Raw LLM Responses

I think there is a very simple solution we are all missing on this issue, but first, a few valid points to correct your perspective: -- The data sets are not created or controlled by AI developers, so blaming David Holz or the guys at Stability AI will solve NOTHING because *they do not create the data sets.* -- AI image generators (diffusion models) do not * I repeat * DO * NOT * steal people's artwork or mix bits of existing works. They look at images in the data set as a *reference point,* then algorithmically construct a new image (for those of you familiar with Filters and Fractal Noise, you may understand what I'm getting at there. Im a 3D artist and I work with procedurals all day. I use them to recreate the look of granite, steel, dirt, plant leaves, clouds, terrain formations, etc. Am I stealing from nature when I recreate the look of it using procedurals? No. I am programming the computer to simulate those materials mathematically in a way that is more accurate than what I could produce freehand. THIS is how AI image generators work too... I've created hundreds of images using Midjourney (which I merely use for reference variations) and every single upscale has similar telltale "puzzle pattern" or "brush stroke" bits in it... Which are very visually similar to Perlin or Worley distributions. This boils down to being the digital equivalent of how * human artists * can go to Google, Artstation, or Deviantart, CGSociety, etc, and screenshot or download other people's artworks and use them as reference when creating their own work. Legally speaking, as long as the new image is at least 70% unique, it can be protected under copyright. THIS IS NOT ILLEGAL unless the "artist" tries to claim a plaigeristic work as their own. BUT: Because AI-generated images are typically made and stored on a server, it should be very easy to merely differentiate AI-generated text-to-image works from original handmade artwork in terms of metadata, and simply pass legislation preventing AI-generated work from being copyrightable.