Two manuscripts landed on my desk in the same week not long ago, both nonfiction, both from first-time authors who had never worked with an editor before. One read like the writer had spent years arguing with themselves in the margins of other people’s books — opinionated, uneven in places, occasionally repeating itself the way people do when they’re circling an idea they haven’t fully pinned down. The other read like a well-organised encyclopedia entry: clean, correct, and interchangeable with a thousand other essays on the same subject.
Neither writer had used an AI tool that I could prove. Only one of them sounded like a person.
That distinction has become the subject of a lot of anxious conversation in publishing, marketing, and journalism this year, and most of it is aimed at the wrong target.
The backlash is real, and it’s aimed at the wrong thing
Brands have noticed that readers and listeners are tired of synthetic content, and they’re saying so loudly. iHeartMedia rolled out a “guaranteed human” pledge late last year, promising no AI-generated on-air personalities and no AI-generated music, after its own research found that 90% of listeners wanted their media made by people. Apple TV’s “Pluribus” closes its credits with the line “This show was made by humans.” Merriam-Webster named “slop” — the AI-generated filler now clogging feeds, slide decks, and even real-estate listings — its word of the year, as CNN’s business desk reported in a piece arguing 2026 will be remembered as the year “100% human” became a marketing claim in its own right.
The instinct behind all of this is correct: people can tell when something was made without a person actually caring about it. But “human-made” as a label solves an attribution problem, not a quality one. Paul Harrison, a marketing academic at Deakin University, made this point clearly in a recent piece for The Conversation: when the same image is randomly labelled “human created” or “AI created,” people consistently rate the one they believe was human-made as more beautiful and more meaningful — even though nothing about the object itself has changed. The judgment isn’t about what’s on the page. It’s about what the reader has been told about its origin.
That’s a useful thing to know if you’re running a marketing department. It’s close to useless if you’re actually trying to write something that doesn’t read like slop, because a label can’t do that work for you. A disclaimer that says a human wrote this sentence doesn’t make the sentence any less generic.
Every writer already has a fingerprint — the question is whether they’re using it
The tools for detecting authorship by style, not by claim, have existed far longer than generative AI has been a concern. In 1964, statisticians Frederick Mosteller and David Wallace settled a two-century-old dispute over who wrote twelve of the Federalist Papers — Alexander Hamilton or James Madison — by measuring how often each man used small, unremarkable words like “upon,” “also,” and “by.” Neither author was consciously choosing those words for effect. The pattern was buried below the level either of them was paying attention to, and it was distinctive enough to attribute authorship with odds of 100 to 1 in Madison’s favour on the most contested paper. That method, now called stylometry, is still taught as a foundational technique for anyone trying to establish who actually wrote a disputed text.
What that tells editors is something most of us already suspected from reading manuscripts for a living: every writer has an idiolect, a set of habits in word choice, sentence rhythm, and structure so consistent it functions like a signature, whether or not the writer has ever noticed it. The writer who sounds unmistakably human on the page isn’t the one performing humanity for the reader’s benefit. They’re the one whose sentences still carry the fingerprint that was already there before any of this became a marketing problem.
What gets lost first is friction, not correctness
In manuscripts, the giveaway is rarely a factual error or a grammatical slip — generic writing is usually perfectly correct. What goes missing is friction: the slightly awkward turn of phrase that reveals someone thinking in real time, the aside that doesn’t quite fit the paragraph’s thesis but earns its place because it’s true, the sentence that runs a beat too long because the writer was still working out what they meant as they wrote it. Smoothed-over prose reads easily and says nothing in particular. It’s the same instinct that makes a hotel lobby feel like every other hotel lobby — nothing wrong with it, nothing to remember either.
Writers who’ve been at this a long time tend to develop an ear for their own friction and know which parts of it to protect. That’s a different skill from having strong opinions or a big personality; some of the most distinctive prose I’ve edited belongs to quiet, methodical writers whose sentences just move in a way no one else’s do. The distinctiveness isn’t loudness. It’s consistency of a particular, hard-to-fake kind — the kind stylometry measures and readers feel without being able to name.
One test I use with a manuscript that feels flat: read a paragraph aloud and ask whether it sounds like a specific person talking, or like the average of everyone who has ever written about that subject. Writers rarely need this pointed out more than once. Most of them already know, on some level, which of their own sentences are doing real work and which ones were written to fill space between the ideas that mattered.
The writers who never lost their voice never had to fake it
The authors whose work still reads as human right now, in the middle of all this anxiety about authorship, aren’t the ones who’ve adopted a policy against AI tools on principle. They’re the ones who were already writing in a voice specific enough that smoothing it out would have been an obvious loss — to them and to their readers. That specificity was never a defence mechanism against a technology that didn’t exist yet. It was just what happened when someone kept writing like themselves for long enough that the habit became load-bearing.
Editors spend a lot of time doing the opposite of what generic prose does: restoring the friction a nervous first-time writer has sanded away because they assumed it looked unprofessional, or pointing out the one paragraph in a manuscript where the writer’s actual voice shows up and asking why the rest of the book doesn’t sound like that. It’s rarely a question of adding personality. It’s a question of noticing where a writer already had one and stopping them from editing it out.
The two manuscripts on my desk that week weren’t a referendum on AI. They were a reminder of something editors have always known: a distinctive voice was never a stylistic flourish added on top of good writing. It was the writing.
