Making AI-Assisted Copy Sound Human
By ABD Web Tools Editorial · 2026-02-09 · 7 min read
Using a language model to produce a first draft is now normal, and it is not a problem in itself. The problem is publishing the first draft. Machine-written prose has a consistent set of habits, and readers notice them long before they can name them. The good news is that the habits are mechanical, which means they can be edited out systematically.
The tells, listed plainly
Once you know them, you cannot unsee them:
- Sentences of near-identical length, giving the text a flat rhythm.
- Triads everywhere — every list has exactly three items.
- Hedged, unfalsifiable claims that no one would dispute.
- Section endings that restate the section opening.
- Vocabulary tics: delve, leverage, landscape, tapestry, testament, robust.
- No specifics: no numbers, no names, no dates, no anecdotes.
Pass one: cut the throat-clearing
Delete the introductory sentence of every paragraph that merely announces the paragraph. Typically 15–20% of a draft disappears in this pass and nothing of value is lost. If a paragraph collapses entirely, it had no content.
Advertisement
Pass two: add what only you know
This is the pass that matters most for both readers and search quality guidelines. Insert concrete detail a model could not have produced: the number you measured, the client situation, the thing you tried that failed, the tool version where the behaviour changed. Two or three genuine specifics per thousand words transforms a generic article into a credible one.
Pass three: vary the rhythm
Read the draft aloud. Where you run out of breath, split the sentence. Where three sentences in a row have the same shape, merge two and leave one short. A four-word sentence after a long one carries real emphasis, and models rarely produce it unprompted.
Pass four: verify every claim
Check statistics, quotations, product capabilities and links. Fabricated citations are the most damaging failure mode of AI-assisted writing, and the one readers are least forgiving about. If you cannot verify a claim, cut it or attribute it honestly as an estimate.
What good use of a model looks like
Outlining, rephrasing a clumsy paragraph, generating counterarguments to your own position, compressing a long section, and producing variations of a headline. These are all editing accelerants. Producing final prose that nobody reviews is not.
Disclosure and policy
Search engines do not penalise AI assistance as such; they penalise unhelpful, unoriginal content regardless of how it was produced. Advertising and content-monetisation policies take a similar view: the requirement is original value for the reader. A draft that passes the four passes above meets that bar. One that does not, would not have met it if a human had written it either.
Frequently asked questions
Will AI-assisted content rank?
Yes, if it is genuinely useful, accurate and original. Search guidance focuses on quality and usefulness rather than the tool used to produce it.
Are AI-detection scores reliable?
Not especially. They produce both false positives on edited human writing and false negatives on carefully prompted output. Judge the writing, not the score.
Should I disclose AI assistance?
Disclose when it helps the reader judge reliability — for example on medical, legal or financial content. Consistency matters more than the specific wording.
How much editing is enough?
When you can point to at least two things in the piece that only you could have written, you are close.
Advertisement
