I apologize, but that was my understanding. The AI is trained on images and then draws upon those images when the prompts are given. Maybe I'm mistaken, but I was under the impression that it basically creates an image from what it was trained on. My logic went that it's taking bits and pieces from existing images to make the new one, so maybe I misunderstood the process or I chose the wrong word when I said remix.
As for Googling it, Google has been increasingly frustrating to use ever since people figured out how to game their algorithms. I do a search for "nsfw ai art generator paid subscription" and I get half a dozen articles from mainstream commentators and then three dozen spam links. It's not the be all and end all of search engines that it used to be.
Yes, google is getting more useless by the year.
Anyways, you're correct - it uses bits and pieces of from existing images when it checks its own accuracy in the feedback loop.
All these programs are is simply a filtering machine which tries different inputs, passes them through a filter, then checks the result against an existing image, compares, then adjusts the filter.
As a very simplified example, say a program is meant to generate a human face, so it starts by saying "I think there's a tan pixel at x=100, y=100, with an HSV of 34,33,82
. My filter says its neighbors must be within 2 shades hue: 32 < H < 36 - lets try a value of 33 for its neighbors. The program then compares that to a real image, noting the real image's pixel at that location has neighbors with an average of 3 shades of hue difference. It plugs that back into its own filter, and tries again with the new numbers, and then compares again. It does this again and again with many different real images to try and adjust its filters to produce the best replications.
So no, it's not neccesarily "starting" from an existing image, but without existing images to provide feedback, the whole process wouldn't work.
As Woody554 summarized well earlier in the thread, there's no actual "thinking" going on as we would understand it, it's just iterating and interpolating, again and again and again. "Artifical intelligence" is kind of a misnomer, I'd call it more data processing, analysis, and interpolation.