Editing · 9 min
How to humanize AI text without chasing a detector score
Chasing a detector score is the wrong target. The right target is writing that a careful reader recognises as considered human work. Do that, and the detector score takes care of itself.
- 1
Step one — accept that detectors are not the judge
Detector tools report a probability, not a verdict. They are wrong both ways: false positives on careful human writing are well documented, and false negatives on thoughtfully edited AI text are equally common. Editing toward a specific score number leads to writing tricks that look human to a detector and obviously off to a human reader — random punctuation, deliberate typos, inserted parenthetical asides. None of these improve the writing. They make it worse in a way that protects against the wrong audit.
- 2
Step two — replace the framing opener
Almost every AI draft opens with a sentence that establishes context rather than making a point. 'In today's...,' 'When it comes to...,' 'It is widely recognised that...' Cut it. Start with the thing the reader actually came for — the claim, the question, the scene, the recommendation. The single most reliable signal that text reads as machine-generated is the throat-clearing opener, and the single most reliable improvement is to remove it.
- 3
Step three — vary sentence length deliberately
AI text clusters its sentence lengths in a narrow band — usually sixteen to twenty-four words. Human writing distributes more widely. Read your draft aloud. Anywhere you hear monotony, break a long sentence or merge two short ones. Aim for one very short sentence per paragraph and one longer developmental sentence per paragraph, with the rest in the middle. The variation should feel natural, not patterned.
- 4
Step four — replace generic nouns with specific ones
'Solutions, experiences, opportunities, capabilities, insights, strategies, considerations.' AI text leans on these because they are statistically safe; specific nouns require knowing the specific situation. Every paragraph should have at least one replacement of a generic noun with the concrete thing it stands for — not 'engagement,' the reply rate; not 'experiences,' the moment a user noticed; not 'considerations,' the actual trade-off. Specificity is the single largest source of human-feel.
- 5
Step five — cut hedging until you take a position
Read your draft and count the hedges per paragraph — 'can,' 'may,' 'often,' 'in some cases,' 'generally,' 'arguably.' Most paragraphs in unedited AI text carry three or four. Keep the hedges where genuine uncertainty exists. Cut the rest. The piece should make claims it can defend, with hedges reserved for the places hedging is honest, not as the default register of every sentence.
- 6
Step six — add one detail no model would invent
AI text reaches for plausible examples. Human text reaches for specific ones — the meeting where a particular thing was said, the screenshot you actually have, the colleague whose objection you remember. Add one such detail per major section. It does not need to be dramatic; it needs to be specific. Specific detail is the property AI cannot manufacture without retrieval, and it is the property that most clearly signals 'a person wrote this.'
- 7
Step seven — fix the ending
AI endings summarise. 'In conclusion, the above considerations suggest...' Strong endings either commit to a recommendation, name a question worth carrying away, or zoom out to a broader implication. Replace the summary. This is often the single edit that moves a piece from acceptable to memorable.
- 8
Step eight — read once for voice, not content
Final pass: read the draft only listening for voice. Does it sound like the person whose byline is on it? Where it does not, the fix is usually micro — a different word choice, a small turn of phrase, a quietly opinionated aside that a real person would write. The cumulative effect is what separates 'this reads as human' from 'this reads as carefully written by someone in particular.'
Common pitfalls
- •Deliberately inserting typos or grammar errors. Detectors are tuned for this; it backfires.
- •Random sentence-length shuffling. The goal is meaningful rhythm, not noise.
- •Adding parenthetical 'human' asides (lol, right?, etc.) that do not match the piece's register.
- •Synonym-swapping with a thesaurus. It produces stiff, unnatural prose worse than the original.
- •Editing for a detector score on a different draft than the one you will publish — the published version is what matters.
Frequently asked
What detector score should I aim for?
None specifically. Aim for writing that reads as considered human work. The score will follow more reliably than the other way round.
Will these changes always pass detectors?
No tool will reliably pass every detector — including text actually written by humans. The right standard is whether the writing is good, not whether it passes a probabilistic check.
Is humanising AI text ethical?
Editing AI drafts to read well is normal writing practice. Concealing AI use from contexts where the audience reasonably expects original human work is a separate question, addressed in the ethics article.
Run the workflow on a real draft
Open the rewrite engine and apply the steps to writing you actually have to ship today.
Start rewriting