It just trains the program to look for similarities, and then make the program mash the images together.

For example, you show the program an image of an Apple and label it “Apple”. The program doesn’t “know” that it’s an apple, it just looks for patterns that looks like that image. If you make the program go through thousands of images, then it might have thousands of ways of drawing an apple.

In the same way, you might make the program “learn” a “bear eating”. So if you input “a bear eating an apple”, then you might output an image where it might look like a bear is about to eat an apple.

Obviously, humans are much more flexible than that. All it has to do is learn what an apple and a bear is, and it can draw an infinite pattern of art with those two items. Or it can be a completely new style of art. Fact is, we don’t “learn” how to draw stick figures for example, yet we do. We don’t just imitate reality like the AI does.

We THINK that we’re drawing people when we’re drawing stick figures. They’re our imperfect yet approximate figures of humans. We’re not copying, we’re estimating. That’s what it means by drawing from scratch.