"Humanize" has been captured by tools that add random errors and synonyms to AI text so that a detector scores it as human. That is not humanizing. It is laundering. Real humanizing is the editorial work of making a text honestly belong to the person whose name is on it: their knowledge, their position, their voice, their responsibility. Here is how a writer does it, and why it is the only version that lasts.

First, what the platforms actually want

Google's guidance on AI-generated content is explicit and has not changed in three years: using AI is not against its guidelines; using automation primarily to manipulate rankings is. It rewards content that shows experience, expertise, authoritativeness and trust, however it was produced. It recommends bylines where a reader would expect one, and AI disclosures where a reader might ask "how was this made?". It advises against listing AI as an author.

Read that carefully. Nothing there rewards fooling a detector. Everything there rewards a real person who knows the subject, says something specific, and stands behind it. "Humanizing" that adds noise to dodge detection does not produce any of those signals. It also produces worse text. The responsible method produces all of them, and it happens to be what good editing has always been.

The method

1. Decide who is speaking, and write it down

AI text has no speaker. It sounds like the average of everyone. The first job is to choose one: a named person, with a role, a relationship to the reader, and a reason to be writing this. "Priya, founder, writing to customers who have been with us two years, because we are changing prices." That sentence governs every edit that follows. If you cannot write it, the text has no author yet, and no amount of word-swapping will give it one.

2. Add what the model could not know

This is the largest single move and the one that separates humanizing from paraphrasing. The model did not know the real reason for the decision, the customer who complained last month, what was tried before, what the number actually was, what the author is worried about. Every paragraph should carry at least one thing only this author could have written. If a paragraph has nothing like that, ask whether it should exist.

3. Take a position

Models hedge by default: "it depends", "there are many factors", "both approaches have merits". A person with a name has a view. Find the places where the text sits on the fence and decide which side the author is actually on. Then say it plainly. Taking a position is also where responsibility enters: the author is now claiming something, and must be able to back it.

4. Remove the model patterns

These are not forbidden words; they are tells of text with no author. Cut or rewrite when you see them:

PatternWhat to do instead
Triads everywhere ("clear, concise and compelling")Pick the one word you mean.
Empty openers ("In today's fast-paced world")Start with the specific thing.
Summarising the paragraph you just wroteDelete the summary.
"It's important to note that", "It's worth mentioning"Just note it.
Balanced hedges on every claimDecide. Hedge only where you genuinely do not know.
Generic examples ("a company might")The actual company, the actual event.
Rhetorical questions as transitionsState the transition.
Adjective stacks and intensifiers ("truly", "incredibly", "seamlessly")Cut them. The noun does the work.
Perfectly parallel structure in every listLet items be different lengths if they are different sizes.

5. Verify every claim, then own it

Models invent statistics, quotes, dates and names with total confidence. Every specific claim gets a primary source or gets cut. This is also the moment to ask the author: are you willing to be quoted on this? If not, it goes. Humanized text is text a human is accountable for.

6. Match the register to the relationship

Formality, warmth, directness and humour are not settings; they follow from who is speaking to whom. A founder writing to long-standing customers is warmer and more direct than a compliance notice. Read the text as the named author, to the named reader, and fix every sentence that would feel wrong to say aloud in that room.

7. Read it aloud

Rhythm is where AI text most reliably fails: uniform sentence lengths, uniform paragraph shapes, a cadence that never varies. Reading aloud exposes it. Vary sentence length on purpose. Let one paragraph be a single line. Cut the sentence you stumble on.

8. Disclose, when a reader would ask

Following Google's own guidance: where a reader might reasonably wonder how the text was made, say so, briefly and honestly. "Drafted with AI, rewritten and fact-checked by Priya." This costs nothing with a reader who trusts you and protects you with one who does not. It is also what a job record on After AI Work produces automatically.

The test at the end is not "does a detector think a human wrote this?" It is "would the named author say every sentence of this, out loud, to the intended reader, and stand behind it?"

Why the tricks fail

Detector-evasion tools optimise for one thing: a lower score on today's detectors. They do it by degrading the text: odd synonyms, broken rhythm, inserted errors. The result reads worse to every human and carries no more knowledge, no more position, and no more accountability than the draft. Detectors change; readers do not. A text with a real author does not need to evade anything, because there is nothing to hide: the AI draft was a starting point and a person finished it.

A definition of done

Named author and intended reader written at the top. Every paragraph carries at least one thing only the author could know. Positions taken where the draft hedged. Model patterns removed (see table). Every specific claim sourced or cut; author has confirmed they will stand behind each. Register matches the relationship. Reads aloud without stumbles. Disclosure line added where a reader would expect it.

What to give the writer

The AI draft, three or four samples of how the author actually writes or speaks (emails are ideal), who the reader is, the real reason the piece exists, and the facts the model was missing. The writer's job is judgement; the more of the author they can see, the better the judgement.

How this was made: written by the After AI Work team from editing work on our own platform. Google's position from "Google Search's guidance about AI-generated content" and "Creating helpful, reliable, people-first content", both current Google Search Central guidance as of 2026. AI tools were used to draft the pattern table; every row was rewritten and checked by a person. The irony is noted.