It's mainly used for various kinds of black hat SEO (comment spam, automatically generating thousand-page websites and so on). From what I've seen, the quality ranges from atrocious (e.g. simply running random words through a thesaurus) to passable, but the goal is just to make the text look reasonably unique to a machine, not to skirt copyright law. If you were given the output and a list of potential source texts, you would not have any trouble figuring out which was the parent.