Well, yeah! It's a probalistic model, and extremely biased - it has to be, so that it can predict the correct token.