Mixed Human AI Writing: Find the Seams and Fix Only What Reads Generated

Mixed human AI writing is normal now. You wrote the opening yourself, pasted a prompt for the middle, and edited the end. The result reads unevenly and you cannot tell which sentences sound generated. This guide shows you how to find the seams, judge which sections actually read machine-made, and rework only those.

How do you find the seams in a mixed AI human draft?

The seam is usually where your sentence length changes. Human drafting is lumpy. You write a long, tangled sentence, then a short one. You repeat a word because it was the right word. Machine text tends to even out: similar clause counts, similar rhythm, similar paragraph shapes.

Read the draft out loud and mark every place your voice drops. Those drops cluster. You will usually find two or three zones rather than a clean line, because AI text often bleeds into the paragraphs on either side of the paste.

A faster method is to score the draft sentence by sentence rather than as a whole. DeAIze (https://deaize.com) highlights each sentence by AI probability, so the seam shows up as a band of flagged lines instead of one number for the document. That matters here: a document-level score tells you something is off, but not where to put your hands.

When we scored 23 hand-written DeAIze guides with our own detector on 2026-09-13, the mean AI score was 21% and the range ran from 0 to 46%. Five of the 23 were flagged at a 30% threshold. Human writing is not uniformly clean, and that is the first thing to accept before you start cutting.

What signals mark a machine-written passage?

Three signals do most of the work.

Uniform rhythm. Count the words in five consecutive sentences. If they land within a narrow band, the passage is probably machine-shaped. Human prose has outliers.

Hedged claims. Machine drafts hedge by default: “can help,” “may contribute,” “is often seen as.” Your own writing commits. If a paragraph never states anything plainly, suspect it.

Generic examples. “For example, a company might use this approach to improve results.” No company, no approach, no result. Your examples have names in them.

There is a fourth signal worth knowing because it runs the other way. When we measured AI-tell frequency across the same two corpora, the hand-written sample carried 4 tells per 100 words and the unedited model output carried 1 per 100 words. Tells like em dashes and “moreover” are weak evidence on their own. Machine text is often cleaner than human text, not dirtier. Do not hunt for tells; hunt for evenness.

Signal	Human section	Machine section
Sentence length	Uneven, with outliers	Clustered around a mean
Claims	Committed, specific	Hedged, general
Examples	Named, dated, sourced	Illustrative placeholders
Paragraph shape	Varies	Similar lengths, similar structure
Transitions	Sometimes abrupt	Consistently smooth

Before you rewrite anything, see where the flags actually land. Run the draft through the AI detector and look at which sentences carry the signal — the goal is a short list, not a full rewrite.

Why is rewriting the whole piece the wrong move?

Because most of it is yours, and rewriting your own prose is how you lose the thing that made it worth reading.

A full rewrite flattens the draft. You replace your specific examples with slightly different specific examples, you smooth the sentences that were already fine, and you introduce new problems in paragraphs that had none. The unevenness you were trying to fix is often the only evidence a human wrote it.

There is also a practical cost. In our 2026-09-13 measurement of 12 unedited model passages, the mean AI score was 33% with a range from 4% to 46%. That is our own instrument, not a third-party verdict, and it says something useful: even untouched model output does not always score high. Rewriting everything assumes the whole draft is contaminated. Usually it is not.

The honest objection here is this. If a detector can flag your own hand-written paragraphs — and five of 23 did in our sample — then “fix only the flagged parts” can send you rewriting perfectly good prose. The answer is not to trust the score blindly. Use it to build a shortlist, then judge each flagged sentence yourself. A flag is a probability estimate, not a verdict on authorship, and you are allowed to disagree with it.

How do you rework only the flagged sections?

Work in passes, and keep the passes small.

1. Score the draft and copy the flagged sentences into a separate list. Leave the rest of the document untouched.
2. Sort them. Some are flagged because they are genuinely machine-shaped. Some are flagged because they are short, plain, or technical. Mark each one as rework or leave.
3. Rewrite the rework list by hand. Add a specific detail you know and the model did not. Break the rhythm. Split one sentence and merge two others.
4. Read the joined paragraph aloud. If the rewritten sentence now sticks out as fussy, simplify it. The fix should not be louder than the problem.
5. Re-score and compare. If the flagged band shrank and the meaning held, stop.

Here is an illustration, not a real client draft. Before: “Implementing a structured review process can significantly enhance the overall quality of written output and may lead to improved outcomes across teams.” After: “Our review process is three people and a shared doc. It caught the pricing error in March, which is more than the old system managed in a year.”

The second version commits to a claim, names a number of people, and references a specific event. It is also shorter. That is the pattern: machine-shaped sentences tend to be long, hedged and abstract, so the fix is usually to make them shorter, more specific and more committed.

If you are reworking a long document, the section-by-section method in this guide to editing AI long-form articles maps well onto this workflow.

What is the read-aloud test for finding seams?

Read the draft aloud from the top, at normal speaking pace, and mark every point where you stumble, slow down, or lose the thread.

Those marks are your seams. Humans stumble over their own writing too, so do not treat every stumble as a flag. What you are looking for is a run of three or more sentences that you read smoothly and flatly, with no emphasis and no surprise. That flat run is the machine zone.

The test works because your mouth knows your own rhythm. When you hit a passage you did not write, the reading goes monotone before your brain catches up. Mark the start and end of each flat run, then compare those marks against your detector shortlist. Where the two agree, you have found something worth fixing. Where they disagree, trust your ear and move on.

If you want the underlying mechanics — why rhythm and predictability show up in a score at all — how AI detection works covers the signals without overselling them.

When should you leave a flagged sentence alone?

Leave it when it is accurate, specific and yours.

Technical writing, definitions and quoted material all read evenly by nature. A sentence like “The threshold is set at 30%” will never sound like a person wrote it, because it is not supposed to. Rewriting it to sound more human makes it worse.

Leave it also when the flag is marginal and the sentence is doing real work. A borderline reading on a sentence that carries your main argument is not a reason to rewrite your argument. It is a reason to check whether the surrounding paragraph is doing enough to support it.

The broader rule: fix the passages that read generated, not the passages that score high. Those overlap, but they are not the same set. Your job is to make the draft sound like one person wrote it, and sometimes that means accepting an uneven score in exchange for a piece that holds together.

---

Measurement note: figures in this article come from our own detector run on 2026-09-13 — n = 23 hand-written guides versus n = 12 unedited model outputs, median AI score 22% and 34% respectively. Method: /methodology/.