Sep 2, 2021
The power of surreal logics: spawning crime story ideas with OpenAI
© The Guy Bourdin Estate, 2014/Courtesy A+C
Imagine the following (dull) crime scene:
The victim is lying in a pool of blood. A whisky glass can be seen, shattered to pieces. There’s a letter lying next to the body.
Take a few seconds to imagine what might have happened there.
Now let’s evaluate the following hypothesis:
The victim was killed with a broken whisky glass. The letter is written to the victim’s husband, who is an alcoholic. When he fails to respond to the letter, the letter is mailed to him again, with the whisky glass.
You might be somewhat puzzled by this explanation. It is, in fact, AI-generated.
Welcome to the GPT-3-empowered outskirts of the Real.
Welcome, to OpenAI Beta.
Here, users are offered a “playground” where they can play and experiment with some flabbergasting GPT-3 models (or engines) that can understand and generate natural language. The model can be used for anything, from creating recipes based on some ingredients to talking with verbose and virtuoso virtual assistants.
But back to the whisky.
As a fiction writer, I find the engine’s suggestion intruiging. The elements of the crime-scene (the letter, the whisky glass, the body) are connected with an inherent logic that goes beyond the human intuition. One can feel the leap, the total lack of conventional “thinking”. When working at detective and crime stories, this is of course what you’re looking for. What the model suggested is improbable, but not at all unlikely. In a strange way, it seduces me, and as a writer I am challenged to take it seriously.
If I want, I can even bend its surrealism into an interesting plot: a succesful politician receives a written warning from someone who threatens to disclose his secret alcohol addiction. The man ignores the letter and burns it. A few weeks later, a parcel with the same letter and a provocative whiskey glass is sent to him, but alas, his wife gets to open it. When the politician, already in an agitated state, gets home and hears what his wife has discovered, he believes his wife is behind the ruse, and…
Whether this story is worth it to be written or not, is another question. That’s also not the purpose of this exercise. In the title, I use the word spawn on purpose: new crime scene compositions can be fed into the engine again and again. Every combination will generate unique ideas, some of which can be transformed into a pitch, a dramatic scene or just an interesting idea for a character’s development in another story you’re working on.
Consider it an extra dimension to brainstorming. What’s especially powerful about the engines is their intrinsic necessity to inject meaning into the hypothesis. It almost feels like a greed, or some kind of fatal necessity… When the elements are so random that a realistic coherence is out of the question, the results can therefore be twisted, surreal or unwillingly poetic, like an aurora borealis of meaning.
Those of you who start working with these GPT-3 models will quickly understand how important it is to feed the engine with qualitative data. In fact, the poorer the sample text or instruction, the weaker and more repetitive the results. Crime setting ideas that are far too obscure on the other hand, lead to useless hypotheses, or the incoherent details are just left out. It’s a matter of balance and trial and error.
Now let’s look at some other AI-generated answers to crime riddles I came up with. I believe each of them can be incorporated in creative writing practices, either to spice up existing stories or kick off new ones.
Solve a murder mystery.
Crime scene: a dead man, a bloody knife, flowers, a sleeping dog, an open fridge, an unfinished painting.
What happened?
The man was killed with the knife by the woman. She painted the picture of the dead man, then put flowers on his body. The dog woke up and ate part of the painting.
The woman killed the man and hid the body.
Apart from the goofy detail about the dog, I love the idea that the woman first killed and then painted the victim. The chronology of the death and the painting seem to be somehow mixed up here, and I like the unexpected romantic and surreal outcome.
Solve a murder mystery.
Crime scene: the victim is found next to a river. He shows no signs of drowning. There are white feathers on his body. He has no shoes. A bike is found a bit further.
What happened here?
A man came to the river to try to wash away his sins. He fell in and drowned.
The victim’s feet were recently washed.
The victim was trying to wash away his sins. The river washed the sins away, and the man drowned.
The engine correlated the bare feet (and probably) the feathers into something spiritual. Rather than a crime story, this could become a poem or a beautiful short story about a man who committed a crime and came to grips with his sins. The tragic ending is a cylical answer to his own past.
Solve a murder mystery.
Crime scene: Two men are found dead in a forest. They are naked and both have scars and wounds on their legs and back. They are holding each other.
What happened here?
The men were injured by a bear. The bear then found the man’s clothes and ran away.
Another fine example of how the engine breaks the scheme of expectations. Rather than a gay hate crime, we might have two friends here who wents swimming and bumped into a bear. They must have clung to each other in their last moment, in a desperate attempt to protect each other. Again, it’s not about the story that I will write based on this answer, but more about how this particular image stimulates me as a writer. It is stored somewhere in my own mental notebook and might one day find its way into a short story, in a completely different form.
Earlier on, I referred to the results of this exercise as seductive, much in the sense of Baudrillard’s concept of seduction and the zero signifier: “That which doesn’t mean anything has every reason never to be forgotten.” With the OpenAI playground, the fiction writer can see this at work. The results the GPT-3 models come up with are sometimes weird and twisted, but also often surprisingly humane and touching.
