AI Writing Habits That Make Content Sound Generic (and how to fix them)

I have read somewhere past a thousand AI first drafts in the last year. Client blog posts, product pages, cold outreach, onboarding emails, a best-man speech once, at 1am, in a panic. They came out of ChatGPT, Claude, Gemini, and whatever my clients happened to paste into the doc before handing it over. After enough of them, you stop reading the sentences and start seeing the wiring underneath: the same small set of moves, repeated no matter the topic, the tool, or the person who typed the prompt.

The strange part is that the drafts are rarely wrong. They are grammatical. They are organized. They stay on topic and hit the word count. The problem is quieter than a mistake. They read like they were written by someone who has never once had a strong feeling about the subject, and who is being paid by the buzzword to hide that fact.

A generic draft is not badly written. It is written by no one in particular, about nothing in particular, for no one who was actually asked.

What follows is a field guide to those moves. Not theory. Each habit below is one I flag on real work, and each fix is an edit I have made this month, usually in the margin, usually in about ten seconds once you know what you are looking at. Learn to see the ten of them and most drafts stop sounding like a machine and start sounding like a person had an opinion.

The ten habits, with the fix

Machine draft on the left, the rewrite on the right. The words in red are the tell.

Reaching for the fancy word

SOUNDS LIKE leverage · robust · navigate · seamless · testament · realm · foster

The single loudest signal. A model reaches for the dressier synonym because dressier words filled the polished, corporate text it learned from. The result is prose in costume. If a plain word carries the meaning, the longer one is not sophistication. It is padding with a bigger vocabulary.

The “not just X, it’s Y” move

SOUNDS LIKE “it isn’t merely ___, it’s ___ that resonates on a deeper level”

This construction promises depth and delivers a shrug. It gestures at a profound distinction, then names something vague enough to mean anything. Nine times out of ten the second half is where the sentence quietly gives up. Cut the setup and say the actual thing.

The rule-of-three tic

SOUNDS LIKE “simple, powerful, and effective” · “clear, concise, and compelling”

Three items in a row have a pleasant rhythm, so the model uses three for everything: three adjectives, three benefits, three closing thoughts. Read a page of it and the rhythm turns into a metronome you cannot unhear. Vary your counts. Sometimes two lands harder, and sometimes one blunt word does the most work of all.

Throat-clearing before the point

SOUNDS LIKE “in this section we’ll explore” · “let’s dive in” · “it’s worth noting that”

The draft narrates its own structure instead of just having one. It tells you it is about to make a point, warms up, clears its throat, and only then arrives. Readers do not need the announcement. Delete the runway and start at the moment the idea actually begins.

The weather-report opener

SOUNDS LIKE “in today’s fast-paced, ever-changing digital world”

You know this one. The opening line so general it could sit on top of any article ever written, about any subject, in any year. It says nothing because it was built to fit everything. Open instead with something only your specific piece could say: a number, a scene, a claim someone might argue with.

Authority with no source

SOUNDS LIKE “studies show” · “experts agree” · “research suggests” · “it’s widely known”

The model reaches for the shape of evidence with no evidence attached, because that shape appears constantly in what it read. A real citation earns trust. A phantom one spends it, and readers learned to smell the difference a long time ago. Name the source or cut the claim.

Hedging on every side of every claim

SOUNDS LIKE “can often, in many cases, depending on the situation, generally speaking”

A model is tuned to be careful and agreeable, so it qualifies itself into vapor. Every statement gets a cushion until nothing has a spine. Readers do not want a survey of maybes. They want your call. Make the claim plainly, then name the one real situation where it breaks.

Dead verbs and buried nouns

SOUNDS LIKE “provide assistance” · “make an improvement to” · “utilize” · “facilitate”

The verb gets turned into a noun, then a weaker helper verb gets bolted on to carry it. “Provide assistance” is “help” wearing a suit. “Facilitate optimization” is two nouns standing where one strong verb belongs. Find the buried action and turn it back into a verb doing its job.

Punctuation played for drama

SOUNDS LIKE the long dash dropped into every third sentence for a theatrical pause

This is the tell most people have started to notice, and for good reason. The model loves a dramatic interruption: the long dash for a pause, the aside tucked inside commas, the trailing suspense, the exclamation it has not earned. Use one and it lands. Use nine and the reader goes numb to all of them. For the record, this whole document uses zero of the long dashes on purpose, and the restraint is the point. Full stops still exist. So do colons and semicolons.

The conclusion that repeats the article

SOUNDS LIKE “in conclusion, as we’ve seen, by following these tips you’ll be on your way”

The last paragraph loops back and lists what you just read, adding nothing. A summary is not an ending. It is a receipt. Leave the reader with a move to make, a consequence to sit with, or a sharper version of the point than you dared put up front.

The cheat sheet

Keep this next to the draft. Scan the middle column, and if you spot yourself, apply the right column.

Why does AI write like this?

Worth understanding, because it tells you which edits a better prompt can handle and which ones only you can.

None of these habits are bugs. They are what a model does well, pointed in the wrong direction. It was trained to be safe, agreeable, and fluent, and it learned from an internet whose center of gravity is competent, forgettable content. Ask it to write, and it hands back the average of everything it has seen. The average of everything is, almost by definition, generic.

Which points to the real fix, the one no prompt fully solves. You can strip out every buzzword and still be left with clean, correct, empty prose. The buzzwords are the symptom. The cure is you adding the thing the model structurally cannot: a real example, a real number, a real opinion, a real scar. A sharper prompt gets you a cleaner draft. It does not get you a point of view.

A 10-minute pass that removes most of the tells

Run this on anything a machine helped you write, in this order, before you publish.

1. Read it out loud. Anything you would not say to a colleague across a desk, cut it or swap it for what you would say.

2. Delete the first paragraph. Nine times out of ten the piece actually starts in paragraph two.

3. Search the draft for leverage, delve, robust, navigate, seamless, foster. Replace each one or defend it.

4. Find every studies show and experts agree. Name the source, or remove the sentence.

5. Break the rhythm on purpose. Put one very short sentence next to a long one so the page does not tick like a metronome.

6. Add one concrete detail only you could know: a number, a name, a date, a specific thing that went wrong.

7. Cut every it’s worth noting, it’s important to, and in order to. The sentence survives without them.

8. Reread the last line. If it only sums up, replace it with a move the reader can make tomorrow.