There's been some interesting threads about stylometry over the years [1]. The top link was quite decent at unmasking HN alt accounts with basic ngram analysis whipped up in one day [2].
I remember this! it got me pretty good. I have a bad habit of generating alts because I forget the password.
It makes me wonder if we could use non-instruct LLMs to slightly alter the wording of text while keeping the meaning the same meaning. Perhaps by using perplexity or some other metric. I don't know, maybe you compare the distance of the "meaning" vectors.
You might also want to have some "style" vector associated with each pseudonym. For example, I might want it to produce british english under a certain pseudonym, and simulate an ESL speaker under another.
Essentially, you would want some way of re-styling text. The basic way to do this would be to run the same sylometry tools the hunter uses and manually make synonym word/phrase subsitutions to lower your similarity.
It's a cat and mouse game, but I think the mouse eventually wins. Consider a program that translates your english writing programmaticially into a low-entropy symbolic form and then translates them back to english in a procedural manner. Basicially you design an intermediary language that cannot contain style. It would be boring to read but it would remove all the style.
Yep, I was thinking the exact same thing. Write something, get your llm to rewrite it to "standard Steve Yegge form" to hide your identity. If we all do the same we are probably safe.
If that's as effective as comments suggest, then it seems indeed unmasking pseudonyms has only been a matter of effort for a while.
I still think people underestimate the power of even minor inconvenience. While you can't just click a button to reveal all pseudonyms of someone (e.g. you need to download some obscure tool or even perform statistical analysis yourself), I think this provides significant (and surprising?) protection for pseudonymous individuals, I'd say even (although to a reduced extent) for sophisticated threats like state-level actors. I hope LLMs continue to be unable to do so for the near future (I just tested and LLMs can't do it with a simple prompt).
Which is why I think privacy safeguards still work quite well even while being technically mostly bypassable.
So probably tools like this should be kept private if possible.
Looks like the site was taken down. I was curious who wrote the most like me, given that I don't have any alt accounts, but I guess weighing privacy over my curiosity is a good thing in the big picture
It makes me wonder if we could use non-instruct LLMs to slightly alter the wording of text while keeping the meaning the same meaning. Perhaps by using perplexity or some other metric. I don't know, maybe you compare the distance of the "meaning" vectors.
You might also want to have some "style" vector associated with each pseudonym. For example, I might want it to produce british english under a certain pseudonym, and simulate an ESL speaker under another.
Essentially, you would want some way of re-styling text. The basic way to do this would be to run the same sylometry tools the hunter uses and manually make synonym word/phrase subsitutions to lower your similarity.
It's a cat and mouse game, but I think the mouse eventually wins. Consider a program that translates your english writing programmaticially into a low-entropy symbolic form and then translates them back to english in a procedural manner. Basicially you design an intermediary language that cannot contain style. It would be boring to read but it would remove all the style.