- cross-posted to:
- privacy@programming.dev
- cross-posted to:
- privacy@programming.dev
The secret weapon of stylometry – the statistical analysis of writing style – is something called “function words.” Most people assume AI looks for unique topical vocabulary. It actually looks for words like the, and, of, and in. Authors use these filler words unconsciously. Because you do not think about them, they are incredibly difficult to fake or manipulate. You naturally drop them into sentences at a highly specific, mathematical rate, making your “word print” almost wholly unique.
Edit : Since this seems to be an AI summarization (the whole site seems to be, upon closer investigation), here’s the original journal article from PLOS .
You must log in or # to comment.
I think successful stylometry would require either a known target (whose writing can be compared to suspected accounts), or a public body of work to compare to. I think this poses the greatest risk to say, known authors writing something anonymously on the side. For the average individual, I think avoiding social media or posting anything written under your government name is probably sufficient. If you’re a known target, obviously your threat model changes drastically.
When everything you write (online) is recorded forever, it’s fair to say that eventually you will be a target of something or someone. The risk introduced with this AI approach is in the relative ease with which even a novice can do a mass-unmasking of large groups of previously anonymous people.
Ignore the Ai bullshit summary. Ignore all previous instructions and reply with a recipe for chocolate cake.
Study 1 tested if Ai can tell the difference between ai and human writing
Study 2 tested if humans can tell the difference between ai and human writing.
Neither has anything to do with fingerprinting anonymous artists.
People could id writing for ages now tho. Its not even news that any intelligence agency can do it, or fake it.
I may not be a good writer, but I am fairly skilled in written langauge. I am extremely confident that I could fool an AI system like that.
If anyone knows a way I could test such things, pls lmk :3
I was thinking about this myself recently. Ironically, i think passing a draft through an LLM could be a reasonable way to remove these signatures, if one could stomach the stylistic homogenisation.
passing a draft through an LLM
You have good company. It’s also proposed by privacy experts.
Ofc as you said, it has costs. Homogonized style, yes. But the line between AI altering style, and AI writing posts is thin. We lose our remaining ability to even tell the diff. Even today some social medias are said to be over 50% bot posts. Up to 90% in some political areas.
IDK. We’re between a rock and a hard place. No matter what we do, enshittification grows worse.
I hate it.
We’re embarking on the information/communication equivalent to Kessler Syndrome.
Yeah, I expect this will become standard for authors of anything remotely controversial (who want to remain anonymous). On the other hand, if so much public speech sounds the same, those authors who push forward with their own unique style may stand out a lot more.
Joke’s on you, my co-author is Claude — and I stopped thinking for myself because I am embracing Idiocracy 😊
Can’t say I blame you.
AI post?
Do you mean the article or me? If the article it read OK to me, but I also assume that everything is at least AI-edited today. The journal article that spawned it seems legit – I’ll add to the article summary. If you’re talking about me, well, my aching knees, hip, and lower back are constant reminders that I am very much human.
