November 2023

ChatGPT Darling Killer

Michael Bettencourt | Scene4 Magazine

Michael Bettencourt

While the writing advice to "kill your darlings" has a roster of many fathers (Faulkner, Wilde, King, Chesterton, Welty, Chekov), its true sire is Arthur Quiller-Couch, who coined it in "On Style," a 1914 lecture included in his published 1913-1914 Cambridge lectures, "On the Art of Writing," where he inveighed against "extraneous Ornament":

    If you here require a practical rule of me, I will present you with this: Whenever you feel an impulse to perpetrate a piece of exceptionally fine writing, obey it—whole-heartedly—and delete it before sending your manuscript to press. Murder your darlings.[1]

These days, available to all writers, is a free and competent darling-killer: ChatGPT. (I'm using version 3.5 with a free account at OpenAI; ChatGPT 4 is available through the Bing search engine, which, as an aside, is a pleasure to use for search.)

First, some caveats. ChatGPT does not think, at least as a human thinks. Which means that when you, as the writer, prompt it to "streamline and simplify the following text," it makes its editorial decisions based on comparing the text you offer against its training materials, not on anything like the lived experience of a multidecade editor with acquired tastes and biases. In other words, it has no cognitive or aesthetic context for what it offers back to you, which is only the results of the statistical probabilities generated by its algorithm.

It follows, then, that ChatGPT does not have a "voice" in the way that that word is understood regarding human writers. If you asked it to describe its voice, it might likely reply that it is ChatGPT and therefore cannot have a voice, but it can list for you the constituent elements of a voice and what others have said about what it is and why it is important.

Given these limits, however, what I've found that it does very well is what I call "stress-test" my writing, that is, cleanse it of tics and looseness that get in the way of saying what I want to say as clearly and as quickly as I can say it.


A quick example of what I mean: I ran the above sentence through ChatGPT with this prompt:

    Please use your decades-long editorial expertise to streamline and simplify the following sentence: "Given these limits, however, what I've found that it does very well is what I call "stress-test" my writing, that is , cleanse it of tics and looseness that get in the way of saying what I want to say as clearly and as quickly as I can say it."

The result:

    Despite these limitations, it excels at stress-testing my writing, helping me eliminate any stylistic quirks or verbosity that hinder clear and concise communication.

I like this but want to keep something of the original, so this is the revised sentence:

    Despite these limitations, it excels at "stress-testing" my writing, helping me eliminate any stylistic tics or looseness that hinder clear and concise communication.

ChatGPT and I have partnered to create something that I think is better than the original. More on this later.


In other words, it planes down my style in a very Strunk & White way:
"A sentence should contain no unnecessary words, a paragraph no unnecessary sentences, for the same reason that a drawing should have no unnecessary lines and a machine no unnecessary parts."

Another and longer example. Here are the first two paragraphs from an essay I wrote about how the term "participatory fascism" applies to modern American politics:

    It Can't Happen Here by Sinclair Lewis came into the public domain in 2023, so I wanted to see if it could be adapted to something. (It has been adapted as a stage play but not a movie – and it had great success as a play.) I romped through the book, whose tone I would describe as whimsically vicious with a side order of jeremiad, and it certainly gave a good feel for the fears and frustrations of the times as people confronted an ascendant authoritarianism and many said, "Hmm, that doesn't sound so bad."

    Much like George Orwell, Lewis conceived of the authoritarian regime as based on physical punishment and intellectual oppression designed to flatten all differences while still using those differences to divide and conquer (such as Windrip's abolition of the Negro). Orwell crystallized his vision of how this authoritarianism would work in the image of a boot stamping on a human face forever, and Lewis follows in that same vein. (163 words)

When I finished the essay, I ran it through ChatGPT, and here is what it did with those opening paragraphs:

    "It Can't Happen Here" by Sinclair Lewis entered the public domain in 2023, prompting my curiosity about potential adaptations. The book, characterized by a whimsical viciousness with jeremiad undertones, gives a good feel for the fears and frustrations of an era grappling with rising authoritarianism. It's akin to George Orwell's vision, depicting a regime that employs physical punishment and intellectual suppression to both homogenize and divide, similar to Windrip's abolition of the Negro. (73 words)

I quite like the changes:

It deleted digressions or digression-like instances to keep the writing on the trunk line. For instance, gone is the parenthetical statement about other adaptations, and it brings the mention of how Windrop abolished the Negro from behind the parentheses to nicely round off the paragraph.

 It tightened the prose. My third sentence kind of lollops-along, trying for a little humor ("romped"), more or less gets around to describing the book's tone ("whimsically vicious with a side order of jeremiad"—and why "side order"? what purpose does the culinary reference serve?), and ambles its way to a soft landing with a bit of ventriloquism thrown in for a faux flair.

And I lollygag into the next paragraph as if the reader has nothing but time to devote to my maunderings.

ChatGPT made short work of that. It distilled the beginning of the third sentence into a very nice concentrate, saying better what I was mucking about trying to say at all: "...characterized by a whimsical viciousness with jeremiad undertones…" Perfect.

Not content with that, it took the rest of that sentence and the whole of the next paragraph and concisioned them neatly.

Notice that it deleted my reference to Orwell's image of a boot stamping on a human face, but later, ChatGPT brought that image back in a very neat way.

I was making a case that Aldous Huxley, not Orwell, had the better argument about how the repression would happen: through spectacle rather than assault, through worrying people about losing their comforts rather than by taking those comforts away. Of course, I was wending my way through my presentation—wending here, wending there, wending back and around—but ChatGPT made straight the way and connected Orwell and the boots sharply to cement my point:

    Meanwhile, we can keep all the trappings of a democracy as long as those trappings are limited to voting, making modest campaign donations (while dark money floods the system), engaging in elections as spectacles, and indulging in excessive consumerism while drowning in debt.

    In this scenario, it's not the jackboot on the face but the allure of 50% off Doc Martens that prevails.

My original sentence was "Not the jackboot on the face but Doc Martens for 50% off while they last." ChatGPT added the bit that my prose needed to make the point stick: "In this scenario…that prevails."

Does this bother me? Not a bit, for a few reasons.

I've used ChatGPT to conjure material from scratch for my job as a communications person for a university's development department. For instance, I want a solicitation email with a particular theme requesting a donation, and with very few exceptions, ChatGPT will generate functional but not inspired prose no matter how precise I make the prompt.

And no wonder: because it's drawing from thousands of solicitation examples for its output, ranging from the mediocre to the sublime, it will give me the useful mean, the serviceable average, without any assessment of its effectiveness. Remember: ChatGPT is not built to be creative but to give you the most probable statistical outcome that answers the prompt you ask.

So, I don't use it for that. Instead, I use it as an assistant to brainstorm options to phrases, sentences, or documents when I'm stuck. It usually doesn't unknot the problem outright, but if I tweak the prompt to regenerate responses, ChatGPT often provides more options than I could come up with on my own, and out of that welter, something usually arises that gives me a way forward that I wouldn't have thought of myself.

ChatGPT is a superb summarizer. I am still astounded by the way it can take a complex document and extract its core points in a clean prose. It is also a good summarizer if you want a range of opinions or options. Let's say I want to find the five best practices for writing a good solicitation email, so I prompt it to do that, and it gives me that information (as best as it can based on its training). You can then ask it to explain more deeply any one of these practices, and it will (again, limited by its training).

If I wanted to, I could sharpen the prompt to say that not only do I want the thumbnail descriptions but that I also want ChatGPT to extract from each practice its top three principles and the top three ways to put those principles into action, and to put all of that into a table that I can use for a presentation. And it will do that (again, limited to its training).

These results are not gospel; you need to verify them. But you also save so much time doing it this way rather than opening up Google and firing off your search terms and then trolling through the results and so on.

ChatGPT is good for generating email subject lines and pre-header text based on your parameters: it must say X, or it must say X in the first three words, or it should have a joke, and so on. And it will generate as many of them as you want (though, after two or three cycles, it just begins recycling what it's already produced—even ChatGPT can get bored).

In short, ChatGPT, built as it is right now, is a ready and willing editorial assistant. It helps me keep my writing honest, it takes on some of the more prosaic tasks in my work (I mean, generating 10 to 12 subject lines for every email quickly drains me of anything original or catchy), and, when needed, can tutor me on some aspect of something more efficiently than if do the search myself.

It's difficult to know how to end this essay because no one at this point can identify all the ripples generative AI will cause as it falls into the pool of human life. Generative AI is/will be dangerous, magical, disruptive, surprising, revealing, obscurantist, pedestrian, and so on and so on and so on. I just know that I have found a tool that, within the limits that bind it, has prompted me to reconsider things in ways that add value to what I produce without nullifying my own mind and soul.

Will it become so sentient that it will take over my job? In my job as a writer of fundraising documents, I usually rely on confined and tested approaches that are probably algorithmable, and I could see them simulating what I do well enough to put me out of a job. (Some are even suggesting that AI could handle the routine bits of being a CEO—that will happen as well.)

The dangers of generative AI, like ChatGPT, lie more in systems that make decisions about people's lives, which Dan Quillen points out brilliantly in his book Resisting AI. In that realm, people should rightly fear how AI will be used to preserve the current power relations in our society and, quite literally, determine who will live and who will die.

But at my lower level, I'm content with the technology and look forward to how it is going to upset and unseat the conventional wisdoms we indulge about life, the universe and everything.

[1] -what-writer-really-said-to-murder-your-babies.html


Share This Page

View readers' comments in Letters to the Editor


Michael Bettencourt is an essayist and a playwright.
He writes a monthly column and is
a Senior Writer and columnist for Scene4.
Continued thanks to his "prime mate"
and wife, María-Beatriz.
For more of his columns, articles, and media,
check the Archives.

©2023 Michael Bettencourt
©2023 Publication Scene4 Magazine




and creates


November 2023

  Sections Cover · This Issue · inFocus · inView · inSight · Perspectives · Special Issues
  Columns Adler · Alenier · Alpaugh · Bettencourt · Jones · Luce · Marcott · Walsh 
  Information Masthead · Your Support · Prior Issues · Submissions · Archives · Books
  Connections Contact Us · Comments · Subscribe · Advertising · Privacy · Terms · Letters

|  Search Issue | Search Archives | Share Page |

Scene4 (ISSN 1932-3603), published monthly by Scene4 Magazine
of Arts and Culture. Copyright © 2000-2023 Aviar-Dka Ltd – Aviar Media Llc.

Thai Airways at Scene4 Magazine