I once made a program that analysed text and put up a frequency table for letter triplets. Using this table I then created nonsense text. A sample text file of 1 kB was enough to produce a table that could create some nonsense text that unmisstakeabely was English. Or German, Swedish or what ever language I sampled.

If you have sampled some megas of plain English text, you would have a pretty accurate table of triplet frequencies. Then you could sample another mega of text and pick sentences from that text. For each sentence, you pick each triplet and sum its frequency. You have to divide the sum with the number of triplets in the sentence, which gives you an average. Compute the average of all averages, and the standard deviation of the averages. This will give a picture of what is expected in real English. The standard deviation tells how reliable this method is. This method would OTOH recognise my nonsense English as real English.

Johan