I've been looking over google's patents on this topic (and plan on writing more in detail when I've a better notion of how they all fit together.)
But the use of a hashes seems one of their top methods.
As I understand it, they won't try to look at and compare words. Rather they'll create a hash based on things like words per sentence, sentences per paragraph, paragraphs per article, sentences per article, etc.
This way, an article spun merely by replacing one word with another would have roughly the same (identical?) hash value.
That's why I really want to see article spinning software that can delete sentences and paragraphs, change sentence-ending puncuation to semi-colons (changes number of sentences and sentence length), and shuffle parts of different articles together.
The says of simply spinning words are over.