How do you tell when something is AI art?

The AI art.detectors seem to mostly identify old D&D art as AI generated, weirdly enough.

Because it was. 70's D&D illustrators were time travellers in disguise. Same with authors. No real human would list a vouge, a guisarme and a bec-de-corbin in their weapon list. All was just written by ChatGPT-18 in 2054 using the prompt "Enhance my cow-boys and indians game by providing a few guidelines on how to play the game and disguise it as medievalish, ((((more details:1.9))))". Lack of enough negative prompts led to the infamous harlot description table, a telltale sign of the usual AI horniness.
 
Last edited:

log in or register to remove this ad

Why do people keep correcting my question? :D

Because there is no unique answer to your question, first because you didn't define AI art.

I can't identify for sure the raven image that was posted before as AI, but I had a hunch because it was trying to be photorealistic but had a very strong blur in the distance, which is a flaw of many models of a specific implementation of an image-generation AI. Nothing conclusive, just a hunch. But it was a method effective because it was a generation of a photograph-like image, made from scratch. It wouldn't have worked if there was a real photograph being improved/modified/touched up with AI. Would it be AI-art in your definition? The means to detect AI art depends (even if the case there are ones) depend on how the AI works, and therefore we need more definition to answer your question effectively.

Also, one does need to know the goal behind the question, because the answer will depend on it. If it's just to win "guess if it's real or AI generated" contests so prevalent these days, a wide test, with acceptable accuracy, will be a good answer. If it is something with higher stake, like sueing someone into oblivion, you'd need a 100% effective test that would modify the answer. Without this information, one need to rephrase your question to answer it. Or accept that the question can't be answered, which isn't leading to a discussion.

I proposed to ask, for each image, variants of the same character including some details that would be very difficult to have an AI engine trained on, like eating sushi upside down. It is not perfect of course, but it would be a hard test of consistency (a current weak point of many generative AIs) and it would be easier for a carbon-based artist to start back from a previous sketch. Inability to do that would make me tell something is AI art. Does it work for your use-case?
 
Last edited:


Ryujin

Legend
That and the strong bokeh. But I guess you didn't try to remove it.
I had nothing to do with that crow image apart from seeing it pop up on several crow based Facebook pages, without it being flagged as AI. The focus plane is pretty much in keeping with how a real close-up shot of the sort would look and I'm guessing whoever created the script for the image also purposefully introduced artificial motion blur to the crow on the right. There was a fair bit of effort put into it.
 


I don't know what that means. AI art is art created by AI. I'm at a loss as to what definition you're seeking. 🤷

If I do an image from scratch using my pencil, and I digitize it and I notice a few blotches of ink and tell the AI "remove the inkstain" (a possible use of the AI program called llama-cleaner), should the result be "AI art" and fail the test, or should the result be "human-made" and pass the test?

Do you intend to detect 100% made AI art only without human intervention? Do you intend to detect accurately only "low effort AI art" (prompt typed in a common online generation website) or more involved process based on an initially generated image, that was later modified and improved, possibly with AI tools? (which I guess would be most AI-art being professionally made).

The answer to these will give different answers to your question, ranging for "incohrent number of fingers" (which was a good sign to tell low-effort generations, because while you're right that humans can be bad at drawing hands, there is a small chance they'd be able to draw perfect hands... and yet make it with six fingers on the left hand and four on the right hand, as AI often do) to the more involved one I proposed earlier.
 
Last edited:

I had nothing to do with that crow image apart from seeing it pop up on several crow based Facebook pages, without it being flagged as AI. The focus plane is pretty much in keeping with how a real close-up shot of the sort would look and I'm guessing whoever created the script for the image also purposefully introduced artificial motion blur to the crow on the right. There was a fair bit of effort put into it.

OK, I'd have said the blur on the crow on the right, on the same plane as the first row, was an artifact. But your explanation is very possible.


Like I said it was just a hunch, because SD photorealistic models tend to produce unwanted blur more often than not when not prompted not to do that.
 
Last edited:

Ryujin

Legend
Because there is no unique answer to your question, first because you didn't define AI art.

I can't identify for sure the raven image that was posted before as AI, but I had a hunch because it was trying to be photorealistic but had a very strong blur in the distance, which is a flaw of many models of a specific implementation of an image-generation AI. Nothing conclusive, just a hunch. But it was a method effective because it was a generation of a photograph-like image, made from scratch. It wouldn't have worked if there was a real photograph being improved/modified/touched up with AI. Would it be AI-art in your definition? The means to detect AI art depends (even if the case there are ones) depend on how the AI works, and therefore we need more definition to answer your question effectively.

Also, one does need to know the goal behind the question, because the answer will depend on it. If it's just to win "guess if it's real or AI generated" contests so prevalent these days, a wide test, with acceptable accuracy, will be a good answer. If it is something with higher stake, like sueing someone into oblivion, you'd need a 100% effective test that would modify the answer. Without this information, one need to rephrase your question to answer it. Or accept that the question can't be answered, which isn't leading to a discussion.

I proposed to ask, for each image, variants of the same character including some details that would be very difficult to have an AI engine trained on, like eating sushi upside down. It is not perfect of course, but it would be a hard test of consistency (a current weak point of many generative AIs) and it would be easier for a carbon-based artist to start back from a previous sketch. Inability to do that would make me tell something is AI art. Does it work for your use-case?
Such blur can be introduced, in a real image, via an in-camera filter. It occurs naturally when using wide aperture lenses and is magnified when using extreme close focus, as with a macro lens. That, in itself, isn't a give-away that something is AI. I have obtained similar results, if not exactly the same, myself on many occasions. Even the motorcycle picture that I posted shows a similar effect, from a fairly wide aperture lens, despite it being shot with a telephoto lens at some distance. If I'd shot it at F2.8, rather than F4.5, the effect would be pronounced enough that the 2nd through 4th bikes would be so out of focus that you couldn't read their numbers.
 

Remove ads

Top