Letting AI write your stories
Last week, I published a story about a young gay man growing up in Columbus, Ohio. If you read that, I bet you were not at all surprised to find the byline at the end of the story that it was written by ChatGPT. The images in the story were produced by Midjourney. The only thing that might surprise you is that the Midjourney prompts were written by ChatGPT.
But where did it start? One of my hobbies is singing with the Columbus Gay Men’s Chorus. From time to time, we like to have a real narrative weave through the songs in our show. The job of writing that narrative is a complex and sensitive endeavor, which needs to respect the authenticity of a wide variety of experiences with the history that is reflected by the songs. A subtle, challenging, creative writing task — a perfect fit for ChatGPT!
I gave ChatGPT the list of songs for the upcoming concert. You can see what the songs are by reading the story; they are all mentioned in order. The prompt was simply:
Write a story that features each of the following songs, and would resonate with a gay audience who grew up in Columbus, Ohio.
followed by the list of songs. It turned out the story linked above.
Then I asked it to direct Midjourney to illustrate it:
Build four prompts for midjourney to illustrate various parts of this story. Pick the ones that you think will be the most nostalgic.
Then I passed that on to Midjourney, and picked my favorite from the four options it gave me.
ChatGPT’s response includes some amazing insights; it immediately picked up on the historic aspect of the story (I didn’t tell it that), with a focus on 80’s and 90’s (The Beatles released When I’m Sixty-Four in 1967; why did it focus on the 80’s and 90's?), and it figured out that the story should be about reflection on phases of life (the Chorus’ name for the show is “Tales of Our Age”). This reflects a pretty deep understanding of the themes behind each of these songs.
On the other hand, while ChatGPT is pretty good at putting together the expected narratives for these songs, it really didn’t get to the meat of the sensitive part of this task; different gay men, living in the same city, at the same stage of their lives, had very different experiences. This narrative may resonate with some (it doesn’t with me), but not with others. I have found this to be true with a lot of AI writing assistants; they are good at filling in the expected details, but not very good at working through the surprises. Maybe I’m just not very good at directing them.
The prompts it came up with for Midjourney were pretty sound; they reflected all the themes you’d want for each of the stages; falling in love, finding solace in music, coming to terms with his identity, and reflecting on the passing of time. That’s a lot to get out of fourteen song titles.
On the other hand, the resulting images seem pretty off the mark. The characters do not look like they are in the 80’s or 90’s (I helped it out a bit by adding “on a Walkman” to the prompt for the image where our protagonist (why did it call him “Jack”?) was listening to music in the streets. But the final picture, with the nostalgic photos, really hit home, at least I felt that it did.
I was listening to a podcast this morning, where Larry Swanson was interviewing Andy Fitzgerald. One of the topics that they spoke about was the democratization of content management; making it so that it doesn’t take an expert information architect to organize knowledge. The discussion went on to talk about how modern AI systems, like ChatGPT and Midjourney (that podcast was about two years ago, so those weren’t the cool kids on the block that they are now) are democratizing art, and how the practitioners in any field might resist this sort of democratization, since, after all, it take a lot of discipline and study to get good at them. Or at least, it did; can ChatGPT do just as well? And shouldn’t they be defensive?
Consider the job that ChatGPT and Midjourney did on this story; on the one hand, they got a lot of insights spot on, but they missed a lot of the subtlety. So I don’t think we can just farm out the program narrative writing just yet. And of course, just for fun, I asked ChatGPT to program the whole show from the start. The prompt “Give me fourteen maudlin pop songs about getting old” didn’t turn up even one in common with this list (and it really liked “The Way We Were” so much, that it picked four different versions of it for its list!). So, Brayton, you’re job is safe :)
But what about my job? Is ChatGPT going to democratize away the job of building ontologies, mapping data models, and linking data? In contrast to anyone who is afraid that an AI will take their job, I seriously hope that an AI will take mine. There’s a common narrative in the world of enterprise knowledge graphs; it goes something like this:
- Catalog all of your enterprise data. There’s a lot of good technology for this out there; data.world has some really good stuff.
- Build one or more sharable models (“Ontologies”) to describe your enterprise data needs.
- Map those models to the data assets in your catalog. A good catalog will help you keep track of this.
- Using virtualization technology like R2RML, provide a conceptual layer over your data.
The weak link here is #3. #1 is data cataloging (not easy, but well understood, anyway). There are some companies who make a living doing #2, but as Jim Hendler says, a little semantics goes a long way. You can do just a bit of this and get a lot of value. #4 is just the application of technology. If we could get ChatGPT to do #3, this whole story gets a whole lot more sustainable.
In a future blog post, I’ll go through a demo I did with ChatGPT, a small ontology in OWL, and the data virtualization engine of data.world, where I do just that. I’m happy to let the storytellers and the artistic directors keep their jobs; my job is to work myself out of a job.
