AI: The Blade Itself Incites to Deeds of Hackery

We’re peppered with marketing slop aimed at convincing us that AI is the way forward, that it will inevitably surpass human capabilities, and soon. For the purposes of writing and editing, there is no evidence to support that claim. I expect that there never will be, and that the reasons for that limitation are inherent to AI.

I don’t truck with generative AI or LLMs. Not as an editor of manuscripts, not as a writer of blog posts, not as a father of grade-school kids. I’ve encouraged my sons to replace “AI” with “heroin” whenever they encounter it at school. “Dude, I cranked out my whole paper on heroin. Wonder if the teacher can tell.” “It’s okay to get some starting ideas using heroin, but we’ll be doing some writing in class, too. Without–eyes on me, people–without using heroin.”

The technical limitations of generative AI are enough for me to dismiss it as an editing tool. My objections on principle are even sturdier and less subject to the effects of technological innovation. Each relies on an interested layman’s understanding of how LLMs work and what we can expect from them.

What We Talk About When we Talk About AI

Large language models are a subset of AI. For the purposes of writing and editing, AI is typically a synecdoche for LLMs like GPT and Claude. AI agents–largely autonomous software systems that can refine their understanding of initial conditions and interact with other digital systems to complete complex tasks–move LLMs like Claude a step or ten closer to something resembling AI proper. But that’s beside the point for our purposes. Call it a concession to a desperate spree of debt-fueled marketing, but there’s no point in pretending that AI and LLM aren’t operationally interchangeable terms here.

LLMs work by reducing text to a series of tokens–word fragments of four characters or so, along with individually tokenized symbols and punctuation characters. Peanut is tokenized by GPT-5, for instance, as pe anut. Each token is assigned an index number, which itself is assigned a vector: a series of numbers each of which represents a token’s place in one of many dimensions. A typical vector these days might have 384 dimensions represented by 384 different numbers–that’s the current default for common applications of Claude. Each dimension represents a different galaxy of associations and proximities. Peanut might be associated closely with butter in one dimension and with Tillman in another. (In the second dimension, they should also be closely related to NFL Hall of Fame.) Adding dimensions theoretically adds nuance, but at the cost of computing effort, time, and money.

Vector dimensions are calculated against a pre-training corpus of material fed into the LLM. This corpus can be anything, but it helps if it’s massive. Pre-training can take the better part of a year, at which point the LLM still needs to be fine-tuned, typically with some human supervision. This fine-tuning typically includes human scoring of the LLM’s response against target output that has been confirmed by human operators to be valid, appropriately written, and relevant. Reinforcement learning does away with target responses and provides human or algorithmic feedback on the LLM’s output; over time, this helps the system avoid poorly scored responses. The process continues until the LLM has reached its benchmarks.

All of which makes LLMs powerful tools for things like pharmaceutical research and materials science. Feed an LLM a corpus of pre-training material representing all the biochemical data you can put your hands on, for example, and it can suggest research paths that greatly reduce the trial and error associated with discovering new drugs. Good for the researchers, who get to focus more quickly on a more relevant range of options, and good for us.

But what’s behind LLMs used to facilitate writing and editing? The ideal pre-training corpus in our field would be everything ever written, and for reasons I’ll get into below, that still wouldn’t be enough. Still, general-purpose LLMs are typically pre-trained on corpora consisting of trillions of words. Books are just one source of that material; social media posts are a notoriously significant contributor, as are public datasets and simple web crawling.

And who fine-tunes that sort of LLM? Post-training is increasingly conducted with a minimum of active human supervision (of course that’s the trend). Under algorithmic post-training schemes like direct preference optimization, preference datasets (sub-corpora of model output) take the place of human scorers and formulae replace human feedback. Who chooses those?

Before any editor uses an LLM, they should have compelling answers to those questions–answers that clearly support the editor’s professional goals. Which, yep, is another way of saying that no editor should use an LLM.

Why LLMs Can’t Edit

An LLM’s editing or writing work, then, amounts to a calculation of the likeliest relationships among semantic fragments, confirmed by an analysis of a group of texts whose criteria for selection are notoriously loose and all but undocumented. It gets there by way of a genuinely impressive array of evolving methods, but at the end of the day it’s just an expensive parlor trick.

And not even a terribly convincing one. Generative AI might be advanced enough to pass a Turing test…but only on the test’s original 1950 terms, with its original 1950 subjects. As the presence of generative AI has grown, so has our ability to see it for what it is. The way it clamps on to certain constructions (“not X but Y”). Its inability to generate real rhythm. Its addiction to lists and simple paragraphs. The em dashes–good lord, the em dashes! These aren’t just potshots at a still-emerging technology; they’re implicit confirmations that we still know what good writing looks like. How else would we know how stilted and childish AI-generated writing seems? Who knew that the Turing test was adaptive?

---

It follows that all we’ve ever known about this game is that human sensibilities outpace digital computations for the purposes of writing and editing, especially the writing and editing of fiction. Every LLM developed for our field is better read than any human author or editor on earth. None of them knows, say, the harrowing architecture of Walter Kempowski’s All for Nothing: the seemingly random encounters that slowly take terrifying shape, the winnowing of everyone and everything down to a young boy and his savior or executioner, the immense tension in the last few lines.

Not, that is, until some human reads the book, reflects on its structure, writes up those thoughts, publishes them such that they appear in digital form, and has them scraped and stolen.

Because, like all fiction, All for Nothing turns on the themes of love and death. Few novels make that turning so explicit, and many leave at least one of those themes implicit. But every work of fiction, like every aesthetic creation, reflects George Steiner’s conclusion in Grammars of Creation that “without the arts, the human psyche would stand naked in the face of personal extinction. Wherein would be the logic of madness and despair.” Aside from authentic religious experience, Steiner claims, only artistic creation “enable[s] humans to participate in the truth-fiction, in the pragmatic metaphor of eternity, of liberation from the eradicating dictates of biological-historical time, which is to say: death.”

Unlike, say, music, fiction appeals directly to the impermanent things of our world–the objects, the people, our experiences and our reflections on them–for its material. Every work of fiction enables our liberation from the dictates of time by describing and exploring various forms and modes of attachment to the world and the people in it. That attachment, given shape and meaning within the scope of a life or a story, is love.

Digital systems do not share our awareness of death or time, or the ingenuity that frees us, however fleetingly, on however narrowly scoped a basis, through attachment to the people and things of our perishing world. No artificial simulacrum of human intelligence is capable of appreciating those factors, which just happen to be central to us wetbrains.

AI is a Hack. Not the Cool Kind.

Scott Miller (of Game Theory and The Loud Family) is my favorite songwriter.¹ His fan base is devoted but small, and some of his lyrics wrestle with the space where the full scope of his fame should have been. At some point years ago–I can’t find it in the archives of his wonderful Ask Scott exchanges, so it might’ve been chatter before a show–he said the single most valuable thing I’ve ever heard about popularity. Necessarily paraphrasing and ploddifying here:

Popular music isn’t what everyone’s listening to. You can’t make more of that and affect people the same way. Truly popular stuff is what people want to hear but don’t quite know how to ask for.

That’s the sweet spot that we all want to hit. Its opposite is hack work.

By hack work, I mean writing that’s conceived and produced specifically to meet the author’s idea of what will sell. This doesn’t mean that good authors are completely oblivious to commercial concerns. Books that sell allow other books to be published. If you want lots of readers, if you want to shift units, you’re not ignoble and you’re not a hack. Not necessarily.

You might be writing potboilers, and potboilers aren’t hack work. Off the top of my head, Dostoevsky and Dickens wrote plenty of potboilers, but their commercial and artistic sensibilities were distinct. They had a good idea of what would sell, and they knew that they could strike a chord that the reading public would find intriguing even though they hadn’t explicitly demanded it. Besides, there’s nothing wrong with keeping a roof over your head and food on the table.

Or you could be very good at writing within a particular genre, in a particular style. You know that genre down to its pores, and so do most of your readers. When you drum up a new story, do you keep it within the confines of your previous work? The same characters only more typical, the same plotlines only more predictable? If so, you’ve fallen prey to hack-thinking, and to some extent it probably bothers you. So you get a little impatient with what’s merely worked before, and have a little fun by tweaking your own expectations of the game. Problem solved, hack work avoided. Your readers will have fun with your little subversions, too.

And for what it’s worth you’ll have that much in common with just about every great writer ever. Squint at Dickens and he’s pure melodrama; his eye for detail balances out the sugar. His much better contemporary, George Eliot–in whose future works Dickens expressed “a perfect confidence in their making me wiser and better” before he’d finished her first book–knew all the formulas and tropes so well that she was able to question them and at times turn them back on themselves without losing her readers’ attention.

Dickens knew what it was like to be poor and abandoned; Eliot knew what it was like to be an odd middle-class duck. Those factors trace through their works like scars left by a lightning strike, and they probably point to an important reason why those two well-read kids grew up to be such artfully impatient writers, such knowing tweakers of traditional modes. But what if they hadn’t had such distinctly formative childhoods? What if they hadn’t had childhoods at all? What if you gave Kaspar Hauser three trillion words to read and asked him to produce a novel? What if you then gave him a draft of your book and asked him what he thought? Hauser’s long gone, but we’ve got AI.

AI can only write to the highest degrees of linguistic probability within the constraints of the directions it’s given. In that central way, it’s by definition a hack. Use it to write, and you’re asking for hack work. Use it to edit, and you’re asking to make your own writing more like hack work. It might be saleable, but it’ll never grab a human soul. And it’ll never be truly popular in the healthiest and most enduring sense.

Better than a random selection of all the writing a VC-fueled monstrosity can steal, you want an editor who’s read widely and with some discrimination. Who knows why they like what they like and has a sense of the distance between that stuff and writing that leaves them cold. Whose can read with empathy and professional discernment even things not in their wheelhouse–who, to borrow myself as an example, can understand what makes Dickens special even while vastly preferring Eliot. And who can respond to your writing with a thoughtful eye toward making it the new and precious thing it can be. The kind of editor who puts some bones and blood into it.

1. I’d say something provocative like “my idea of the greatest songwriter in the Western popular idiom,” and I’d mean it, but that’d invite arguments over worthies like Cole Porter and Robert Schumann and Guillaume de Machaud. To be fair, Scott added passages of ars nova to some of his songs, and I’ve never heard a note of jangle pop in Machaud’s. ↩︎