Comment by GTP

We do know that the open web consitutes the bulk of the trainig data, although we don't get to know the specific webpages that got used. Plus some more selected sources, like books, of which again we only know that those are books but not which books were used. So it's just a matter of probability that there was a good amount of SEO spam as well.