Pandemic Without a Vaccine

Yes. I said, “Pandemic.” Go ahead and cringe. The 2020s have given us more exposure to this word than anyone rightly deserves in their lifetime. But the pandemic I’m speaking of could be just as devastating to author careers as the dreaded “C-word” was to human lives. I’m talking about AI. Go ahead and cringe again. I just did.

As I prepared my manuscript for a new round of professional editing, I made the mistake (not really a mistake) of catching up on the publishing news making rounds in the community, and all I can say is, “Wow.” Well, that’s a lie; I have much more to say.

It centers around three harrowing publishing stories that should make us all shake. The first was the abrupt cancellation of Mia Ballard’s horror novel, Shy Girl, after it was accused of being AI-generated. The complaints came from amateur reviewers and were broadcast across multiple platforms. The second was a short story by author Jamir Nazir called, The Serpent in the Grove, which won the Commonwealth Short Story Prize this year. Again, the online community called out the writing as AI-generated based on suspicious patterns in the prose. The third case involved debut author Jerry Falade’s $2.4 million book deal for his work Call Me, I’ll Hide the Body, which was withdrawn by two of the Big Five houses this year. From what I can tell, that accusation of AI-generated text was made in-house by publishers.

These examples are terrifying.  And in my mind, for one single reason above all: AI-detection software. The first two examples relied heavily on this analysis, used to critique, in one case, a pirated manuscript—all for what these products advertise as: confident detection of AI-writing. How confident? Damned if I know. Well, that’s a lie; I have some receipts.

I don't use AI to write manuscripts. Period. Coming from an academic background, I find the notion revolting. I do, however, use Grammarly to identify errors like improper punctuation, unintended tense shifts, and ridiculous sentence structures that aren’t even recognized as sentences. I also use it as a spell checker.

So, can I be confident my prose won’t be tagged as robotic? I decided to try it out. I bit the bullet and bought a month-long subscription to Originality.ai, one of the more respected AI-detection platforms, and pumped multiple chapters through its meat grinder. The results amused, shocked, and frightened me all in the same heartbeat.

The overall assessment of my polished manuscript is that it’s 80% likely a human being wrote it. 80%? WTF? I guess my computer took over while I was asleep—which coincidentally amounts to 20% of the day for me. To be fair, the percentages don’t work that way. It’s not saying 80% of the manuscript is me and the other 20% is AI. It’s making an overall assessment that there’s a 20% chance AI was used somewhere. But where does that number come from? Hold on tight. Rough waters ahead.

As part of its analysis, Originality.ai provides a detailed, line-by-line breakdown, assigning an AI-probability score to each sentence—or whatever segments its algorithm determines. This is where things start to get strange. Here are five example phrases from one of my paragraphs: 

• It was a chaotic free-for-all from [protagonist]’s point of view, but she knew her friend well and suspected an underlying method, only an artist could appreciate, in the madness she encountered.

• [Friend] was an artist in the purest form.

• She delved into virtually every medium.

  
  

• It didn’t matter. [Protagonist] recognized them all.

• Pencil drawing, charcoal, oil painting, watercolors, clay sculpting, oil pastel and acrylic painting, photography, and even metal sculpting.

And here are the results:

• First passage: Likely original (99% confidence). Well, thank you. The sentence is a bit clunky. That's probably the reason. Should I rewrite it? Will it reduce my humanity score?

• Second passage: Likely AI (91% confidence). Umm.

• Third passage: Likely AI (94% confidence). :o(

• Fourth passage: Likely AI (74% confidence). Sigh.

• Fifth passage: Likely original (56% confidence). So why not 57%? Or 58%?

  
  

If you want to get real crazy, consider this innocuous-seeming bit of prose:

•  “No. I see your point.”

  
  

The word “No” is likely original (100% confidence). So go ahead and use this wonderful interjection with confidence (I think. I’m not a lawyer). The phrase, “I see your point,” is a different story: Likely original (54% confidence).

These results should be chilling to all of us. None of these phrases fall under the AI-writing forensic indicators we’ve been warned about. No rule-of-threes, no nonsensical metaphors, no “It’s not X, it’s Y.” It seems that my 80% likelihood of being human is derived from a collection of individual, line-level judgments that themselves seem nonsensical. Is this how the Big Five are analyzing manuscripts? I sure as hell hope not.

So what are we to do? If you know, please tell us. There are years and years of hard work out there, stuck in laptops, created through anguish, tears, and sacrifice, all prepared to come forth into the world’s eye, and their fates depend heavily on sketchy procedures and inhuman declarations of humanity. Until the industry comes up with a vaccine, tread carefully and wear your writer’s mask.

P.S. One of the paragraphs in this post came from a prompt asking ChatGPT to reword my original paragraph. Can you guess which paragraph? I bet not. And I’m not telling anyone either. In real writing, no one could ever possibly know. Take that!