Can AI generate movie pitches that are indistinguishable from (or even better than) real movies?
That’s what I’m going to find out. To do so, I am using AI to generate a bunch of fake movie pitches and gathering a bunch of real ones. Then, I’m going to present them to real people and have them answer two questions about each pitch:
- Would you watch this movie?
- Is this pitch “real or robot?”
From there, the results will be tabulated and hopefully I’ll have answers to a few questions. Can people tell if a pitch is real? Can AI generate pitches for good movies? Does people care?
This is, of course, a dumb product.
There’s no revenue potential. The humor is likely limited to just me and my friends. And it will probably take more work than anticipated.
The upside to doing dumb projects is since there’s no real pressure to succeed, there’s a lot of room to experiment and fail and learn.
The origin of this project dates back several years when I got into using LSTM neural networks to try and generate text. My first project was to see if I could clone my personal Twitter account and see if I could replicate myself with a robot. The second was a pitch generator. Neither achieved results worth sharing despite the process taking weeks. Back then I learned that realistic text generation was difficult.
Or it was before GPT3.
One last twist
In my notes I started referring to this project as “Pitchmaster 2000.” It turns out, there’s a short film from 1993 that’s shockingly similar. The synopsis?
“At the fictitious New Century Studios so many scripts come in every day that the acceptance, processing and assessment of them have been completely automated.” 
And yes, it is also similar to South Park S08E05 “Awesom-O” where Cartman pretends to be a robot and just spits out generic movie pitches, most of which star Adam Sandler.