I'm working on a much longer article on the topic of all the shortcomings of using large language models to help you write porn, but I just wrote a prompt on JanitorAI that totally sent the script into a tailspin.
I'm using the LLM to help me structure and build a story about sexual politics in NYC before and during the 2020 lockdown.
The main love triangle in the story is between the first-person character (Jester,) his aspiring-yoga-teacher girlfriend, Zephyr, and Zephyr's hot bisexual Korean-American roommate who calls herself Jenna.
This is the prompt I've written:
(It's three nights later. Zephyr and I are naked in bed together after another night of intense love-making.)
*Still, I don't let it go. A few nights later when Zephyr and I are lying in each other's arms, our bodies cooling from another bout of intense sex. As casually as I can, I ask* So, have you and Jenna ever hooked up?
*Zephyr smirks down at me and asks, "Why? Do you want to hook up with my hot Korean roommate?"*
*Something in her tone gives me pause.* Somehow, that feels like a trick question with no right answer.
The added subtext is that Zephyr and Jenna grew up together on an artist's collective where non-traditional relationships were somewhat common. Zephyr is afraid that Jester is only dating her because he expects her to be into three-ways and such.
The LLM doesn't get that subtext, but it also doesn't punt. Instead, as I've tried to modify that prompt to get the story back on track, I've had the bot assume (on subsequent runs) that (a) I am the one asking the question about the hot Korean roommate (b) I am speaking to another person off-camera and Zephyr herself is now my hot Korean roommate (c) I am a hot Korean girl instead of the white guy I've been all story.
The short moral of this posting is that, the more I learn about LLMs, the less I worry about them taking over the world. With a lot of coaching, they can be helpful for structuring and pacing, but their biggest use for me has been in prompting me to really think about what story I'm trying to tell when they start telling an entirely different one.
(On the plus side, I do now have four short stories is various states of composition that I used LLMs to pre-draft. That's four more short stories than I've been able to write in years.)