This post is part of LifehackersExposing AIseries.
It’s not going anywhere anytime soon.
Most AI art online is circulating without a clear label warning users the images are not real.
That’s not really what’s happening here, though.
AI art generatorsareindeed trained on huge datasets of images, from art to photographs.
However, these tools cannot “see” these images the way humans do.
Instead, they break these images down pixel by pixel.
It won’t see an apple; it will see a group of pixels and their values.
Overtime, it will learn that this grouping of these particular pixel values tends to be an apple.
Modern AI image generators, like DALL-E,use what’s known as diffusionto learn and generate images.
Count the fingers
AI art is getting better, but it still isn’t perfect.
While these tools are capable of generating high-quality images with realistic lighting, they still struggle with smaller details.
Perhaps most famously, “human” subjects in AI art tend to have the wrong number of fingers.
In any case, fingers and the hands they’re attached to are often messed up.
Any repeating patterns, especially when it concerns details with humans, can be generated incorrectly with AI art.
Most of us don’t haveperfectsmiles, but AI teeth are on another level.
It doesn’t knowhowhands actually workit can only pull from the hands it does see.
We’re experiencing the limits of that knowledge here.
But it’s not just the subjects: Everything in the image is fair game for this blending.
Look at the image I generated with DALL-E below.
(Not to mention, her fingers are blending into one.)
In many cases, AI art that incorporates writing will lookoff.
Now, this side of AI art is rapidly improving.
It’s producing art based on the relationships it’s forged from all its training data.
It pulls from its training to replicate these requests to the best of its abilities.
Look toward the background, and things get even weirder.
A man gives the impression of kneeling, drinking his red cup as if it were a bottle.
Who is building cat sculptures out of Oreos?
Why are there so many sad artists making sandcastles of Jesus?
Images are often overexposed, or feature dramatic lighting, which makes subjects particularly bright.
(Even if the subject’s hands weren’t off, as well.)
Or to stop elements in the photo from blending together?
Or to stop adding weird shit to the backgrounds of images?
