How to humanize AI text: edit the claim, not the vocabulary.

A generated draft is assembled from an average of everything the model has read. That averaging is why it sounds plausible and why it says so little. Every sentence is the most expected sentence given the one before it.

Humanizing, properly done, replaces expectation with particularity. A named tool, a real number, a constraint that came from experience, an opinion that could be wrong. The prose stops sounding machine-made because it starts containing things a machine could not have known.

That is why the process is editorial rather than mechanical. No rewrite tool can add the fact you left out, and no synonym swap can remove the absence of a point.

First, unsupported generality. The draft says a solution is robust and scalable. It never says what breaks under load, or which workload it survived.

Second, structural noise. Announcing connectives, a summary paragraph that repeats the introduction, headings that promise a section the section does not deliver. This is scaffolding left in the finished building.

Third, flat rhythm. Sentences of similar length and weight, paragraph after paragraph, with no short line to land a point and no long one to carry an argument. Readers experience this as boredom long before they could name the cause.

Fix those three and the question of whether a model helped draft the piece mostly stops mattering to the reader, which is the only test that affects your results.

Read a paragraph and underline the clause that asserts something. If nothing is underlined, the paragraph is decoration and should be deleted whole. This one move removes more AI-sounding text than any word list.

If something is underlined, ask what would have to be true for it to be correct, then write that. A sentence that says the migration cut p95 latency from 900ms to 240ms is not a sentence a model produces without being told, and it is not a sentence a reader skips.

Where you cannot supply the fact, cut the claim instead of softening it. A hedge is not honesty, it is a claim trying to avoid being tested.

These six changes handle most drafts, in this order.

Replacing delve with dig and tapestry with mix is cosmetic. The sentence keeps its shape, which means it keeps its emptiness, and the reader is no better off. Worse, the search-and-replace produces odd collocations that read as a second-rate thesaurus.

The tell is not the word itself, it is the ratio of abstraction to information. Crucial is fine in a sentence that explains what depends on the thing. Crucial is noise in a sentence that says the thing is crucial and moves on.

Keep a short banned list for your own house style if it helps the team edit faster, but treat it as a prompt to look at the sentence, not a substitute for looking at it.

Generated prose has low sentence-length variance. Every sentence is a medium one, which is why a page of it feels like a wall even when every sentence is grammatical.

Count the words in five consecutive sentences. If they all land between eighteen and twenty-six, the paragraph is flat regardless of who wrote it. Break one to six words. Let another run to forty if it is carrying a real argument.

Punctuation does some of this work too. A colon before a short list, a full stop where a comma would be grammatical, a paragraph of one sentence after three long ones. None of it is decoration, it is pacing.

The temptation when humanizing is to reach for colour: a war story, a client anecdote, a statistic that sounds about right. That is how a fact error enters a page that was previously merely dull, and a wrong number is a much more expensive problem than a bland paragraph.

Judgement is safer than anecdote. What you would recommend, what you would avoid, which trade-off you would take, what surprised you. Those are yours, they cannot be wrong in the way a number can, and they are exactly what a model cannot supply.

Mark anything you cannot source as a claim to check later, and check it before publishing. Humanized text reads as more authoritative, which raises the cost of any error it contains.

The ear catches what the eye forgives: three sentences starting with the same word, a clause that runs out of air, a transition that exists only to move you along.

Read the first paragraph and the last paragraph aloud. If they say the same thing, delete the last one. Then read the headings in sequence. If the headings alone do not tell the argument, the structure is decorative.

This takes ninety seconds and finds more than a full pass of automated checks, because the checks can only measure prose you have already written.

Measure before and after, and compare flags per thousand words rather than the raw count. If the density has not moved, the edit rearranged the prose rather than changing it.

Then read the sentences you changed. A lower score with a sentence that no longer parses is a loss, and no measurement will tell you that. Numbers narrow the reading, they do not replace it.

Keep the pair of reports with the draft if the piece will be reviewed by someone else. Showing a before-and-after is a cleaner account of your work than asserting that you edited it.

If more than half the paragraphs need deleting, stop editing and write from your own notes. Salvaging a structure you do not believe in takes longer than building the right one, and the result keeps the original's shape.

The signal is simple: if you cannot state the page's argument in one sentence without using its own headings, you are editing something you have not decided yet. Decide first. The prose follows quickly once the claim exists.

That is also the honest answer to the awkward version of this question. Humanizing is not how you make generated text pass as human. It is how you make it good, and sometimes the fastest route to good is a blank page and your own material.