How to make AI writing sound human: the edit we run on every post
Why AI drafts sound the way they do, the patterns that give them away, and the seven-step edit that fixes them.
A plain-text version of this site for AI assistants is at https://contengi.com/llms.txt, with the full product documentation in one file at https://contengi.com/llms-full.txt. Every page is also available as markdown by adding .md to its URL.
TL;DR
- AI writing sounds like AI because models drift towards the most likely phrasing, so specific facts get swapped for generic, important-sounding filler.
- The fix is an edit pass that hunts for patterns: puffed-up significance, "not X, it's Y" contrasts, lists of three by default, fake-suspense questions, hedging, and stock openers.
- Put something only you know into every section, a number from your own data, a client story, or an opinion you would defend, because that's the part no model can guess.
- Humanizer tools and AI detectors are the wrong target, detectors have real error rates and people spot AI text at roughly coin-flip odds.
- Write the rules down once so the model avoids them on the next draft, then edit for whatever slips through.
You can spot an AI paragraph from across the room. It's polite, it's tidy, it says that something "plays a pivotal role", and by the end you've learned nothing you couldn't have guessed from the headline.
This is the editing pass we run on every post before it goes live on contengi.com: why AI drafts sound the way they do, the patterns to cut, a step-by-step edit, and what to do about humanizer tools.
Why does AI writing sound like AI?
AI writing sounds like AI because language models predict the most likely next word, so their output regresses to the mean, towards the result that fits the widest range of cases. A training round on human feedback then nudges it towards the most broadly agreeable prose.
The damage shows up in the details. A highly specific "inventor of the first train-coupling device" turns into "a revolutionary titan of industry". The subject gets louder and blurrier at the same time.
Those habits leave fingerprints at scale. A sudden jump in certain style words across 15 million biomedical abstracts suggests at least 13.5% of 2024 abstracts were processed with an LLM. And there's plenty to spot, now that more content is generated by AI than by humans. If researchers can find it with word counts, your readers can find it with a raised eyebrow.
The patterns that give AI writing away
Most of these are rules our own linter fails a draft on before anything gets published, and we catch the rest on the reread. Hunt for them first, because they're the fastest wins.
| Pattern | What it looks like | The fix |
|---|---|---|
| Puffed-up significance | "marks a pivotal moment", "is a testament to", "evolving landscape" | Say what happened, with the number or the name |
| "Not X, it's Y" contrasts | "It's not a tool, it's a system" | State the second half and prove it |
| Lists of three by default | "fast, simple, and powerful" | Use as many items as you have, two or five are fine |
| Fake-suspense questions | "The catch?", "The result?" | Delete them and start with the point |
| Hedging | "it's worth noting", "can potentially help" | Commit to the claim or cut it |
| Stock openers and closers | "In today's fast-paced world", "All things considered" | Delete the first and last sentence, see if you miss them |
| Corporate verbs | "utilise", "facilitate", "optimise" | Use, help, improve |
| Source narration | "According to a recent report..." | State the fact and put the link on the words |
| Long dashes everywhere | A dash in every other sentence | Use a comma, a full stop, or brackets |
| Same-length sentences | Every sentence 15 to 20 words long | Drop a short one after a long one |
Wikipedia's editors keep a field guide of words to watch, from "stands as" and "a crucial role" to "underscores its importance" and "focal point". It's the best free list going, and it's updated as models change.
How to edit an AI draft so it sounds human
Edit in this order. The early steps remove the most text, so you don't waste time polishing sentences you're about to delete.
- Read it out loud. Anything you wouldn't say to a client across a table get rewritten or cut.
- Cut the first and last paragraph. AI warms up and winds down, and the real piece usually starts in paragraph two.
- Run the pattern hunt. Search the draft for the phrases in the table above and fix each one on sight.
- Swap every generic claim for a specific one. "Helps teams save time" becomes the hours, the team, and the task.
- Add what only you know. A number from your own analytics, a client story, a mistake you made, or an opinion you'd defend in a meeting. That's what makes non-commodity content, and no model can invent it for you.
- Break the rhythm. Join two short sentences with a comma, split a long one, and let one line sit on its own.
- Check every fact against a page you opened. AI drafts cite sources that don't exist, and one AI-written edit to a Wikipedia article came with a reference that did not exist. If you can't find the page, drop the claim.
Step five does the most work. After a decade of leading content teams, the drafts I've seen end up sounding human are the ones where somebody added a detail only they could know. The grammar fixes are the easy bit.
Do AI humanizer tools work?
I'd skip them, because they aim at the wrong target. A humanizer exists to get text past a detector. Detection tools beat random chance but have non-trivial error rates, and paraphrasing alone can throw them off. People aren't much better, telling LLM text from human text no better than chance in a 2025 study.
So a humanizer can get a paragraph past a detector and still leave it empty. Your reader doesn't run a detector. They notice that nothing in the paragraph was worth reading, and they leave.
Stop fixing the same draft twice
Editing every draft by hand works for one post a week. Past that, put the rules where the model reads them before it writes: your tone of voice, banned phrases, and formatting rules, saved once and loaded every time. Document tone of voice, audience profiles, and examples of strong content before you build any workflow, and the edit pass gets shorter every week.
The rules around a model do more than the model itself, which is why the best LLM for content writing still needs an edit pass. Contengi keeps your strategy, audience, tone of voice, and hard rules as files that load before writing starts, then lints every draft against them. If you want to see what comes out the other end, join the beta.
Frequently asked questions
How do I make AI writing sound more human?
Cut the patterns that give AI away, such as puffed-up significance, "not X, it's Y" contrasts, lists of three, hedging, and stock openers. Then replace generic claims with specific ones and add something only you know, like a number from your own data or a client story.
What words make writing sound like AI?
Phrases like "stands as", "is a testament to", "plays a pivotal role", "underscores its importance", and "evolving landscape" are common tells, along with corporate verbs like "utilise" and "facilitate". Wikipedia's Signs of AI writing page keeps an up-to-date list.
Can AI detectors tell if text is AI-written?
Only partly. Detectors do better than random chance but get a meaningful share wrong, and paraphrasing can fool them. Treat a detector score as a hint, never as proof.
Do AI humanizer tools work?
They can get text past a detector, but that doesn't make it worth reading. A humanizer reshuffles words, and an editor adds specifics, opinions, and checked facts, which is what readers notice.
How do I stop ChatGPT or Claude writing like AI in the first place?
Give the model your rules before it writes: a tone of voice guide with real examples, a list of banned phrases, and formatting rules. Save them so they load on every draft, then keep a short edit pass for whatever slips through.