5519832271289953
    rauljara: 0.5418566694533273
    waterlesscloud: 0.5378996309342633
    damoncali: 0.5292014150349463
    gruseom: 0.5290151637991445
    kemiller2002: 0.5254174524920762
    jfengel: 0.5231938496089998
    jamesaguilar: 0.5229081613163672
    houseabsolute: 0.5219738531025365
    danssig: 0.5195368367601849
    austenallred: 0.519343009683366
    loewenskind: 0.5177030083877397
    baguasquirrel: 0.5153841099708854
    asdfasgasdgasdg: 0.5146704002447524
    aptwebapps: 0.5144149629369845
    allenbrunson: 0.512802806408646
    danielweber: 0.5123620795710832

Not today.

You fail, I win.