IMO, the only thing to be taken seriously with text generators should be natural language processing.

• take this fat block of text and give me a bullet point list.
• what are synonyms for X?
• copy-paste a big TOS and tell me the key takeaways that are anti-customer.
• take these documents and make one coherent document about one page long.
• etc.

The problem is that even with things like this, it frequently fails because it hyperfixates on some details while completely glossing over others, and it’s completely random if it does that or if it’s good, and this uncertainty basically necessitates that you check everything it outputs, negating much of the productivity that you gain.

I once used it for a Python script, and I generated three outputs out of these three generations only. One regex function ended up in my real script, but I got the idea to use regex from it. And I used its output, which actually worked.