I bet this is because LLMs stole her art for training, so now that the LLM users have pumped out thousands of slop-art aping hers down to the colour palate and brushstrokes, her originals also get identified as also AI.
This is exactly it. Also the reason that inputting an excerpt of a classical text (cfr. Dracula, Frankenstein, …) returns “likely AI” results - it’s literally what it was trained on.
While I don’t think you’re entirely wrong, I feel like this sentiment will lead people to believing AI content checkers more. The content checkers are wrong even for newer stuff. I always default to believing something is human made because I never want to accuse a real artist for using AI.
I don’t follow, if content checkers are wrong for not being able to tell original art from AI slop because they have been trained on datasets that include the original art, why wouldn’t they also be wrong for the same reason on newer art? They would still likely absorb that newer art into their datasets.
Okay after writing it out I get what you mean, but I am thinking that the turnaround time of absorbing new art, using it to produce AI slop requests a few dozen/hundred/million times, and then calling the original works AI because the AI sloptificator can no longer tell the difference will be a constant, unless there is some way to track OC across multiple AI agents in the panslopticon.
AI content checkers will always be wrong. They’ll always sometimes flag non-AI images as AI and AI images as human made some percentage of the time. I don’t want artists to have to prove that their art is real, I want them to be able to make and share the content they want.
No, words have meanings, I hate AI as much as the average lemming but “LLMs stole art” is just a horrible way to describe what actually happened. It makes it sound like LLMs are going rogue and collecting art themselves. Humans decided to train an image generation model specifically on art like the one in the OP, and just glued an LLM on top of the final model. LLMs are not the ones that do this. I wasn’t even trying to defend anything or anyone here, but I guess this place is too toxic to even use correct phrasing.
I guess this place is too toxic to even use correct phrasing.
I feel you on a daily basis here, however I do think you could consider what the tone of your first response was like. There are a lot of AI supporters out there and I genuinely think you sounded like one there.
But you could have said something along the lines of “LLM companies steal data, don’t accidentally give credence to the marketing idea that these things are sentient”
The irony of him complaining about inaccuracy and lack of clarity in someones post, but being so inaccurate and unclear with his complaint is just… amazing.
I bet this is because LLMs stole her art for training, so now that the LLM users have pumped out thousands of slop-art aping hers down to the colour palate and brushstrokes, her originals also get identified as also AI.
This is exactly it. Also the reason that inputting an excerpt of a classical text (cfr. Dracula, Frankenstein, …) returns “likely AI” results - it’s literally what it was trained on.
For anyone doubting Kyrgizion’s claim:
While I don’t think you’re entirely wrong, I feel like this sentiment will lead people to believing AI content checkers more. The content checkers are wrong even for newer stuff. I always default to believing something is human made because I never want to accuse a real artist for using AI.
I don’t follow, if content checkers are wrong for not being able to tell original art from AI slop because they have been trained on datasets that include the original art, why wouldn’t they also be wrong for the same reason on newer art? They would still likely absorb that newer art into their datasets.
Okay after writing it out I get what you mean, but I am thinking that the turnaround time of absorbing new art, using it to produce AI slop requests a few dozen/hundred/million times, and then calling the original works AI because the AI sloptificator can no longer tell the difference will be a constant, unless there is some way to track OC across multiple AI agents in the panslopticon.
AI content checkers will always be wrong. They’ll always sometimes flag non-AI images as AI and AI images as human made some percentage of the time. I don’t want artists to have to prove that their art is real, I want them to be able to make and share the content they want.
LLMs don’t steal art for training, image generation models are fed datasets by humans, which are then invoked by LLMs when prompted
Arguing semantics is a classic strategy of people who don’t have a valid argument to make.
Literally the “guns don’t kill people” argument.
No, words have meanings, I hate AI as much as the average lemming but “LLMs stole art” is just a horrible way to describe what actually happened. It makes it sound like LLMs are going rogue and collecting art themselves. Humans decided to train an image generation model specifically on art like the one in the OP, and just glued an LLM on top of the final model. LLMs are not the ones that do this. I wasn’t even trying to defend anything or anyone here, but I guess this place is too toxic to even use correct phrasing.
I feel you on a daily basis here, however I do think you could consider what the tone of your first response was like. There are a lot of AI supporters out there and I genuinely think you sounded like one there.
But you could have said something along the lines of “LLM companies steal data, don’t accidentally give credence to the marketing idea that these things are sentient”
The irony of him complaining about inaccuracy and lack of clarity in someones post, but being so inaccurate and unclear with his complaint is just… amazing.
Exactly what does any of that change?
It’s tracing.
Why do you think people are upset about ai “stealing” art styles then?