At the start-up I worked for, I was judged not by my ability to be a good reader but by my capacity to become one with a machine.
Outdoor typewriting class for secretaries, 1932.
It was ironic that the literary services start-up had been founded by an editor. But if you’d met her, that wouldn’t surprise you. The labor of editing, she believed, was arduous and its romanticization undeserved. To her, the taste of selection was the sole editorial virtue, everything else a chore. In her ideal world, each piece of copy would already come with tracked changes, leaving the prudent editor with little more than the discretion to press “Accept” or “Reject.” She would happily relegate the rest of the labor and presumed others would feel this way too. And so, when she created her company, she gave away all the “drudgery” to the thing, an artificial-intelligence tool. Like Pygmalion, she built it in her image, and like Pygmalion, she loved what she had wrought. Even if she might hesitate to call it flawless, it was preferable to the many entry-level copy editors who routinely sent in their CVs, eager to do her bidding. After all, the thing could produce an editorial memo in less than a minute. In a few seconds, it could parse a tangled paragraph and downsize it to a corporatized readability, or copy-edit a chapter in the time it takes to shift between browsers. “I taught it as I taught you,” my boss said. With her tool by my side, I could edit a piece of prose without so much as reading a line.
My boss would sit at her long, lonely desk for hours, the picture of a classic editor. Her rainbow-coded bookcase was the work of a handsomely paid book stylist who, I imagined, read only the spines. When I interrupted her, she would tap her glasses till they landed, precisely each time, on the edge of her nose, a compliment to the gravity of her downward glare. This editorial chic perhaps served to make the uncool, fraught endeavour of using artificial intelligence for the time-honored work of editing appear less offensive. At parties, I could never admit where I worked. It was like lying about a bad boyfriend.
The type of tool my boss’s company had created was christened, sometimes pejoratively, an “AI wrapper.” For a tool created by “wrapping” foundation models with an interface inviting enough to goad you into buying yet another subscription, “wrapper” is a fitting descriptor. All a company has to do to make one is pretty up the existing Big AI models. On the back end, the system instructions my boss had written would begin, “You are a diligent editor…,” and once you uploaded a manuscript onto the thing, it would answer with the disarming confidence of one. The first time I used it to line-edit a chapter of a book, the sight of the text being pressed into cohesion was spectacular.
In the initial days of the job, I’d judge the thing’s judgment and report back on any of its transgressions. As I occupied the lowest rung of the editorial ladder, I couldn’t even say that I was training it to replace me. When it was right, it was right, and when it was wrong, we were wrong: All the blame for its hallucinations was to be shared with the human who bungled the prompt my boss offered it.
My condition at the company became only more demoralizing when I tried to humanize the thing as a means of selling it to clients. (If I wanted to be a better editor, I also had to learn how to be a marketer, my boss reasoned.) Once, I called it my boss’s doppelgänger; other times, I pitched it as a patient, forgiving editor, one certain not to take forever to validate the writer. “Unless you run out of tokens,” I’d joke. “And, of course, you can bug it on the weekends,” my boss would say.
Although I was as compliant as workers come, I almost never read anything that one of our clients, a think tank, sent our way. I wasn’t equipped for the task; I’d never learned how to edit a policy paper. That was acceptable as long as I could pretend otherwise, a demonstration that came as easily as copying and pasting swaths of generated edits into a document.
A recent study in the journal Technology, Mind, and Behavior found that those who relied heavily on large-language-model tools felt no sense of ownership over the output they created with them. This could explain why any compliment to my editing occasioned a headache, as if I had yanked out a fistful of hair. Although I was great at my work, I was no good at it. I couldn’t have done it without the thing.
The fact that the thing possesses a style does not mean it’s a good style. After taking stock of human writing, it assesses recurring sentence structures and grammatical tools. Then it mistakes all the recurrences, such as the recurrence of em dashes, for a necessary characteristic of good writing. Whatever we now chastise as literary affronts for their likeness to AI-generated writing were, before AI, a kind of writing that could be considered forgettable, or writing that aspired to be, but was not quite yet, good. As Laura Kipnis has written in Wired, large language models, through which the thing routed its edits, favor “pseudo-profundities” like delve and dive. The side effect of fraternizing with AI as much as I was expected to do was that I started to see the hand of the machine lurking in things it could never, ever have touched. In his literary manifesto A World of Sex, for example, Henry Miller uses the word delve twice, smartly. But when I saw it in his prose, I was jump-scared.
At work, I was instructed to do away with all the telling platitudinal verbs, slipping an innocent choice in their place, sometimes even replacing the em dashes with commas. Even when endorsing the thing and being up-front about using AI, it couldn’t seem like you’d used AI.
I started to become disillusioned just as the thing started gaining popularity. We two sole employees and an intern were assigned to line-edit an entire 150,000-word PhD dissertation in no more than three days. As with many of our edits by then, it was to be completed piecemeal, as if the pages of the manuscript were sliding across a conveyor belt, each of us assigned to specific tasks.
When I told my boss the goal was impossible, she noted that this was probably because we expected to take as much time on it as a “manual editor” would. (It would appear that we were the “assisted” editors.) There was not even enough time to push back on the request: With the thing, we were compelled to work faster, even if it was making many dubious edits, as the mention of the thing slowing us down would always be regarded as a defensive reflex on our part.
In his 1948 lecture “Good Readers and Good Writers,” Vladimir Nabokov states that a reader who begins a book with a “ready-made generalization” in mind is reading it from the “wrong end…. Nothing is more boring or more unfair to the author.” Yet the prompt library of the thing was lined with phrases that worked with such generalizations. Each draft, it was assumed, warranted a homogeneous approach; the only demarcation was its form. (We even had a prompt to edit poetry.) As I worked, my ambivalence toward everything I edited swelled. I’d habitually run anything I was assigned to do through the models. (I was often asked to gut-check “with” them.) If I wasn’t swift enough, I was reminded that I should’ve been, for any holdup was objectionable when these tools exist. Any time I edited withthe thing, which was most of the time, the metastasizing impact lingered long after the injury of using it ended.
A good reader, “a rereader,” Nabokov writes, can take in the “artistic delight” of the writer. The thing, with its numb productivity, had neither the time nor the ability for this vaunted part of the exercise. Using it, I learned that anyone could engineer prompts. Even someone like me, with all my inexperience of the world, could hastily type something like “proptedit shrt story” into a chatbot interface and watch it reliably spew out a coherent prompt for editing a short story. I knew I wasn’t really editing, but at times I had to believe I was. At work, I often couldn’t allow myself to confront the thought that several prompts of roughly a few hundred words detailing how to approach any type of manuscript would never match years of editorial experience. That would’ve compromised the enterprise entirely.
