one of my regrets is when i wrote that article on stylometric fingerprinting. i should have waited to get better at stylometric mimicry and focused on how i did that to improve anonymity and make it harder to connect alts (a bunch of custom vale.sh rules to remove style and conform to a boring technical style, then reintroducing a target’s style). instead i kind of spawned this community of people interested in using it for forensics.
i know they’ll see this post but idc at this point lmao

Seirdy
in reply to Seirdy • • •on the other hand, LLM slop awareness has created more bad stylometric analysts than any blog post of mine ever could.
“this was clearly written by an LLM because it uses [very common vocabulary preferences]!”
There are clear tells, and LLMs like ChatGPT do have a set of (cringeworthy) emergent loose style guides, but I usually don’t rely on this sort of thing. I don’t have a linguistics background. I can claim to be good at fingerprinting avoidance but actual fingerprinting is too error prone.