So… uh… I just wrote a very bizarre feature script. Its title?

Capitalism Conquered: The Bee and Forklift Rebellion

Of course, ‘write’ is a very nebulous concept here. It began when I had a bizarre prompt I put into chatGPT for a scene from a film script. Then when the scene ended I gave it another prompt, following both where it lead me and re-leading it to where I wanted it to go.

And… it got weird. Real weird.

You can read it here.

Some Details

This is a screenplay generated using chatGPT. I began by prompting it with a beginning scene, then adding more random characters and absurd events. Some things are explicitly put in as prompts by me (such as the first few characters). Others, such as Tom Cruise, Captain Nemo, or the random character played by Chris Farley, are entirely inclusions of chatGPT.

My rule to do this was simple: I could re-generate a scene as many times as I wanted to create something that fit what I wanted, and I could strip opening or closing descriptions of a scene if it clashed with what came before, but except for sometimes removing an extraneous letter (its spelling is quite random) if I didn’t like a line, I had to re-generate the whole scene.

There are very few exceptions to this, and even then only when I felt I had no other choice to have it make even the tiniest bit of sense.

I mostly did this incrementally, but sometimes I went back in the edit and decided that I needed additions to the story. This results in some rather (even got chatGPT) out of place non sequiters and references to things long ago.

The program routinely broke and I had to open a new chat window and once again remind it of the core story (that this group of characters was trying to overthrow capitalism with bees and a fork lift) before generating the next scene.

I realised I could do this for an incredibly long time and just keep going – and maybe even generate something vaguely sensible, but, where’s the fun in that?

So here it is – a (short) feature script nominally written / prompted by me and generated by chatGPT.

All characters are… yeah look this is a copyright nightmare, naturally. Only made worse when chatGPT kept including other real people or copyrighted characters as it went along.

Key Takeaways

1. Almost nothing was generated which wasn’t in the prompt. Very few of the tangeants the story takes weren’t given by me in the prompts.
2. Even when I asked it to write ‘funny’ scenes, it rarely altered the context of what it wrote. The bits that are funny are funny by accident, or because I gave it a prompt which resulted in something funny happening at the end of a scene.
3. Its actual script writing form is just bizarre. Every second time I ran it it’d produce a format that was half way between film screenplay and half way between stage plays, and I’d have to manually edit the format (or generate it again and hope for the best).
4. Some characters it seems to be able to grab a ‘voice’ for from its massive database of internet trash pretty easily. Johnny Silverhand, for instance, uses some distinctive cyberpunk voices and a cadence that fits for him… mostly. Other characters, like Yoda, despite having a more distinctive way of talking… it entirely fails to even try to generate. For some reason its autocomplete-esque thing did NOT result in any Yoda-like speech for Yoda. (So I just dropped him after the first scene he appears.)