The human writers are constrained by what they think they can get away with. The automated writers don't even work that way. You can see from the examples used

The human writers are constrained by what they think they can get away with. The automated writers don't even work that way. You can see from the examples used in discussions here, the more limited the domain of discourse the more convincing the samples are (except the hello world one that looked like a mash-up of every programming tutorial ever written). I think this thing could do a great job making fake announcements to air travelers about flight delays. I think the synthetic texts are already convincing enough that the natural desire to try to impart meaning to writings/utterances will go a long way to convincing forum user "victims" that something is being said that needs to be thought about. And here in this discussion, we know that the fakes are fakes (except for one comment which is unlabeled and I'm not sure about), in an aggression against a forum the fakes won't be labeled.