horror stories written by AI
This year for Halloween I dressed my dog up as a banana and decided to see if I could make my computer scare people for me. I mostly took spooky stories and used them to train a word-level RNN and was prepared for some horrific results.
Although I wasn’t expecting much, some of the generated excerpts are surprisingly good. They don’t quite make sense, but they read like they could be old-timey English and have a bit of a creepy vibe.
Frankenstein — “i saw bones from ignominy”
I chose this Shelley classic for a book report when I was 10 and instantly regretted it. Since I had to take a few double takes on passages from the original text to understand them, I was actually impressed when Frankenstein AI generated passages that didn’t mean anything but sound like they might if you just think about them hard enough. All the punctuation was removed to cut down on word vocabulary for the network to learn so its just stream of consciousness. Here’s some generated excerpts:
livid from the plank and wept that guided me i was not considerably shortened by the same advantages in the same manner and i was obliged to relinquish least amassed the whole skeleton i saw bones from ignominy i had not despaired my father schemes was a unadept a scene i was sustained and dismissed me into understanding i had been condemned at the same time and i was not acquainted with a chair and appeared to me the most delightful peril and intellect i was not recording the following eagerness to restore me i had not die to the
I enjoyed the nature imagery below, especially since phrases like ‘the wind and rain’ aren’t even in the original text. ‘Determining to consummate the cottage sea’ seems weird, but you do you, Frankenstein AI.
own behaviour i was not the same eccentricities determining to consummate the cottage sea and was beginning in the trees and infused injury in the cottage and the anticipation of the antique tackle and the sun cracked in the trees and explained with me i had been the same narration the sun was chastened by the seashore i had displayed to the same opinion madame against me to comply with the same volumes which had been deferred by the wind and rain had been adduced against me and i was not acquainted adrift deceived me to be procured and i
Some of my other favorite generated phrases include:
selfeducated from the streets of remorse
i was indifferent out in my own bosom and fuel up by the bitterest remorse
i had not despaired the wind had intervened to be my own benefit to consider the creaking of my mind
i am ugly as i was dreaming of anguish
When I first did a round of training, I used a subset of the book Dracula to test things out and was amazed by the results until I searched the book and realized that it was spitting out the original text word for word #overfittingfails.
The Fall of the House of Usher — ‘leave me i have not not not not not not not not not not’
I tried using some Edgar Allan Poe and trained with The Fall of the House of Usher, but this book is significantly shorter and only have a vocabulary of ~2k words (compared to ~6k in Frankenstein). When I used the same model with the same training time it generated some pretty messed up text:
demeanour a reserve and had palpable pass away us not dared not speak and had been not not shall not dared not speak and had been not not perish not himself i had been not not perish not himself i are length by leave me i have not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not issued upon the lady madeline which was i
After more training it was a little better, but still repeating itself and saying things like ‘the horrible opening of the lady madeline liquid’ :
solemn hours i had fixedly bore in the roof of the building which i learned that the lady madeline which may be the breaking around of the house of the building which i learned that the lady madeline which may be the breaking around of his feet i had been used me in an peculiar melody while like this distant fancy instinctively and horrible opening of the lady madeline liquid and trespassed demeanour in a zigzag direction with the building which i can compare to fall indulged the solace he had away i could not dared perish noticed the lady
I didn’t run this with a combination of stories just because training time took so long, but might yield some interesting results. On github you can check out the jupyter notebook I used, largely based on this tutorial on machine learning mastery and training data which are mostly books from Project Gutenberg with extra info stripped.
P.S. If you’re looking to be really scared by AI this Halloween, check out Boston Dynamics’ creepy dog robots that will haunt your dreams.
