What Makes Good Writing?

Part II: A Tale of Two Openings
In Part I of this series (see Speakeasy#8), I laid out some basic principles of what constitutes good writing, including (but not limited to) variety, rhythm, precision, momentum, originality, and humor. In order for Heresy Press to consider a submission, the text must work first and foremost on the sentence level.

Since Heresy Press is re-opening its submission window on December 1, 2024, the following is meant as guidance to fledgling writers and aspiring novelists, in order to increase their chances of passing the triage stage and making it from the slush pile onto the desk of editors, awaiting serious examination.

I cannot stress enough that the opening sentences of a submission have to be flawlessly written: polished, rhythmically pleasurable, free of needless words, and, above all, vivid. To illustrate a good and a bad example of what I mean, I provide two different opening passages of roughly the same length:

1. “Leaning forward against the weight of the pack, his skillet lashed over the groundsheet and rolled blankets on top, he walked sometimes on the wood of the crossties, sometimes on the sharp rocks of the ballast.”
2. “He switched off the lights, but they continued burning brightly.
   Odd, he thought. What do you know about that?
   He took a second swipe, but his hand raked air—the switch was gone.”*

Pause for a moment to decide which opening passage presents the greater incentive to keep reading.

Dynamic writing draws energy from full verbs. But in example 1 we encounter two lengthy adverbial phrases containing present and past participles—“leaning,” “lashed,” and “rolled”—before the main clause (“he walked…”) arrives, thus draining the sentence of its forward momentum. There’s a general sense of clutter and wordiness (“on the wood of the crossties” should simply read “on the wooden crossties.”). A bit of description would also be helpful to let the reader know who is walking there. I further wonder if it can be taken for granted that a general readership would know the term “ballast,” derived from railroad jargon. Finally, there’s no striking figure of speech in this sentence.

Now compare this to the second opening passage: First of all, full verbs predominate (“switched,” “continued,” “thought,” “took,” “raked”). And while the personal pronoun “he” is also used without descriptive qualifiers, this is compensated for by the presence of “voice” in the internal monologue—“what do you know about that?”—an idiomatic turn of phrase indicative of personality. Moreover, the strong verb “raked” in the fourth sentence serves as an evocative metaphor. Victor Shklovsky called this function of figurative language “defamiliarization,” insofar as a well-chosen figure of speech lends a mundane object or action a new, unexpected aspect. While sentence 1 labors under deadweight—wordy adverbials and prepositional phrases—the second passage propels us forward into the rest of the novel, something further indicated by the dash.

* Quoted from Deadpan by Richard Walter.

Unleashing AI on the Slush Pile—Boon or Bane?

By Bernard Schweizer
One of the big topics at this year’s Frankfurt Book Fair was the role of AI in literature and publishing. The industry’s response to AI tools is mixed, ranging from bullish embrace in some quarters—mainly on the production, translation, advertising and data transformation sides—to skepticism over AI’s potential to generate genuine artistic content, and even to outright opposition against the illegal training of Large Language Models on copyrighted content.

Critics of generative AI maintain that “AI struggles to ascertain relevance or context, an instead employs hyperbole or a false sense of urgency to impart meaning or relevance” (Ed Nawotka, Publishers Weekly). But if AI is not (yet) able to write literary-grade material, what about AI’s ability to discern creative talent among works submitted to publishers? In other words, could AI be tasked to sift through the dreaded slush pile, isolating the pearls and weeding out the work of hacks and amateurs? This would save both time and money since assessing a 100,000 word manuscripts can take up to 30 hours, easily costing $1000 in terms of salary. By contrast, running a 100,000 word manuscript through AI costs the publisher $5 and takes only minutes.

But would such an AI slush-pile reader be any good? To find out I agreed to do an experiment with the intention of vetting a commercially available AI editor (which shall remain anonymous in this essay). I asked the system to run five texts:

• two novels that I had already published (The Hermit and Animal: Notes from a Labyrinth)
• two texts that I had rejected due to stylistic gaffes and various literary shortcomings
• one manuscript that had me sitting on the fence.

A major outcome of this experiment is that the system failed to detect any difference between raw submissions and highly polished, already-published texts. Not only did the system recommend four of the manuscripts—including two that I had roundly rejected—for acceptance after only “minor revisions,” but the one manuscript that it deemed in need of “major revisions” was one that had already been published.

WEAKNESSES OF THE AI READER:

1. In terms of style, the algorithm could not distinguish between flat-footed sentences and sentences that sing. The system ended up praising the writing of one sample as “strong,” “sophisticated” and “high quality” even though the style was convoluted, illogical, riddled with awkward wording, compromised by needless adjectives and adverbs, and sprinkled with non-sequiturs.
2. The system lacks real critical discernment and skews positive. How useful is a slush pile reader that more or less likes everything that you give it? And if it does wax critical, it tends to find fault with the wrong things! For instance, it highly praised a work that was smooth but unexciting and predictable, featuring flat characters, but it criticized a real work of art—Animal—because the system couldn’t appreciate the book’s unconventional structure, taking issue with its deliberate discontinuity and kaleidoscopic, experimental form. The system also objected to the use of profanities and other strong language in Animal while suggesting that the text be submitted to a sensitivity reader because of the presence of a transsexual character. It appears that this AI has been programmed to make left-progressive (woke) judgments.
3. The system is primed to make suggestions for improvement even when the work is already fully developed, polished, and—as near as it can be—perfect. I have no doubt that this AI would find areas of improvement in Hamlet. But when over-confidence is paired with lack of discernment we have a problem. Not only are some of its suggestions redundant, but others are counterproductive. With regard to Katarina Grishakova’s The Hermit, the system suggested that in the interest of “balance” the main character should be more attuned to the positive aspects of his life as a Wall Street trader (The Hermit is a tale of disillusionment). This got me to thinking: if a copy of Dostoyevsky’s Notes from Underground were put through the system, it would probably recommend that the “Underground Man” be made more well-adjusted and mentally balanced. But would we still want to read Notes from Underground if the protagonist were a jolly good fellow?
4. Finally, the system gave high marks to one text that I had rejected, indicating that “the interactions and family dynamics are presented in a way that feels genuine and relatable.” How a machine can find a literary text “relatable” is another question, but what the AI system found “relatable” was contradicted by two readers reports that criticized the novel for its outdated dialogue and a certain inauthenticity in character interactions. What evaluation am I inclined to trust more—the human assessment or the machine-generated one? The question answers itself.

What’s needed, obviously, is an algorithm sensitive to literary style. The ideal editorial AI would be programmed to fulfill Salman Rushdie’s expectations for valuable literature: “Great art … is never created in the safe middle ground but always at the edge … it challenges, questions, overturns assumptions, unsettles moral codes … it can be shocking, or ugly, or controversial.” But since today’s AI programmers have not even solved the problem of false positives, they are obviously not about to filter out only prose fiction that crackles with Rushdie-worthy originality and subversion.

What they can do to improve the performance of this AI in the short term is to give it parameters as to what kind of work a publisher is looking for. Is this a press that caters to a specific audience? Is this a press that wants to mollify its readers and avoid ruffling any feathers? Or is this a press more like Heresy that takes on work normally passed over by the mainstream publishers, placing literary quality above ideological requirements? In other words, the developers should instruct the AI to look for aspects of “fit” between submission and publisher.

Despite its obvious shortcomings, the AI system was not completely devoid of merit: It was really good at generating informed summaries of the content it was fed; it gave surprisingly insightful analyses of the main characters; it was spot on in diagnosing the correct literary genre; and it generated catchy promotional blurbs. In addition, the system produced a longer list of comparative titles (and reasons why they relate to the work at hand) than I could come up with.

My verdict is this: If tasked with the job to simply digest submissions by synthesizing their content and providing competent summaries, I give it an A-. If tasked to spot a promising gem among the majority of mediocre submission in the slush pile, it gets a D. Can the system be trained to perform better by approaching a work of art as a living, unpredictable, revolutionary, and sometimes even “ugly” thing? I doubt it, and something in me hopes that the answer will always remain a firm “no.”

While the two-legged acquisitions editor is thus clearly superior to its algorithmic counterpart, should AI ever become the go-to tool for assessing literary submissions—mainly because of the enormous savings in cost and time—I foresee a future in which the literary marketplace will be flooded with the kind of mediocre stuff that not only gets high marks from AI but that may actually be written by AI. A world in which literature is written and judged by AI strikes me as dystopian at best. The scorched-earth digitization of human creativity is a folly worth resisting.