From my own dog fooding, I was able to come to these conclusions:
- if you provide no input, it's going to fabricate (trying as much as possible to fill in all the blanks)
- if you provide a decent amount of input, then it'll just act as a tool to fill out missing words
The second part is a bit tricky, and it's what we've tried to improve on with the UI. Even if you provide enough info, it can still be too terse to begin with. And there's a chance it'll fabricate things.
That's why we've added tools to edit any output and to expand on sections. The idea being to get you 90% of the way there, and allow you to add finishing touches and edits and still be truthful. Basically, an iterative process.
Admittedly, it is work in progress.
Do you have any examples of fabrications and stuff that was so off the mark compared to what you would've expected? Appreciate it's a big ask, but genuinely curious.