A nightmare scenario previously only imagined by AI researchers, where AI image generators accidentally spit out non-consensual pornography of real people, is now reality.
I scrolled a bit more, and saw another AI-generated nude of someone who looked like Margot Robbie.
Ever since I was a little girl, I’ve always wanted to become an anatomically deformed eldritch horror haunting the collective nightmare of humanity. And also, look pretty.
The AI sexbot revolution can be a net positive. With all the weirdos being indoors and hooked on their digital soma, only people who want genuine connection with other human beings will be on the dating market.
“I get good chat going, the AI is set up properly, very good start, like 10 messages in or so but then suddenly the AI decides I should cum and end it all,” another user said. “The thing is that the sex part haven't even started yet.”
Once there’s a way to simulate touch and this tech actually works, I bet some people will never leave their house and have a VR headset on for days on end.
There is a way to simulate touch in VR chat and other VR platforms through haptic feedback. There are ways of turning picture of humans into 3d models (See the Econ github). And, of course, there are ways of making AI pictures for free. All you need is someone to link these together.
I won't lie, I checked it out from curiosity. Scrolled for maybe five seconds on their homepage and found a woman in a yoga pose with her head full on backwards on her body, completely 180° around.
Honestly the market for this kind of thing is pretty sad.
Seriously, speaking as a man, I have no idea why on Earth some of us men are willing to pay for things like the stuff described in the article?!
I remember thinking the same thing a few years ago about Belle Delphine and her “gamer girl bath water”. I get it, I get horny too and have no idea how to approach women, but seriously, there is so much porn of all kinds out there entirely for free.
This is why all my AI bots are offline and on my own web server. Not only can they never randomly morph into hideous abominations, no amount of fearmongering moralism can take it away.
Yet another reason that we cannot allow ML companies to set a precedent that “it’s fine to use non-consensual training data, because the model only ‘learns’ from it and never reproduces an exact replica of any single input”.
Also, this was not surprising:
Dillon said that DreamGF has a team of between 20-25 developers, mostly in Bulgaria, and that they previously worked at an NFT company.
My complaint actually hinges on models not emitting an exact replica. That would be obvious infringement. In cases like DreamGF, they would be wide-open to lawsuits from very wealthy people whose primary asset is their right of publicity.
What these ML companies are doing is: They are identifying where the line of definite infringement lies, and aiming their business as close to that line as possible.
Well, setting aside that there’s no law against being a total asshole, so like… We don’t have to make a bad behavior illegal in order to complain about it…
There’s the letter of the law, and there’s the intent. We start with a shared cultural attitude of how we should treat each other, and then we turn that into a quantifiable, objective rule that we can enforce through law.
We can try to make the law match our cultural attitudes as closely as possible, but there will always be gaps.
Now, I’ve got my own beef with how our IP and publicity laws work, and I’d like them to be more permissive in many ways. Much of IP law is exploitative, takes advantage of creators more than it protects them, and has lagged way behind where our social norms are these days.
But these ML companies aren’t interested in abiding by any social norms at all. Only paying lip service to current laws, which were written in a time before these “AI” services were even a possibility – skating by on technicalities, like a little brother poking the air 2 inches from your face and taunting “I’m not touching you! I’m not touching you!”
Were there nude images of Jennifer Lopez in the training set? Did it output copies of those nude images of Jennifer Lopez from the training set?
The article says that images were generated of "someone who looked like Margot Robbie, and another image of Lopez." Looking like someone is very different from producing an "exact replica" of a training input.
I think the idea is that if the model doesn't know what Jennifer Lopez is, it couldn't make imitations of her naked.
Realistically that ship has sailed and AI is capable enough now that even if the data wasn't there it could be pretty easily added.
It will need to become a simple fact of life. If we can imagine something now, we can have pictures of it. There is no putting this back in the bottle.
"Knowing what Jennifer Lopez looks like" is a very distinct thing from "reproducing an exact replica" of training data. OP appears to be arguing that the former is not true because he thinks the latter is true, but it's actually the opposite. That's the crux of what I'm arguing here, OP is simply factually wrong about his position.
Edit: OP has pointed out that he doesn't actually think there are exact replicas being produced, which just makes this even more confusing.
OP has pointed out that he doesn't actually think there are exact replicas being produced, which just makes this even more confusing.
Your misread their first comment, I think.
They were saying that DESPITE the common arguments that AI only learns and doesn't copy exactly it might still be good to require consent for people's content to be in training data.
“If the models are trained on images of specific individuals, the models can reproduce images that resemble those people. In the worst case, the model may even directly output verbatim copies of images from the training set,”
Firstly, that paper was written in January and examined a Stable Diffusion model that was already obsolete due to its poor training even back then. Secondly, even with that poor model they had to move heaven and earth to find a handful of examples out of hundreds of millions of training examples where they could get a blurry replica out.
Secondly, as mentioned, that model is long obsolete due to issues exactly like this. Modern models work better in part because they have better curated training sets that eliminate this sort of "overfitting." There's no indication in this article that the website in question is using one of those old models, it's just presented as a hypothetical concern.
The cat is out of the bag. Everything what those services do is like child’s play compared to the stuff you can do at home on your own PC. This ain’t ever going to stop, it’s just going to get more powerful.
I recommend playing around with Bing Image Creator/DALLE-3 for a bit to get an idea of what is possible these days. It will filter celebrities and NSFW content of course, but the amount of flexibility it has and quality it produces is mind boggling.
Add comment