How to make AI writing sound human: an editor's checklist and prompts

AI writing sounds human when it carries your specific knowledge, opinions and examples, and when an editor has removed the patterns models fall back on. This guide explains why AI text reads as AI, gives you a pattern-by-pattern editing checklist, a voice guide you can paste into any assistant, and the parts of writing you should keep for yourself.

How do you make AI writing sound human?

You make AI writing sound human by putting yourself into it before and after the model writes. Before: give it your actual knowledge, opinions, numbers and stories, so it has something specific to say. After: edit out the patterns language models default to, add details only you know, and read the result aloud.

Most "make it sound human" advice focuses on the second half, swapping words and varying sentence length. That helps a little. The bigger problem with most AI drafts is that they say very little. A model given a topic and no material produces the average of everything written about that topic, and the average reads as generic no matter how you phrase it.

So the work splits into three jobs:

1. Supply substance. Notes, data, examples, a position you hold.
2. Constrain style. A voice guide, a specific reader, a length limit.
3. Edit like an editor. Cut patterns, add specifics, check facts, read aloud.

Why does AI writing sound like AI?

AI writing sounds like AI because models fall back on a recognizable house style: confident, balanced, polished, and vague. Readers learn the pattern quickly, and once they spot it they trust the text less.

The volunteers of WikiProject AI Cleanup maintain a long field guide to signs of AI writing based on the AI-generated drafts they have removed from Wikipedia. Their categories line up with what editors see in marketing copy, essays and emails:

The same guide includes two caveats worth keeping in mind. These signs are not proof: people write this way too. And the list changes as models change; the guide keeps a section of "historical indicators" for patterns that newer models have dropped.

There is a second, quieter reason AI text sounds artificial: it has no point of view. It hedges every claim, balances every argument, and never says "this is wrong" or "I would not do this." Human writing worth reading usually takes a position.

Start with substance, not style

The single most effective change is to stop asking AI to write from a topic and start asking it to write from your material.

Compare two prompts for the same blog post:

The second draft will still need editing, but it contains facts, a timeline, a problem, a solution and a doubt. None of that can come from the model.

A good way to collect material is to talk first. Record yourself explaining the topic to a colleague for five minutes, transcribe it, and give the transcript to the model with the instruction to keep your examples and opinions. Spoken explanations are full of the specific details that make writing feel human.

Below is a transcript of me explaining [topic]. Turn it into a [format] of about[N] words for [reader].Rules:- Keep every specific example, number and opinion from the transcript.- Do not add claims, statistics or examples that are not in the transcript.- Where my reasoning has a gap, insert [TODO: ...] instead of filling it.- Keep my plain wording where possible.

The [TODO] rule matters. Without it, the model fills gaps with plausible filler, which is exactly the generic text you are trying to avoid.

Write a voice guide the model can follow

A voice guide is a short document that tells the model how you write. It works better than "sound human" or "be conversational", which models interpret by switching to a different set of tics.

Keep it under a page. Include who you write for, what you sound like, banned words and structures, and two short samples of your own writing.

VOICE GUIDE: [Your name or brand]Reader: operations managers at mid-size companies. Busy, skeptical, not technical.How I write:- Plain US English. Short sentences. Contractions are fine.- I state an opinion and give the reason. I don't present "both sides" unless  both are genuinely reasonable.- I use concrete numbers and named examples. If I don't have one, I leave a gap.- One idea per paragraph, 1 to 3 sentences.Never use:- These words: intricate, multifaceted, meticulous, vibrant, bustling, nestled,  showcase, underscore, testament, commendable, plus your own pet peeves.- "It's not X, it's Y" or "not just X, but Y".- Lists of exactly three adjectives.- Sentences ending in ", highlighting/underscoring/showcasing ...".- Opening lines like "In today's world". Closing lines that summarize.- Em dashes. Exclamation marks. Emoji.Samples of my writing:[paste 150 to 300 words you wrote][paste another 150 to 300 words you wrote]

Save the guide as custom instructions or a project instruction in your assistant so every draft starts from it. The prompt engineering guide explains why examples (the samples) carry more weight than descriptions.

The editing checklist: four passes

Editing an AI draft goes faster and better in separate passes than all at once. Each pass looks for one kind of problem.

Pass 1: cut

Delete anything that does not carry information.

• The opening paragraph if it only restates the title.
• The closing paragraph if it only summarizes.
• Sentences that claim importance ("This is a key consideration").
• Hedge stacks ("may potentially help to some extent").
• Every adjective you cannot defend.

Expect to cut 20 to 30 percent of a typical draft. The text gets more human as it gets shorter.

Pass 2: add specifics

Go paragraph by paragraph and ask, "Could anyone have written this?" If yes, add something only you can add: a number, a name, a date, a mistake you made, a customer's exact words, a reason you disagree with common advice. This pass does most of the work.

Pass 3: fix the rhythm

• Replace "serves as" with "is" and "utilize" with "use".
• Break up runs of same-length sentences. Mix short and long.
• Turn passive sentences active where the actor matters.
• Remove bold from everything except the one term per section that needs it.
• Change title-case headings to sentence case.

Pass 4: read aloud and check facts

Read the whole piece aloud. You will hear the phrases you would never say. Then check every fact, number, quote and link against a source, because models invent them. Google's guidance for publishers says to manually fact-check and review all AI-generated content before publishing, and that applies well beyond search.

Prompts that help with editing

AI is a useful editor of its own output if you ask for a diagnosis instead of a rewrite. Rewrites tend to replace one set of tics with another.

Review the draft below. Do not rewrite it. Return a table with columns:quote | problem | suggested fixLook for: inflated significance, vague attributions, groups of three that arepadded, "not X but Y" constructions, sentences ending in an -ing clause that addsno information, claims without a number or example, and summary sentences.Limit to the 15 most important issues.

Here are two paragraphs I wrote, then a draft. List the 10 biggest differencesin style between my paragraphs and the draft (sentence length, word choice,how opinions are stated, use of examples). Then suggest edits to the draftthat close those gaps.

Apply the fixes yourself. Making the edits by hand keeps you in control of the voice and forces you to read every sentence.

Do AI humanizers and detectors work?

AI humanizer tools and AI detectors both work poorly, and you should not build your process around either.

Humanizers rewrite text to dodge detectors, mostly by swapping synonyms and varying sentence structure. They do not add substance, and the synonym swaps often make text stranger or change its meaning. A draft that sounds generic still sounds generic after humanizing, with odder word choices.

Detectors estimate whether text was machine-written, and their results are not reliable enough to accuse anyone. Researchers found that popular detectors consistently misclassified writing by non-native English speakers as AI-generated. Wikipedia's AI cleanup guide tells its own editors not to rely solely on detection tools, and notes that humans are bad at telling the difference too.

Detectors answer the wrong question. Ask instead whether a reader who knows the subject would find the text useful and believable. Writing that meets that bar usually reads as human because a human's knowledge is in it.

What should you never outsource to AI?

Some parts of writing lose their value when a model does them. Keep these yourself:

• Your opinion. AI can stress-test an argument. It cannot hold one for you.
• Personal experience. Stories, mistakes, what a customer said. Invented anecdotes are dishonest and readers often sense it.
• Facts you have not checked. Every number, quote and citation.
• High-stakes personal messages. Apologies, condolences, performance feedback, a resignation letter.
• Anything that claims to be your own unaided work where that matters: school assignments under an AI ban, signed statements, some client contracts.

For marketing teams, the AI for marketing guide covers quality control at scale, and the guide to AI for SEO covers what Google rewards. If you publish AI-assisted work, the responsible AI use guide has a disclosure policy you can adapt.

Your next draft: a 15-minute routine

Use this on the next piece you write with AI:

1. Spend five minutes writing or dictating rough notes: your point, your examples, your numbers.
2. Paste your voice guide (or write a short one) and the notes into the assistant with a word limit and a specific reader.
3. Ask for [TODO] markers instead of invented details.
4. Run the four editing passes: cut, add specifics, fix rhythm, read aloud and check facts.
5. Ask the model for a diagnosis table of remaining AI patterns and fix them by hand.
6. Save any new banned phrase you caught into your voice guide.

The same rule holds for any AI output, from blog posts to code: a person who knows the subject reviews it before it ships. That is how Orkio is set up for coding agents too, with agents working in the background and opening pull requests that you review before anything merges.

Frequently asked questions