Tinker a little with the number of words and the threshold for detection of duplicates, and you're probably almost there for a large share of the copy-pasta spam.
I experimented with n-grams a little bit and couldn't find a good value. Low n yields too many false positives, high n doesn't detect spinners, etc. So I'm using a mixture of algorithms and base the decision on the pattern of the results of those algorithms - e.g. if the similarity of two texts using algorithm A is 70%, then union/intersect/otherwise manipulate the texts, run algorithm B, if it scores 90% then run algorithm C to eliminate false positives - made up numbers but you get the idea. Works ok-ish, but as I mentioned it doesn't scale well and I need to do more testing on larger samples.
The difficulty will be to find sources to match against (unsure if scraping Google will be permitted, we'll see).
Google has a search API. Not sure if there is a free tier though.