Interesting preprint paper comparing human writing to AI-generated writing
A new paper by Jenna Russell, Rishanth Rajendhran, Chau Minh Pham, Mohit Iyyer, John Wieting from the University of Maryland and Google DeepMind dropped. They analysed over 60 000 stories written by both humans and AIs. A ginormous sample.
In the paper, AI stories were detected as being more direct and clear. This ended up in flat storylines, clean narratives, and explicit morals.
Human authors have been shown to write implicitly, allowing their texts to be understood in different ways or using their texts as vehicles for researching or expanding the writing prompts, not handing meaning over to the reader.
More importantly, human authors show greater diversity.
The differences are not found just in the writing styles used (from my experience, humans that write clearly get unfairly flagged as using AI), but the narrative structures beneath the surface styles as well.
StoryScope: Investigating idiosyncrasies in AI fiction paper from arXiv
My take
This analysis speaks to the importance of the human experience in writing and the use of subtext in fiction. Machines treat text as statistics, that’s how they predict output. That’s not how you write emotionally impactful stories. You don’t get to math your way into linguistic mastery.
As a reader, I want to read stories that challenge me. I want messy streams of consciousness, original voices, and none of those things can be mimicked. If it’s not there, there’s no amount of machine play-pretend that can come close. Writing fiction is not what machine learning is best used for anyway.
