It's not, but the author did say they have used this test against models when they come out. So it's possible that put the unpublished text into the training da

It's not, but the author did say they have used this test against models when they come out. So it's possible that put the unpublished text into the training data for the next model, somehow linked back to the author's identity