Editing a generated draft is a different job from editing a human first draft, because the flaws are the opposite ones.
A human first draft is usually messy, uneven, and specific — the ideas are there and the prose is rough. A generated draft is the reverse: smooth prose, even rhythm, and nothing in it that only you could have written.
So the editing that works is subtractive and additive rather than corrective. You are not fixing bad sentences. You are removing competent ones that say nothing, and adding the material a model had no access to.
Four passes, in this order, each done separately.
Why separately
Because a single pass catches the easy problems and leaves the expensive ones.
Reading for everything at once means reading for prose, which is what the eye does by default — and generated prose is fine, so the pass finds little and feels complete. The pass that would have caught the fabricated statistic never happens.
Four passes on a 2,000-word piece takes about forty minutes. That is the real cost of AI-assisted content, and budgeting for it is the difference between a saving and a liability.
Pass 1 — Structure
Read only for order and coverage. Ignore every sentence-level problem.
What you are looking for:
- Does the piece answer the question in the title, early? Generated drafts characteristically warm up for three paragraphs before saying anything
- Is anything important missing? A model writes from what is commonly written, so the uncommon and correct point is the one absent
- Is anything repeated? The same idea in the introduction, a body section, and the conclusion, differently worded
- Does each section stand alone? A section that refers backwards — "as mentioned above" — cannot be retrieved and quoted on its own, which matters for both featured snippets and model citation
The most common structural fix: delete the introduction and start with what is now the second or third paragraph. Generated introductions restate the title and promise what follows. The answer is usually already written, three paragraphs down.
Pass 2 — Specificity
The highest-value pass, and the one that turns generic text into something worth publishing.
Read every paragraph and ask three questions:
- Compared to what?
- How would somebody check this?
- What happens if it is wrong?
A paragraph that answers none of the three is not doing work. Either make it specific or cut it.
What the replacement looks like:
| Generic | Specific |
|---|---|
| "This can significantly improve results" | "It fixes the case where the exit condition fires after the payment confirmation, which is why customers keep getting cart emails after buying" |
| "Best practices suggest updating regularly" | "Update anything with a published vulnerability immediately, ahead of any schedule — the window between disclosure and automated exploitation is short" |
| "There are several factors to consider" | Cut. Then list the factors. |
Where the specific version comes from is the constraint: your own testing, a customer's words, a measurement, a documented source, a mistake you made. A model cannot supply any of it, which is why this pass cannot be delegated.
A useful signal: if a sentence would be equally true on a competitor's site, it is not earning its place.
Pass 3 — Voice
Read for register, and know what you are listening for.
Four markers a frequent reader detects:
- Uniform sentence rhythm. Every sentence roughly the same weight. Real writing varies. Short sentences carry emphasis
- Three-item lists applied to everything, including subjects with two parts or five
- Hedged completeness. "It's important to note", "while X, it's also worth considering Y" — prose optimised for not being wrong rather than for being right
- Summary sentences that restate without adding. Usually at the end of a section, occasionally at the end of every section
Three phrases worth deleting on sight: "in today's fast-paced world", "it's important to note that", and any sentence beginning "whether you're a X or a Y".
The fix is deletion, not rewriting. Most generated drafts improve by roughly 30% deletion, and what goes is the connective and summary material rather than the substance. Rewriting a bland sentence usually produces a slightly better bland sentence.
One thing not to do: ask a model to "make this sound more human". It converges on a different generic register — more contractions, more rhetorical questions, more forced informality — which reads as generated by a model that was told to sound human. That is worse, because it is generic and trying.
Pass 4 — The cut
Read once more with the only instruction being to remove.
What comes out:
- The introduction, or most of it
- Transitions that announce — "Now let's look at", "With that in mind"
- Section summaries
- The conclusion, unless it says something the piece has not
- Qualifiers — "very", "quite", "somewhat", "generally speaking"
- Any sentence you could not defend if challenged
Set a target. Removing a fixed proportion forces decisions that "tighten it up" does not. Thirty percent is a workable target on a generated draft and it is almost always achievable without losing anything.
Then read it aloud. It catches what the eye skips — a clause that does not resolve, a rhythm that drags, a claim that sounds fine and feels wrong when you hear yourself say it. This is the one quality step nothing can do for you.
What this does not cover
Facts. Verification is a separate job with a separate method, and it does not belong inside an editing pass — the eye reading for prose does not stop at a plausible statistic. How to run the check.
Whether the piece should exist. Editing cannot rescue a piece with no distinctive claim. That decision belongs before the draft, not after. Where it fits in the workflow.
A note on detection tools
Do not edit to defeat a detector, and do not trust one applied to you.
AI detection is unreliable. OpenAI withdrew its own classifier in 2023 citing low accuracy, and detectors are known to misclassify writing by non-native English speakers — which makes a detector result a source of unfair accusation rather than evidence.
Editing to lower a detector score optimises for the wrong thing. The reader is not running a detector. They are noticing that the piece says nothing specific, which is a different problem with a different fix — and it is the one the four passes address.
Frequently asked questions
How do you edit AI-generated content properly?
In four separate passes: structure, specificity, voice, then a cut. Doing them together means reading for prose, which generated drafts pass — so the expensive problems survive. Four passes on 2,000 words takes about forty minutes.
Why does AI writing sound the same?
Four markers: uniform sentence rhythm with no short sentences, three-item lists applied regardless of the subject, hedging that optimises for not being wrong, and summary sentences that restate without adding. All four come from producing the average of what has been written.
Should I ask AI to make its writing sound more human?
No. It converges on a different generic register — more contractions, more rhetorical questions, forced informality — which reads as a model told to sound human. That is worse than the original, because it is generic and visibly trying.
How much of an AI draft should I delete?
Around 30% on most generated drafts, and what goes is connective and summary material rather than substance. Setting a proportion as a target forces decisions that a vague instruction to tighten does not.
Are AI content detectors accurate?
Not reliably. OpenAI withdrew its own classifier citing low accuracy, and detectors misclassify writing by non-native English speakers. Editing to lower a detector score also optimises for the wrong thing — readers notice that a piece says nothing specific, which detectors do not measure.
What is the most valuable edit to make to AI content?
Replacing general statements with specific ones. Ask of every paragraph: compared to what, how would somebody check this, and what happens if it is wrong. A paragraph answering none of the three is not doing work.
The short version
- Read for structure onlyRead for structure only u2014 order, coverage, repetition, and whether each section stands alone.
- Delete the introductionDelete the introduction if the real opening is two or three paragraphs down.
- Read for specificityRead for specificity , asking of each paragraph: compared to what, how would this be checked, what if it is wrong.
- Replace general statementsReplace general statements with something from your own testing, a customer's words, a measurement or a documented source.
- Cut any sentence that would be equally true on a competitor's siteCut any sentence that would be equally true on a competitor's site
- Read for voiceRead for voice , listening for uniform rhythm, three-item lists, hedging and restating summaries.
- Delete rather than rewriteDelete rather than rewrite the bland sentences.
- Run a final cut with a 30% targetRun a final cut with a 30% target
- Read the whole piece aloudRead the whole piece aloud
- Verify the facts separatelyVerify the facts separately , as its own job.
Free: The 60-Minute Email Authentication Fix
A no-fluff checklist to set up SPF, DKIM & DMARC correctly and pass Gmail & Yahoo's sender requirements.
Muhammad Basim has worked in digital marketing since 2013, focused on email deliverability and AI-assisted content production. He is the author of The Email Deliverability Playbook and The Email Copywriting Playbook.
Related Articles
Does Send Time Optimisation Work?
Send time optimisation models are trained on when recipients opened messages, and open timestamps now include machine activity that has nothing to do with when anyone was reading. Privacy protection pre-fetches message content on its own schedule, regardless of whether the recipient opened it. Security gateways fetch content during pre-delivery scanning, which happens before the […]
Prompting for Marketers: The Part That Actually Matters
Output quality is set mostly by the material you supply, not the phrasing you use. That is the finding people take longest to accept, because it is less interesting than the alternative. A carefully worded request with no context produces a competent generic answer. A plainly worded request with your actual customer language, your positioning […]
What You Grant When You Connect a Tool
Your security perimeter includes every vendor you have ever connected, including the ones you stopped using and never disconnected. A connection is a standing grant. It does not expire because you stopped logging in, it does not lapse because the trial ended, and it does not disappear when you delete the app from your phone. […]