Kill the AI Tells

Someone replied to your newsletter asking if ChatGPT wrote it

You did use it. You also spent forty minutes editing, reordered the whole middle section, and added the story about your dad’s workshop. It still got picked. And not because of the content, which was yours, but because of a handful of small things scattered across 800 words, none of them obvious on its own.

The way I think about it, tells work like an accent. No single vowel gives it away. The pattern does.

There are three layers to it, and most people only ever fix the first one.

Layer one is the words, and there are about twenty that matter

None of these are bad words in themselves. They just turn up far more often in generated text than in the way people actually talk, because models default to the most formal register available, the one that sounds most like Writing with a capital W.

Swapping them matters more than deleting them. Cut a word and you leave a hole. Replace it with the plain version and the sentence usually gets better.

Tell	What it’s doing	Write instead
delve into	Formal synonym for “look at” that nobody says aloud	look at, go through, read
tapestry	Decoration carrying no information	Cut it. It means “several things exist”
a testament to	Borrowed gravity	shows, proves, or name the evidence
landscape	A noun standing in for “situation”	the market, who’s paying for what, or cut
realm	Rarely heard outside fantasy novels	area, field, or name the thing
navigate	Sounds like action, describes none	handle, get through, or say what they do
foster	Institutional	build, encourage, start
myriad	A dressed-up “many”	lots of, or the real number
robust	Strong without saying how	say what it survives
seamless	A claim nobody can check	say what doesn’t break
unlock	Treats ordinary work as a locked door	get, find, start
elevate	Upward motion with no destination	improve, or name what gets better
leverage (verb)	“Use”, in a suit	use
game-changer	Universal praise, zero content	say what changed
in today’s world	Throat-clearing, always deletable	Delete. Start at the next sentence
it’s not just X, it’s Y	Fake profundity	Say Y

A second tier gets the same treatment: crucial, vital, ensure, utilise, furthermore, moreover, comprehensive, holistic, underscore, pivotal. Each one has a shorter, plainer twin. Use the twin.

Layer two is sentence shape, and it survives a word swap

Here’s the frustrating bit. You can replace every flagged word and the draft will still read as generated, because the shapes underneath haven’t changed.

Everything arrives in threes. “Clear, concise and compelling.” Three examples per point, three items per list, every time. The fix is to use two. Or four, with one that doesn’t quite match the others.

Every sentence is roughly the same length. Real writing varies a lot. You might write a long sentence that works through an idea across a few clauses, and then a short one. Read the draft aloud and notice where the rhythm goes flat.

“Not only … but also.” Very few people say this in a kitchen. Split it into two sentences, or drop “not only” and keep the rest.

Every paragraph is three sentences long. Let one paragraph be a single sentence, and let another run to six.

Hedges on every claim. “Can often help to”, “may potentially”, “is generally considered”. Either delete the hedge or commit to it. “This usually works” is better than “this may potentially prove beneficial in certain contexts”.

A question as the opener. “Ever wondered why your prompts stop working?” State the observation instead: “Your prompt worked in November. It stopped in January.”

Em-dashes doing all the joining. Keep it to one per paragraph at most. The rest can become full stops, commas or brackets.

Layer three is structure, and it’s the layer readers feel

The throat-clearing intro. Two paragraphs restating the title before anything actually happens. Find the first real name, number or physical object in the piece, and delete everything above it.

The summary that restates. “In this article, we explored three key strategies.” Cut it. The reader was there.

Bold on everything. Six bolded phrases per screen means none of them are emphasis anymore. Bold the one thing you’d point at with your finger.

The conclusion that concludes nothing. “Ultimately, the key is finding the approach that works best for you.” That sentence fits every article ever published, which is exactly how you know it’s empty. End with an instruction, a number, or the thing that happens next.

Headings that are labels. “Understanding Prompt Consistency” is a filing category. “Your prompt worked once and it won’t again” is a claim. Write headings someone could disagree with.

Run this over the draft before you publish it

Use this on a finished draft, never a first one. Trying to edit and write at the same time tends to produce neither.

Read the draft below as an editor. Do not rewrite anything. Report only.

1. WORDS. Quote every sentence containing any of: delve, tapestry, testament,
   landscape, realm, navigate, foster, myriad, robust, seamless, unlock,
   elevate, leverage, game-changer, in today's world, crucial, ensure, utilise,
   furthermore, moreover, comprehensive, holistic, pivotal. Add any word that
   is the same species even if it is not on this list.

2. SHAPES. Report:
   - every list of exactly three items, quoted
   - the word count of every sentence, in order, as a list of numbers
   - every "not only ... but also"
   - the sentence count of every paragraph, in order
   - every hedge (may, might, can help, often, potentially, generally)
   - em-dash count per paragraph
   - any question used to open a section

3. STRUCTURE. Answer:
   - Where is the first proper noun, number or physical object? Quote
     everything before it.
   - Does the final paragraph contain any information not already stated?
     Quote it, or say no.
   - List every heading. Mark each one LABEL (names a topic) or CLAIM (says
     something arguable).
   - How many bolded phrases are there, and how many screens of text?

No compliments, no summary of the draft, no rewriting.

DRAFT:
[paste]

Then fix what it finds in this order: structure first, then shapes, then words. If you do it the other way round, you’ll carefully rewrite sentences you’re about to delete.

Detectors are guessing. Your reader isn’t.

Nobody runs your newsletter through a classifier. Detectors get it wrong in both directions, flagging careful human writing and passing whole generated articles, and chasing a score will make your writing worse, not better.

The real cost is quieter. Flat writing drains trust in small amounts. Your reader can’t quite name what’s wrong, so they file you a little lower and read the next one a little less closely. That’s why I’d fix the tells: because the writing gets better, not because a tool might score it.

When not to bother

Skip all of this for internal documents. Meeting notes, a handover, a process doc nobody reads twice. Flat is fine there, and fast is better.

Skip it for safety and technical documentation, too. A procedure should be boring, repetitive and predictable, and “sounds like a person” is the wrong goal.

Skip it on a first draft. This is a last pass, done once, near the end.

And skip it when the argument itself is the problem. Tells are the final five per cent. Beautifully polished prose making a weak point is still a weak point, just read more carefully.

What you’ve got now

The three layers, a plain replacement for every flagged word, the shape fixes, and a prompt that finds all of it in a draft in under a minute.

Most of layer one can be stopped before it’s even written. Put the banned list in your voice file, and the model reaches for those words far less often.

Run the prompt over the last thing you published, and see what it finds.