Do not index
"How do I make this sound less like AI wrote it?"
Stop adding personality and start removing patterns. Sounding human is not something you layer onto a draft. It is what is left after you delete the constructions that no person reaching for a point would have chosen. Voice is a subtraction problem, and the reason most edits fail is that people try to solve it by addition.
Pew now has the receipts on which constructions to cut. Across an analysis of 490,000 English-language web pages, em dashes appeared twice as frequently in pages published since 2023, Oxford comma use rose 63%, and negative parallelism such as "it's not just X, it's Y" nearly tripled in frequency, according to Dataconomy's coverage of the study. The words "delve," "interplay," and "testament" more than doubled in text identified as AI-like. Around 35% of web pages published since ChatGPT's launch show signs of AI authorship.
What makes this useful is that the fingerprints are structural, not lexical. Everyone already knows to delete "delve." Nobody talks about negative parallelism, and negative parallelism is the one doing the damage, because it is a rhetorical shape rather than a word. "It's not just a tool, it's a philosophy" reads as insight and contains none. The construction promises a distinction and then delivers a synonym. Readers register that as flat about a second before they can articulate why, and by then they have already scrolled.
Why readers feel it before they can name it
The em dash finding gets the headlines, and it is the least interesting one. An em dash is a punctuation preference and plenty of good writers use them. What the doubling actually indicates is a rhythm. AI drafts default to a specific cadence where every sentence carries a subordinate clause, every paragraph resolves neatly, and no sentence is allowed to be short and blunt. The em dash is a symptom of a sentence that refuses to end. When you strip them out, you are usually forced to break the sentence into two, and the second one lands harder than anything the original was doing.
The same logic applies to the Oxford comma jump of 63%. Nobody dislikes a post because of a serial comma. But a text that is consistently, mechanically correct across thousands of small choices reads as unsupervised, because real people are inconsistent in ways that signal a person made decisions in real time. Perfect is a tell.
The Subtraction Pass
Here is the process I run on every draft, and I call it the Subtraction Pass. It happens in one sitting and takes about six minutes on a 900 word piece.
First, delete every em dash and repair the sentence without adding words. If the sentence cannot survive, it was two sentences pretending to be one. Second, search the draft for "not just" and "isn't just" and kill every instance. Do not rewrite them into something cleverer. Delete the first half and keep the claim. Third, remove any sentence that could open a post on any topic in your category. That is where "in today's landscape" and "consistency compounds" die. Fourth, cut the last paragraph. AI drafts almost always end with a summary that restates the piece, and a real ending is the sentence before that summary started.
Nothing in that list adds anything. That is the point. I have watched ghostwriters charging $5k to $15k per month spend an hour trying to inject a client's personality into a flat draft, when twenty minutes of deletion would have produced something the client actually recognized. Addition produces a costume. Subtraction produces a voice, because what remains is only the material that had a reason to exist.
This is for ghostwriters and content leads editing volume, and for founders who use AI to get a first draft down and then cannot work out why the result reads like everyone else. It is not for people who want a prompt that solves this at generation time. No system prompt fixes negative parallelism reliably, because the model is producing the shape that its training says reads as insightful, and asking it not to do the thing it thinks good writing is will get you compliance for about two paragraphs. Skip this entirely if you are publishing under a company brand where flat is acceptable and nobody is expected to have a position. The Subtraction Pass is only worth the time when the whole point of the content is that a specific person wrote it.
Which is also why this connects back to positioning rather than editing. A draft only survives subtraction if there was something underneath the patterns to begin with. When you strip the constructions out of a post built on a real claim, the claim gets sharper. When you strip them out of a post built on nothing, you are left with four sentences and an obvious problem, which is useful information delivered early. I have made the case that founders should position as practitioners rather than thought leaders, and the editorial version of that argument is this one. Practitioner writing survives deletion because it is made of specifics.
The trajectory here is worth taking seriously. If 35% of the web published since 2022 carries these fingerprints, the patterns are becoming the ambient default of written English, which means the cost of avoiding them rises every year while the value of avoiding them rises faster. Within a few years the people who write in the residual style, the one left after subtraction, will be a small enough group that the style itself functions as identification. That is not a craft argument. It is a distribution argument, and it favors whoever is willing to delete more than everyone else.
