Quite the spicy one aren’t you. I see where you get your username from.
But yes, they exist, and so does the ability to defeat them by training the detection data into a new undetected model. Cat and mouse game, as they say.
Comment on Amazon's Hidden Chatbot Recommends Nazi Books and Lies About Amazon Working Conditions
mods_are_assholes@lemmy.world 8 months agoIt literally exists today and you sit here typing bullshit like ‘is that even possible?’
Get every manner of blocked.
Quite the spicy one aren’t you. I see where you get your username from.
But yes, they exist, and so does the ability to defeat them by training the detection data into a new undetected model. Cat and mouse game, as they say.
So, what you’re saying is you don’t really understand how data security works then.
Because it’s never a ‘one and done’, it’s ALWAYS a cat and mouse game, ALWAYS.
Which is why antivirus companies push definition updates.
And now we do the same with AI detectors, or they become irrelevant.
This is where I get to lol and say you don’t understand AI.
When a kernel privesc vuln is found, it gets fixed. Unless it was improperly fixed, that particular 0day can’t be exploited again.
But when it comes to AI, a GAN’s job is to take the ‘vulnerability’ that was ‘fixed’ and train on it to exploit it again.
And again.
And again.
And again.
It’s funny how people can just link to a wikipedia article about a ten year old thought experiment and think its some kind of mic drop moment. The current AI paradigm is starting to hit its singularity curve and hardly anything that old is anything more than a novelty and largely not applicable to current models, ESPECIALLY when it comes to modeling,
We aren’t seeing this kind of iteritave adversity being used in actual real world attacks, and it is largely useless to train on a patched vulnerability.
But I’m sure you already knew that, maybe your testing me?
T156@lemmy.world 8 months ago
They’re also infamously terrible, being half-correct and prone to detecting non-native English speakers as being AI. To the point where at least one institutition decided to not use a detector.
You’d likely be equally as accurate guessing at random. Not to mention that at the end of the day, they’re just recognising quirks in the generated text. It is not difficult to mask those quirks either by having the prompt put out text in a different style, or for an update to change the generated text, breaking the detection as well.
There is no definite, sure-fire way to determine that text is AI-generated or not. For all you might know, as an AI language model, I could have cooked this comment up using a billion probability nodes loaded up into a typewriter, as it does not go against OpenAI’s policies on generating text.