The Savage Editorbot

When I set AI loose as an editor of my writing, the bot immediately gave me examples of its limitations.

I asked an LLM to edit the first chapter of my book, a comic novel written from the first-person perspective of its protagonist. I asked the bot to keep the voice and tone, but aside from that, I let it decide the extent of editing required.

The first red flag popped up when the bot described what it intended to do. It suggested a line edit, with the goal of “tightening 15-20% of the prose, sharpening the transitions, and trusting the humor more.” I know that AI editors have a bias toward compression, and knowing that it had a target percentage for cuts before it even started, I wasn’t optimistic.

As I expected, the bot was ruthless in pursuit of its predetermined word count. In just the first two pages (which is all the bot wanted to edit, although I had asked it to edit the entire chapter), it cut an entire third of the text. That, in itself, isn’t automatically bad. Maybe I had been unnecessarily verbose. But in my humble opinion as the author, the bot cut some pretty significant stuff.

For example, it changed this:

“I waved my hand in a serpentine motion, then flipped both open hands, fingers toward the floor, as if I were delivering an armload of stuff to the ground.”

Into this:

“I traced a winding path through the air and finished with a flourish.”

The bot clearly doesn’t like similes. But if it had taken the time to read the entire chapter, it might have realized that the narrator’s flair for the dramatic, both in the way he behaves and in the way he tells the story, are quite important, and beginning to demonstrate that on the first page might not be a bad idea.

The bot had warned me that it was going to be intolerant of repetition, especially with the phrase “middle of nowhere,” which it picked out as a particular offender.

So it changed:

“’But it’s not in the middle of nowhere,’ I said, my own voice straining with the effort of persuasion. ‘That’s the beauty of it. That’s the secret. Seems like it’s in the middle of nowhere. Isn’t in the middle of nowhere. See the hook?’”

to:

“‘But that’s the thing. It looks like it’s in the middle of nowhere. It isn’t. That’s the secret.’”

Maybe it’s just me, but I think the second version loses a little something.

It hacked away at expressive words and excess explanation, so:

“She closed one eye and scrunched up the corner of her mouth. This was the face she made when she wanted me to know, without doubt, that she thought I was crazy; it was a signal that she was still in good humor but that she might lose patience with me soon. I pushed onward.”

became:

“She closed one eye and twisted the corner of her mouth. This expression meant two things simultaneously: she thought I was crazy, and she still found me amusing. The second condition rarely lasted long.”

That one eviscerated the meaning of my original passage, but it eliminated 21 words.

This next one, you’re going to have to explain to me. The bot changed:

“Christina is something of a critic of the Ivy League, having gone to graduate school at a large state university in the southeast, and she seizes every opportunity she can to hold my Harvard education against me.”

to:

“Christina has strong opinions about the Ivy League, having attended graduate school at a large state university in the Southeast. She regards my Harvard education the way some people regard hereditary diseases.”

It doesn’t make sense, but it’s shorter, for sure.

Then there’s the inability of an LLM to follow a thought through a text. In the first page of my chapter, there were two uses of the word “refrigerator.” The second use is a punchline that’s a callback to the first use of word. The bot, for some reason, deleted the first “refrigerator” but left in the second one, thereby nullifying the joke and rendering the second “refrigerator” nonsensical. LLMs, you see, have neither a sense of humor nor a memory.

The bot probably would have been willing to edit my whole chapter if I’d begged, but at this point, I’d had enough. I left the LLM to keep doing what it does for other writers.  I, however, am going to use human editors from now on.