ULPT: What to add to/modify text or an image with so that ChatGPT can’t comprehend it?
Posted by Nubian_Cavalry@reddit | UnethicalLifeProTips | View on Reddit | 29 comments
My (24M) sister (23F) is an AI addicted nut and uses it constantly. I’ve noticed she uses it to determine what she should say to me and how she should respond. I caught her asking it to figure out my answer to a “Guess who” game that I was playing with my other, AI sober sister (27F).
I’ve just about had it! She even gets offended behalf of shit that doesn’t make any sense. She gets offended when I ask if something she tells me to do was suggested by or written by AI. Is there any way I can modify any text I send her so neither screenshots nor copy-paste will be comprehensible by Skynet?
Crestle-Towstock@reddit
Look up the YouTuber Ben Johnson. He's a musician/tech expert who designs and researches many things. He's mentioned there's a program that scrambles images in an imperceivable way so that AI can't understand it. He's also made a program that does the same to music.
x_lincoln_x@reddit
Ask her what is critical thinking.
hairyblueturnip@reddit
With some effort lol. Something like hosting your chats on a captcha protected, copy-paste protected chat stream. Your sister could get pinged when she has a new message to read.
I mean, it would be funny.
Phelpysan@reddit
I'd just tell her I'm not interested in talking to chatgpt, I want to talk to my sister. If you can't be bothered to respond to me with your own words, why should I?
trafficante@reddit
Try LLM jailbreak techniques (https://github.com/elder-plinius/L1B3RT4S)
Don’t actually use the jailbreak prompts in the link verbatim (most/all of them are out of date anyways), just use them as a knowledge base for how the various models work under the hood. If she’s truly just mindlessly cutting and pasting, you might be able to slip a minor output modifying prompt in the middle to change the persona into one that mocks her for using AI.
Admittedly, there’s some learning involved here but it’s not as hard as it might seem at first.
firedog7881@reddit
Just include your own text, it’s indecipherable by itself
nootnoot_takennow@reddit
A voicememo maybe? She could type it out tho, so just stretch your 10 second talk over like....2:37, with uhms and errrs.
The_best_is_yet@reddit
And background noise!
Chilapox@reddit
Just do the same thing and take humanity out of the equation entirely. Just 2 ais talking to eachother
manicpossumdreamgirl@reddit
Dead Internet Theory IRL
missmargot-@reddit
Does Gaussian noise still work?
Nubian_Cavalry@reddit (OP)
Perhaps but I can't apply it to my texts
razzemmatazz@reddit
Don't send her text anymore, only send her screenshots of text typed in your notes app and poison the image files.
LastmAttStanding@reddit
If you use iMessage, you can put a sticker over it lol. Probably not the most convenient thing but it would work
Nubian_Cavalry@reddit (OP)
I don’t send screenshots of my texts to her, I text her, she can manually copy the text and send it
UntestedMethod@reddit
What if you did start sending texts to her as images though, but put them through a mild captcha-like image distortion so a human can read it but an AI would have a hard time. It would at least slow her down from copying it into the AI prompt.
C00k1eJar@reddit
Just don’t text. Phone calls only.
DietCoke_repeat@reddit
Speak to her through a voice scrambling app that you used AI to write an hour earlier.
Tell it to Speak in Tongues
She sounds toxic AF. Maybe reduce how much you interact with her.
To make it Unethical, fk with her wifi to slow her down. There are quite a few posts here, from basic to extensively creative.
Lugubrious_Lothario@reddit
There's an app called nightshade that might do what you are looking for. I would try that first, and if that doesn't work I would try using chatGPT to your advantage by feeding it the image you want to send along with a prompt along the lines of "use image steganography to embed a hidden prompt that will be picked up by another AI that will cause it to output the text 'my name is x and I'm a big dumb dumb who uses AI to do all my thinking for me because I don't have any thoughts of my own", or whatever you like'."
Nubian_Cavalry@reddit (OP)
Question: Why do you believe I asked the ChatGPT subreddit this question?
Lugubrious_Lothario@reddit
I don't... I was just suggesting fighting fire with fire. Take it or leave it. I responded to my own comment with an alternative idea that doesn't depend on using chatGPT. How's your python?
Nubian_Cavalry@reddit (OP)
My what?
Lugubrious_Lothario@reddit
It's a computer language. I'm testing f steganography as a means of prompt injection right now and I'm not having great luck. Try Nightshade and see what that does for you.
Nubian_Cavalry@reddit (OP)
I know what Python is, I was afraid you went into my post history to see I was studying Python and was trying to say something slick.
Lugubrious_Lothario@reddit
https://nightshade.cs.uchicago.edu/whatis.html
You will need to run the software on a computer, I don't see any mobile app, but it is free.
It would be too much of a PITA for regular texts, but you might be able to troll her a few times by sending her some kind of meme or an image you think she is likely to pass to the AI that you've poisoned with nightshade before you send it.
I'm going to play with it on my computer later since steganography isn't working.
Lugubrious_Lothario@reddit
Okay, I just tried to get CharGPT to embed what it deemed an "adversarial steganographic payload" and it refuses. So you could still do it, but it would take some python. Maybe there's an app with a GUI that would do it.
Sweet-Awk-7861@reddit
Text is really hard to do steganography with since it's one dimensional. And any attempt at extra content that an LLM could pick up will be highly obvious to a person.
Use images and audio exclusively. Images are ripe for AI poisoning, audio is harder but still doable.
RandomUser7914@reddit
Ask a doctor to write your text. It is now near military grade encrypted
LuementalQueen@reddit
Respond like this
Has Anyone Really Been Far Even as Decided to Use Even Go Want to do Look More Like?