This @spawning_ project lets artists search for their art to see if it was used as AI training data. It’s also an excellent educational tool for demystifying the spectacle of image generation… haveibeentrained.com/

Try putting key words from any prompt you would use in DALLE2, for example, and you’ll see the images associated with training for that image. True “man behind the curtain” moment.

One thing that process makes clear is how relatively little impact a single, or even a hundred, photographs or pieces of art have on the generated image. (That should take nothing away from anyone’s right to opt out, of course.)

But it does show how much the machine recalculates / recalibrates from the original training data. That will open up some exciting avenues for both artistic research & critically engaged art making.

People who suspect that diffusion models just copy and paste images into new arrangements will be very confused when they compare “spirit photography” as a prompt output (left) and as training data (right).

On the other hand, as we poke around, we might find all sorts of strongly correlated associations. This data has been out there already, but kudos for sure for giving it the design features and simplicity that the site offers up.

Anyway follow @spawning_ and some of these folks because this is step one of creating a “do not train with my art” list ⬇️