jump to navigation

When AI is the Author December 9, 2019

Posted by Peter Varhol in Algorithms, Machine Learning.
Tags: , ,
add a comment

I have been a professional writer (among many other things) since 1988.  I’ve certainly written over a couple thousand articles and blog posts, and in my free time have authored two fiction thrillers, with several more on the way.  I have found over the years that I write fast, clearly, and when called for, imaginatively.

Now there are machine learning systems that can do all of that.

I knew that more recent ML systems had been able to automatically write news articles for publication on news sites.  At the beginning of this year, OpenAI, a research foundation committed to ethical uses of AI, announced that they had produced a text generator so good that they weren’t going to release the trained model, fearful that it would be used to spread fake news.  They instead released a much smaller model for researchers to experiment with, as well as a technical paper.

Today, though, they are using the same GPT-2 model to create creative works such as fiction and poetry.  I started wondering if there was nothing machine learning could not do, or at least mimic.

But there’s a catch.  The ML systems have to be given a starting point.  That starting point seems to be the beginning of the story, which has to be provided to these systems.  Once they have a starting point, they can come up with follow-on sentences that can be both creative and factual.

But they can’t do so without that starting point.  They can provide the middle, and perhaps the end, although I suspect that the end would have to be tailored toward a particular circumstance.

Decades ago, I read a short story titled “In Medias Res”.  That refers to a writing technique where the actual writing begins in the middle of the story, and fills in the backstory through several possible techniques, such as flashback.  In this case, however, it was about a commercial writer who could only write the middle of stories.  Other writers, who could see their beginning and end, but not the middle, hired him to write the middle of their stories.  He was troubled that he always came in the middle, and was incapable of writing a complete story.

While I’m guessing that ML techniques will eventually be good enough to compose much of our news and fiction, the advances of today are only capable of in medias res.  So I will continue writing.