Your AI Leaves Fingerprints

I keep a list of words I am not allowed to publish. Nothing scandalous is on it. It holds ordinary words like "elevate" and "landscape," and the one I catch most often in my own drafts is "quietly."

I write with AI in the loop most days, and AI reaches for that word to make a flat sentence feel important. "AI is quietly transforming coaching." Nothing in that sentence helps the reader do anything. The word is performing significance, and it kept slipping into my writing until I put it on the list.

The list exists because AI leaves fingerprints. Punctuation habits and sentence shapes that show up in machine drafts across every tool and every topic, whatever you asked for. Wikipedia's volunteer editors got so tired of cleaning them off the encyclopedia that they published a public field guide to the patterns, and it reads like a description of half the marketing copy on the internet right now.

Your readers are marinating in these patterns, and every one of us scrolling a feed is developing a radar for them. It is not a precise instrument. In a blind test by the marketing software company Bynder, readers slightly preferred an unlabeled AI article, yet once people suspect copy is machine-made, 52 percent become less engaged. The radar works on feel, and in a trust business that feeling is expensive. A client who starts wondering about your emails goes back and rereads everything you have ever sent them.

The stakes moved again in late July. LinkedIn added a "seems like AI slop" button, so readers can now flag posts that feel machine-made, and the platform will privately warn writers when readers flag their posts this way. In the same update it pulled its own AI post-enhancer and replaced it with a plain proofreader, one that checks spelling and grammar and leaves your wording alone. If you market on LinkedIn, machine-flavored posts are now something your readers can report.

Five fingerprints

These are the five I sweep for before anything ships. Fair warning, once you see them you cannot unsee them. When I first read Joseph Campbell's The Hero with a Thousand Faces, Disney films were never the same again. This list does that to your LinkedIn feed.

• The dash habit. God love the em dash. Emily Dickinson built poems out of it, writers have trusted it for two centuries, and it has ended up the scapegoat of this whole discussion. What convicts is the habit — the long dash dropped mid-sentence as a hinge — twice a paragraph — every page — from a writer who never touched it before 2023. This tell became so notorious that OpenAI's Sam Altman publicly celebrated, in November 2025, that ChatGPT would finally obey an instruction to stop. It took the company that built the model three years to talk it out of a punctuation habit.

• Mirrored halves. "This isn't about productivity. It's about presence." Two halves of a sentence reflecting each other a little too neatly, again and again. One earned contrast can carry a paragraph. When every third sentence pivots on the same hinge, the page reads like a machine arguing with someone who is not there.

• Fake-significance words. "Delve." "Elevate." "Unlock." "Tapestry." And mine, "quietly." These words claim a significance the sentence never earns. The pattern is measurable. In 2024, forms of "delve" turned up in biomedical journal abstracts at up to 28 times the expected rate.

• The rule of three. "Clear, confident, and compelling." "For coaches, consultants, and creators." Machine drafts group everything in threes. The rhythm feels polished on a first read, and it flattens real lists, which usually hold two true items or four.

• The blur. A model drafts by pulling toward the average of everything it has read, so your specifics dissolve. Twelve years running a clinic in Leeds becomes "extensive experience in the health sector." The client who told you she got her Sundays back becomes "transformative results." If a paragraph could sit on a competitor's website without a single edit, it has this fingerprint in it.

None of these convicts on its own. Real people write long dashes and triples, and always have, which is why accusing a colleague of using AI over one piece of punctuation is a bad idea. Detection software makes that mistake at industrial scale and routinely flags human writing, so resist becoming the punctuation police. What gives a page away is the cluster. When several of these show up together, the page stops matching the person. Someone who has heard you on a webinar, or sat with you on a discovery call, notices the difference without needing any of the names above.

Why scrubbing is a losing game

The obvious response is a blacklist. Ban the dash and "delve," then scan every draft before it goes out. That was my first version of the fix, and it kept failing.

The first problem is that the tells move. "Delve" was the most famous AI word of 2024, and within months of the startup investor Paul Graham calling it out, its frequency in academic writing started falling. Each retraining suppresses the notorious patterns, and new habits grow where the old ones were patched. A banned-words list is a snapshot of last year's model. Mine has grown every few months since I started keeping it, and I expect this article to date the same way, with some of the patterns above gone from the models by the time you read it.

The second problem runs deeper. I have thrown away entire drafts of my own that passed every word check I run. The vocabulary was clean and the piece still read machine-made, because the machine was in the structure. The mirrored halves were the engine of every section, and no find-and-replace reaches that deep. A page like that needs its writer back, which is what the two fixes below are for.

Read it aloud

The fix I trust most costs nothing and takes about ten minutes. Open the page and read every line out loud, at speaking pace, because your ear catches what your eye forgives. On the screen, "I help ambitious leaders unlock their fullest potential" slides right past. Read it to an empty office and you stop halfway through, because you have never said that sentence to a living person. You will feel a bit ridiculous reading your own website out loud to nobody. Do it anyway.

The standard is a table with a client on the other side of it. Every line either survives being said across that table or it gets rewritten as whatever you would say instead. Say the replacement out loud first, then type what you just said. The sentence that comes out of your mouth will be shorter and more specific than the one it replaces. While you are in there, hold the page to a budget of one long dash and one deliberate contrast. Anything past that is where the cluster starts.

Swap in their words

The second fix hands you the replacement language, and you already own it. Your clients have been describing their problem, and what changed after you worked together, for as long as you have been practicing. It is sitting in emails, session notes, testimonials, and the message someone sent you the day something clicked.

Build a quote bank. Pull fifteen or twenty of those sentences into one document, stripped of names and anything identifying, with permission where the words were private. When the read-aloud test catches a machine line, skip the thesaurus and pull a line from the bank. "Restore work-life balance" becomes "I got my Sundays back." "Overcome limiting beliefs" becomes "I stopped rereading every email five times before hitting send." The swap removes the fingerprint and restores the specific detail the machine blurred out, in words your next client is already using to describe their own life.

You can do the gathering by hand in an evening, or have AI do the digging. The prompt below runs in any capable AI chat, and it holds the machine to the one rule that matters here. It surfaces your clients' words and is not allowed to improve them.

I am building a quote bank: a collection of my clients' exact words,
which I will use to replace generic language in my own copy.

Below I have pasted raw material: testimonial excerpts, emails, and
notes from my work with clients. [Paste everything below this
prompt. More is better. Remove names and anything a client would
not want shared before you paste.]

Go through it and pull out the 15 to 20 sentences where a client
describes, in their own words, either the problem they came to me
with or what changed after we worked together.

Rules:
1. Exact words only. Do not polish, paraphrase, or improve a single
   sentence. Typos and odd phrasing stay, because that is how real
   people talk.
2. Strip out any remaining name or identifying detail.
3. Group the quotes under two headings: "The Problem, In Their
   Words" and "The Change, In Their Words."
4. If you cannot tell whether a sentence is the client's own words
   or my summary of them, leave it out and list it separately so I
   can check.

You are not writing anything. You are surfacing what has already
been said.

Head it off in the draft

The two fixes above repair a page that already exists. The cheaper move is to stop the fingerprints arriving at all, and the shape of that is something I teach as the human-AI sandwich. Your thinking opens the work, AI develops it, and your judgment closes it, which makes you the bread on both sides.

In practice that means three habits. Get your own rough version down first, even two minutes of talking into your phone, and hand the AI that as its raw material, because a model developing your words drifts far less than a model writing from nothing. Give it standing rules in its settings, the way you would brief a new assistant. Mine include keeping my wording where it exists and never using the long dash. And tell it to ask rather than invent. One line in a prompt, something like "if you need a fact you do not have, stop and ask me," cuts off the blur at its source, because blur is what a model filling a gap with averages looks like.

The longer game

The fingerprints are what a model produces when it has nothing specific to work with, and the deeper version of that first slice of bread is a written base the AI can draw on for every draft, the framework you teach and the words your clients use. I have written before about building that base and about what AI can show you once it can see your material. Context improves the first draft without changing who signs off.

None of this is a case against the tools. I write with AI most days and say so, and the images on this site keep their AI watermarks because I am not hiding anything. Whether you tell people about your process is your call. The point of the partnership is using the technology to amplify your ideas, so you become more prolific in your own voice rather than the computer's. Celebrate the human in the work, and protect it with the last pass. Mine is every piece read aloud before it ships.

One page this week

Pick one page on your website. The about page and the services page are good candidates, because they are what people read right before deciding to get in touch. However the page got written, with AI in the loop or long before AI existed, the test is the same. Read it out loud at a normal speaking pace. Mark every line you would never say across a table, along with any claim that has inflated past its evidence. Swap each one for what you would say instead, or better, for words a client said. Start a list of what you catch. Mine is the one from the top of this piece, and every new draft gets checked against it.

You are done when you can read the whole page aloud to a real person without flinching. Ten minutes on one page, and the words people read before they ever speak to you start telling the truth about who answers the phone.