I found it was easy to get mind-blowing results on a single dataset - e.g. 99% accuracy for working out who wrote which emails in the Enron [0] dataset.

The moment the test set was even slightly cross-genre, results became bad.

So this is great in theory, but I would love to see a test set from the same authors outside the context of the Google Code Jam. I'd be surprised if the results were anything like as good.

There are very few real-world cases where you have a large amount of same-genre data for the questioned author. (No one writes a thousand suicide notes / ransom demands / scams)

[0] A large email database often used for authorship tasks https://www.cs.cmu.edu/~enron/