The OpenAI samples are leaps and bounds ahead of traditional Markov-chain generated text. I don't think you can compare the two. It's the fluency and plausibility that gives pause around a public release.

I agree with your last point though - it falls into the same category as paid Russian trolls. I think that's exactly why they were hesitant to release the pre-trained models - they didn't want to make it easier/cheaper for a bad actor to replicate the 2016 election.

It remains to be seen whether their decision will make an iota of a difference. But I understand their motivation.