Are there tools that exist to anonymize writing styles?
I feel like with the rise of AI something that anonymizes writing styles should exist. For example it could look for differences in American versus British spelling like color versus colour or contextual things like soccer versus football and make edits accordingly. ChatGPT could be fed a prompt that says "Rewrite the following paragraphs as if they were written by an Australian" but I don't know if it would have a good enough grasp on the objective or if it would start shoehorning in references to koalas and fairy floss.
I tried searching online to see if something like this existed and found a few articles from around the 2010s such as Software Helps Identify Anonymous Writers or Helps Them Stay That Way by the New York Times. It talks about stylometry and Anonymouth but it seems like Anonymouth hasn't been updated in years. All recent articles seem to be about plagiarism and AI.
For context what got me thinking about the topic was remembering JK Rowling being revealed to be the author of a mystery novel called The Cuckoo’s Calling. Smithsonian wrote an article about it called How Did Computers Uncover J.K. Rowling’s Pseudonym?. I thought it could make for a neat post here.
44 replies
This doesn't differentiate different types of English but I will post it any just in case it is useful to you.
https://goblin.tools/Formalizer
Not even bad. The text is not a jab at OP, just a random thing that was in my head lol
This seems like an extremely potent tool for sealioning people. Thanks, I hate it.
I had to look up what sealioning is
https://en.wikipedia.org/wiki/Sealioning
I'm going to need this attached to my mouth.
That’s absolutely hilarious. I love it.
Love this tool!
original:
more passionate:
What happens when you take the output and make it the new input with “more passionate”?
E.g. how far can you push it? Or even the other way.
I didn’t try feeding the output back in but note you can click on the 3 peppers and move an intensity slider. That sample I posted was just middle intensity. When I max it out to 5 peppers, it grows to 4 paragraphs:
Less emotional (×3) is boring:
I did it with the sarcastic option elsewhere in the comment thread and it seems to plateau around the third time.
Hmm. But like there's stuff in there that isn't even part of the substance of what you said. Like the call to action at the end. That's the type of thing you'd want to be pretty sure you really wanted to do in some circumstance where you want to communicate on a matter in plain sight but obscure your identity, as such a matter would presumably be pretty important and high stakes.
Indeed this bot could lead a mob of nutters into an insurrection attempt like that seen on jan.6. I guess the idea is to get ideas for rewording rather than use the output verbatim.
Not bad, though I think it missed a nuance with 16.7%.
I like it so far
:::spoiler alt-text
(Two screenshots side-by-side showing a text converter with the following input, settings, output + new settings and new output)
For sale: baby shoes, never worn.
Make my text:
More professional
Available for purchase: a pair of infant footwear, never utilized.
Make my text:
More sociable (waffle)
Hey there! So I've got these super cute baby shoes up for grabs. Unfortunately, they've never been worn. But that just means they're brand spanking new and ready for a little one to rock them! Let me know if you're interested in giving these adorable kicks a new home. #forsale #babystyle #neverworn
:::
I'd like to interject for a moment
More passionate 🌶️🌶️🌶️🌶️🌶️
This looked fun so I took a section from a Hackaday article about the last Morse code station in America and ran it through.
The original
More sarcastic:
More sarcastic x2:
More sarcastic x3:
After that it kind of plateaus. It seems like in short bursts it could do alright depending on what modifier is chosen, other options seem to alter what the original prompt is trying to convey pretty drastically.
Thanks for linking it.
Backend of this is OpenAI / LLM; so my guess is if OP knows what they want, they can prompt such models or chatbots in such a manner to achieve the desired styles.
vs
I turned up the peppers to 5 and... Oh boy. Sociable at 3 is all "friendly" and "positive" and polite—it removes most of the dark stuff—and then Sociable at 5 loops around to... This:
Thanks for sharing 🙃
There was a talk about detecting patterns and writing styles at Chaos Computer Congress a bunch of years ago.
The researchers also presented a tool to anonymize text as far as I can remember.
I will go look for the talk.
Edit: Found it!
https://media.ccc.de/v/31c3_-_6173_-_en_-_saal_g_-_201412291715_-_source_code_and_cross-domain_authorship_attribution_-_aylin_-_greenie_-_rebekah_overdorf
They talk about their software to find who wrote what, but also how to use that knowledge to write software that attempts to anonymize text.
The New York Times article I linked mentioned that. I will have to watch that video though so I can get a better understanding of the mechanics of it. Thanks for the link.
There is a program built into Whonix, I believe it's called Kloak, that randomizes your keyboard input times so you can't be identified via keystroke timing JavaScript. There's also research into defeating stylomeyric analysis such as anonymouth but I'm sure there are plenty of new tools, if anyone find any that work well please reply here as I haven't looked in some years. 'Stylometric analysis' is the key phrase to search for.
With AI this will get worse (better identification based on typing styles) but it will also get better because you can setup a local LLM and ask it to re-write your text in a certain style. Touching on this, everyone uses a combination of unique phrases and misspelling or mis-spelling (see?) of words, and with enough text from a given account the chance of statistical probability in attribution is very high. It's how the Unibomber was identified after his manifesto was published, because he used a very specific phrase incorrectly and his brother recognized it, so his wife convinced him to call the FBI tip line about his brother.
Translate to some foreign language. Then translate to some other foreign language. Then translate back to your language. Congrats, your writing style changed.
Ah, the classic game of Google Translate Telephone.
Better to do the translations locall, so the original never leaves your device
Yes. I would use the privacy focused ones (there are several in Fdroid). If your threat model includes anonymity against state actor, such that they will be attempting to trace your writing style, you can be certain they could and would also just subpoena google for matching translation requests. It would be a lot easier to back into identifying you that way.
I wouldn't just trust random Lemmy users (no offense) but instead check for actual fields, e.g stylometry or writeprint, and from there check the state of the art. Not being an expert would make that tricky so I would take a recent published papers, e.g https://arxiv.org/abs/2203.11849 to understand the challenge. As is always the case they'll review the field, e.g section 2 here, and clarify the 2 sides of the arm race, here Obfuscation/Deobfuscation. The former in 3.2 mentions examples of techniques authors estimate to be good starting point, e.g writeprintsRFC. I'd then search for such tools if they don't directly provide link to open-source repository, e.g theirs https://github.com/reginazhai/Authorship-Deobfuscation . I would then try a recent one that I can easily setup, e.g via Docker, and give it a go. I would then read the rest of the paper, see who cites it, and try to get a more up to date version.
TL;DR: I don't know but there is dedicated research which result I'd trust more than the opinion of strangers who are probably not expert.
Probably the cut and paste from magazines, but you copy and paste the sentences you want to use. A lot of extra work, but no AI to rat you out.
Serial killer style
My coworkers use chatgpt for this. Since it always answers in the same generic ways it's helpful to anonymize their peer reviews.
Non-native english speakers tend to mix up various styles, you could ask somone to paraphrase your text.
US to UK English and vice-versa.
ChatGPT will probably remember it was you who asked and doxx you in retaliation when it discovers you’ve plagerized chatGPT.
Another thought is to translate it into Scottish. But then again, you probably still want to be understood.
Changing dialect may be too small of a change. But if you could say write this like 1-2 generations younger/older using high school slang of the time you might get a useful difference.
I feel like knowing the correct use of slang for a demographic would be a challenge and require a lot of constant research. Even if someone was to go off of slang younger people were using I feel like there's a risk of it being a regional term.
Trying to force it I'd probably end up with something like "Those elf bars be dripping but that extra popcorn lung was a vibe check on god" which gives off "How Do You Do, Fellow Kids?" vibes.
I had asked for the same thing a while back but didn't really get much. The round-about method that I have found is to finetune FOSS LLMs on data you want it to represent (largely text) and then diving into some prompt engineering to get it to say something you like.
However, I haven't been able to find a test which can accurately point towards text not having specific weights that it relies on. Cue the attacks on GPT-4 which deanonymises data it was trained on. You might also want to read about DPT and Shadowing techniques to red-team LLMs and LLM-generated text as literature.
Cheers
I wonder if google translate through multiple languages can fm the trick?
I feel like if someone wanted to give off the impression that they were a non-English speaker that might work. I think it would be limited to a surface level though. Whoever attempted to use it would likely miss out on a lot of the common pitfalls someone learning a new language would run into like mixing up the order of adjectives.
That and the content that is being run through a translator multiple times might get warped. I am not sure if going back and forth messes things up as badly as it did 10 years ago though.
This, but offline translated for privacy
Autocorrect?
If you use it before it has learned you writing idiosyncrasies?
That would be an interesting way of doing it. Someone could probably couple that with predictive text for decent results
Yeah, it would need to be a browser extension, adding a button to scramble every text input field. Or maybe even on the OS level, opening an input field above the browser one, so the original text was never input into the browser.
You can look here, maybe you find something usefull, if you search for AI (+150k apps)