How to humanize AI text: a 5-step editing method

What “humanize” actually means

The Killers’ Brandon Flowers built a whole hit single around a question about whether people are truly free or simply conform. I’d be depressed to think we humans have no free will — and that even when we do, we always conform. Conformity is exactly what gives AI writing away. It conforms to a certain vocabulary, a specific structure and cadence. Human writing doesn’t. Every writer has an individual style, just as we each have a unique voice. Some of it is natural; some was laboriously developed through school and professional practice. And some of it is riddled with typos and grammatical errors. Spell check in early word processors helped writers ditch the errors, and bespoke tools over the last two decades helped us fix the rest. Now we have Large Language Models that generate typo-free, grammatically correct writing in seconds. The only problem: it’s sterile. It smells AI-generated. And it is making us suspicious of everything we read online — understandably, when most new web pages now contain AI text.

Which model produced the draft barely matters. ChatGPT, Claude, and Gemini share most of the same tells because they were trained on overlapping text and tuned toward the same conventions. There are subtle differences between models though. ChatGPT leans hardest on chatbot sign-offs and emoji, Claude on the hedged both-sides paragraph — but the method below applies to all of them.

Humanizing writing is not about disguising machine output. If your goal is beating a detector, this guide won’t optimize for that. Humanizing writing is about putting the specific judgment of the writer — usually a domain expert — back into it, along with their style and signature.

Step 1 — Cut what makes no claim (the fastest way to make AI text less AI)

Read each sentence and ask what it asserts. AI drafts are full of sentences that make no claim — they’re pointless. For example: This approach offers significant benefits for organizations of all sizes. Delete them. There is nothing to salvage, so resist the temptation to rewrite. Expect to lose 20–30% of the draft. That’s a good thing: edited-down drafts make your point sharper and your argument more convincing.

Do the same with weak claims. Amazon (the technology company) has a culture of writing copious memos, and instituted writing principles to keep them efficient. One is “no weasel words” — e.g. “nearly all customers,” “will significantly increase revenue.” Weasel words make weak claims. Either cut those sentences or replace the vagueness with numbers.

Step 2 — Put back what only you know

A perfect segue to the next step: add what shows your expertise. A model can’t know the number your team hit last quarter, the objection a customer raised, or your lowlights. AI slop is generic because it’s built from the general case. The fix is the particulars — names, numbers, a specific challenge or customer objection. You can’t automate this, and you can’t outsource the critical thinking that interprets the specifics.

What’s the difference between a LinkedIn post that makes you cringe and one that makes you think? The first has every telltale sign of AI content. The second has a unique insight, ideally backed by data or personal experience.

Step 3 — Fix the rhythm (why AI sentences all run the same length)

Vary the length of your sentences. AI produces monotonous, similar-sized sentences (unless it’s the staccato of AI slop on LinkedIn). It reads uniform and boring. This is easy to fix: find your longest sentence and break it; find three consecutive sentences of similar length and merge two by rewriting them or joining them with a conjunction or punctuation. But beware the trap — overcorrect and you create the telltale staccato cadence.

The master writer Gary Provost demonstrated this with a passage that begins with five-word sentences and builds to a crescendo. Read it — it’s worth the click.

Step 4 — Strip the tells (the words ChatGPT overuses)

The last major editorial step is cleaning up the vocabulary. There’s a reason it comes last: there is no point polishing sentences you should have deleted in Step 1. Fix the structure first, then the words. This is a scanning pass, looking for four things.

First, the overrepresented words — intricate, pivotal, robust, seamless, tapestry — the vocabulary that shows up in AI output far more often than in anything people write unassisted. I keep a running list in ChatGPT / Claude speak, grouped by what each word usually means, so you can swap them easily. Second, the buzz verbs: delve, leverage, unlock, elevate, empower — every one has a plain sibling like examine, use, open, improve, let. Third, the stock framings that open and close AI drafts: “in today’s fast-paced world,” “it’s not just X, it’s Y,” the summary paragraph that restates everything you just read. Fourth, chatbot residue — “Hope this helps!”, “Great question,” the offer to elaborate. It sounds silly, but these end up in final copy from a lazy copy-paste more often than anyone admits.

Do it with find-and-replace if you like; the words are literal enough that search works. Or let a tool do the pass in one shot — this step is the most mechanical of the five, so it’s the most automatable. But if you do only this step, as most “humanizer” tools do, you get writing that is de-flagged but still empty.

Step 5 — Read it aloud

Finally, read the draft aloud. Actually aloud. Muttering it in your head doesn’t work, because your inner voice smooths over the problem areas. Anywhere you stumble, run out of breath, or hear the same sound twice in a row is a spot your reader will feel too. Mark it, and fix the marks at the end.

If reading aloud isn’t practical, have your machine do it. Every phone and laptop has text-to-speech buried in accessibility settings. Hearing your draft in a flat synthetic voice is brutally revealing.

This is the cheapest test in the whole method, and the one professionals actually use. Editors read galleys aloud. Comedians table-read every script. Speechwriters pace the room. They do it because the ear catches what the eye forgives.

Worked example

Before.

Step 1 — Cut what makes no claim. Sentence one is about “businesses” in general, not this company — delete it. In sentence two, “Recognizing this growing demand for intelligent automation” is throat-clearing — delete it; “leading” is a weasel word — cut it; “officially announced its launch of” says launched twice — keep one; “advanced” modifies nothing — cut. Sentence three promises transformation with no verifiable fact — delete. The first pass:

Thirteen words out of sixty-eight. After the cut you can see the paragraph contained exactly one verifiable fact. The slop was hiding the substance.

Step 2 — Put back what only you know. Which is why this step exists. Only the writer knows what the agents do, what they connect to, and what a pilot showed. With illustrative specifics in place:

Step 3 — Fix the rhythm. The two full sentences run to similar lengths. Shorten the first, let the second stretch:

Step 4 — Strip the tells. Most died in Step 1 (“fast-paced,” “innovative,” “leverage,” “seamlessly” never made it in). Scan the survivor against the list: “solutions provider” is borderline, but it’s the company’s own category descriptor — the writer’s call. Otherwise clean.

Final after.

Sixty-eight words to thirty-five, one fact preserved, three specifics added, and one statistic no editor or tool can supply.

How long this takes, and when a tool helps

By hand, this method takes 10–15 minutes for a short post and maybe an hour for a long article. Most of that goes to Steps 1 and 2, where you decide what to cut and what only you can add. That’s as it should be — those two steps are judgment, and best not automated or outsourced. They’re the part that makes the writing yours.

Steps 3 and 4 are mechanical, and mechanical work is what tools are for. UnSlopIt runs the rhythm and vocabulary passes in seconds and hands back a redline showing every change and the reasoning behind it. Paste, review the redline, and spend the saved time on the two steps that matter.

Unslop your draft →

FAQ

Can I humanize AI text for free?

Yes, two ways. The five-step method above works if you want to do it by hand. If you want the mechanical steps done for you, UnSlopIt's instant editor is free for your first three drafts, with three more when you register with an email. No card, no login needed.

How do I handle a long document?

Section by section, not in one pass. UnSlopIt caps each pass at 1,000 words, and the limit is doing you a favor: editing at section level is where you catch the structural problems — the padding, the restated premises. Split at your headings, run each piece, and read the seams afterward so the sections still hand off cleanly.

Does this beat AI detectors?

Sometimes, but that's not what it's for, and I'd recommend not pursuing that goal. Detectors are unreliable in both directions — they flag human writing and pass machine writing often enough that no edit guarantees a result. Besides, what passes today may flag next quarter. This method optimizes for the reader instead: writing with real claims, real specifics, and a human rhythm tends to stop pattern-matching as AI as a side effect of being worth reading.

Will editing change what I meant?

Steps 1, 3, 4, and 5 can't change the meaning — they remove and rearrange your claims, so your facts and stance stay the same. UnSlopIt's editor is built under the same constraint, with a redline so you can verify every change. Step 2 is the exception by design: it changes what you said by adding the specifics and the evidence, and it's a step you do by hand.