GPT is pretty reliable at taking an existing body of text that fits entirely in a prompt and doing something with it--reword, translate, answer questions about it, so forth. Even expanding an outline into written sentences usually works great as long as the outline has all the detail.

It's really only when you let it introduce "facts" based on its body of training that it'll start going off the rails. If he's just running text through without checking the final results, that's really not wise, but the LLM is unlikely to start introducing novel facts or changing figures in response to a conservative request based entirely on the prompt text.

Using it as a writing aid the way he claims he is (not that I necessarily believe all his claims) is probably one of the more responsibles way to use GPT, honestly, because you have full opportunity and ability to vet the output. Hack together something that has all the right content, then make it readable with the AI engine the way you might with a human editor collaborating, while jointly making sure the editor doesn't change the actual facts of the content.

Whether or not he could be hacking anything of value together every 37 hours is another story.