7 Word-Level Habits That Give Away an AI Draft
Before you publish anything a bot helped write, run it against these seven tells — the small verbal habits that flag a draft as machine-made even when the ideas are solid.
Your reader notices before they can explain why
Most people can't articulate why a paragraph feels off. They just feel it — a flatness, a sense that nobody actually wrote this to them. That reaction usually traces back to a handful of verbal tics that language models fall into by default, not because the content is wrong but because the phrasing follows a predictable groove.
This isn't about editing sentence structure or rewriting whole passages — that's a different job. This is a diagnostic pass: a list of specific things to scan for in a draft, the way a proofreader scans for double spaces. Once you know the seven habits, you can spot them in about ninety seconds per page.
I started keeping this list after a client sent back a batch of product descriptions with the note 'these read like a robot wrote them' — even though I'd personally typed every word after cleaning up a model's draft. I went back and found four of the seven habits below, still sitting there from the original output. That's how sneaky they are: you can edit around them for an hour and still leave the actual tells untouched.
The seven tells and what to do about each
| Tell | How it shows up | What to swap in |
|---|---|---|
| Starter repetition | Three or more sentences in a row open with the same word or shape — 'This means...', 'This allows...', 'This creates...' | Rewrite two of the three so the sentence opens with the subject, a question, or a fragment instead. |
| Filler transitions | 'Furthermore,' 'Moreover,' 'Additionally,' stacked at paragraph openings like connective tissue | Cut the word outright. If the logic still holds without it, you didn't need it. |
| Hedge stacking | 'It could potentially be argued that this may, in some cases, be somewhat beneficial' | Pick one hedge, or none. State the claim and let the reader push back if they disagree. |
| Missing contractions | 'It is,' 'do not,' 'cannot' used everywhere a real person would say 'it's,' 'don't,' 'can't' | Contract anything that isn't a formal legal or technical document. |
| Rule-of-three padding | Every claim gets exactly three examples, three adjectives, or three clauses, whether or not three is warranted | Use one strong example instead of three mediocre ones. |
| Summary sentences | A paragraph restates its own point in different words right before ending — 'In short, this shows that...' | Delete the last sentence. If the paragraph still makes its point, the summary was dead weight. |
| Uniform sentence length | Every sentence lands within a few words of the same length, like a metronome | Break the pattern deliberately. One short sentence after two long ones resets the reader's ear. |
Read the draft out loud, phone-recorder running, at normal speaking pace. Anywhere you stumble, slow down, or feel silly saying the words exactly as written is very likely one of the seven tells. Real speech doesn't hedge three times in one sentence or open five sentences with 'This.'
A five-minute scan before you publish
- 1Highlight every sentence opener
Go paragraph by paragraph and mark the first two or three words of each sentence. Any color that repeats back-to-back-to-back is a starter-repetition flag.
- 2Search for the usual suspects
Use find-and-replace (without replacing) to count 'furthermore,' 'moreover,' 'additionally,' 'in conclusion,' and 'it is important to note.' Anything above zero per page deserves a second look.
- 3Circle every hedge word
Mark 'may,' 'could,' 'potentially,' 'somewhat,' and 'arguably.' If two show up in one sentence, cut one.
- 4Run a contraction check
Search for 'is not,' 'are not,' 'do not,' 'cannot,' and 'will not.' Contract the ones that would sound stiff if you said them in conversation.
- 5Read the last sentence of each paragraph in isolation
If it just repeats the paragraph's opening claim in slightly different words, delete it and see whether anything is lost.
Tells that don't make the main seven but still matter
- Every list in the piece has exactly three items, even when two would do or five would be more honest
- Abstract nouns doing the work concrete ones should do — 'utilization' instead of 'use,' 'implementation' instead of 'setup'
- A definition sentence that explains a term the audience already understands, inserted out of caution rather than need
- Transitions that summarize instead of connect — 'Having discussed X, we now turn to Y' where a simple heading would do the job
Why this list works better as a checklist than a rulebook
None of these seven things are wrong in isolation. Real writers hedge sometimes. Real writers use 'furthermore' once in a while. The tell isn't the presence of the habit — it's the density. One hedge per five hundred words is normal caution. Three hedges in one sentence is a pattern, and patterns are what readers (and detection tools) actually pick up on.
That's also why this is a scanning exercise, not a rewriting one. You're not trying to fix the prose here — you're trying to find where the prose needs fixing. Mark the spots, then go do the actual sentence work separately, with fresh eyes, ideally on a different day so you're not just skimming past your own blind spots again.
If you'd rather skip the manual highlighting and searching, running the draft through AI Humanizer Lab's humanizer flags this exact kind of pattern automatically and suggests natural-sounding swaps for each one — a decent way to catch what a tired eye misses at 11pm before a deadline.
Make your writing sound human
Humanize AI-generated text in one click with AI Humanizer Lab.
Try for free