How to Humanize AI Text Manually (Step-by-Step Editing Guide)

To humanize AI text manually, edit it in five short passes instead of fiddling sentence by sentence: add the specifics, opinions and numbers only you know; break the formulaic structure; cut the AI vocabulary and hedges; put your voice back with contractions and concrete verbs; then run a detector and fix what it highlights. On a 600-word draft this takes 20 to 30 minutes once you've done it a couple of times, and it works better than asking the model to "rewrite this to sound human," because the model will mostly rewrite it in its own defaults again.

Can't I just ask ChatGPT or Claude to humanize it?

Partly. A good prompt is a reasonable first pass, and it's worth doing before you touch the draft. Something like:

Rewrite this in plain English for a smart, busy reader. Use contractions. Vary sentence length, including some under eight words. No em dashes, no lists of three, no "not just X but Y," no "may" or "potentially" unless the uncertainty is real. Don't open or close with a summary. Keep every fact exactly as written.

Expect it to fix maybe a third of the problem. A language model produces the most probable continuation given its training and your instructions, and its idea of "human" is itself an average. Ban "delve" and it reaches for "explore"; ban the em dash and it puts a semicolon in the same spot; ban the rule of three and you get pairs everywhere. The rhythm stays even and the content stays generic, because the model has nothing new to add. It doesn't know your numbers, your clients, or your opinion.

Two data points show how sticky the defaults are. OpenAI's CEO announced in November 2025 that ChatGPT would finally obey a custom instruction not to use em dashes, which tells you how long it ignored it. And GPTZero's vocabulary tracker, built by comparing 3.3 million AI and human texts, finds stock phrases AI uses 10 to 200 times more often than people; in its October 2024 list, "play a significant role in shaping" appeared 182 times more often in AI text and "today's fast-paced world" 107 times more often. Those survive most "make it sound human" prompts.

So use the prompt, then do the passes below. For more first-pass options, see our guide to ChatGPT prompts that sound more human.

The five-pass method at a glance

Work through the whole draft once per pass; if you try to do everything in one read, you'll fix words and miss structure.

Pass 1: Read for meaning and add what only you know

This pass does most of the work, and no tool can do it for you. Read the draft once, fast, and at every paragraph ask: what do I actually know about this that isn't here? Add, wherever it's true:

• A number you have. "Response times improved" becomes "response times went from about two days to same-day."
• A specific case. A client, a class, a date, a place. "Many teams struggle with handoffs" becomes "our support team lost tickets every Friday afternoon at shift change."
• An opinion. AI drafts refuse to take sides. You can. "There are pros and cons" becomes "I'd do it again, but not with a team under five people."
• What went wrong. Models write success stories. People mention the part that didn't work.

If a paragraph stays generic after this pass, you probably don't know anything about that point. Cut it. A short piece with three real things in it reads more human than a long one with none.

This is also where academic honesty gets decided. If you can't add anything because the ideas and evidence came from the model, editing the wording won't change that (more below).

Pass 2: Structural pass

AI drafts share a skeleton: an opener that announces the topic, same-length body paragraphs that each make one tidy point, and an ending that restates everything and gestures at the future. Wikipedia's editors list "outline-like conclusions about challenges and future prospects" and rigid section structure among the main signs of AI writing.

Do these in order:

1. Delete the first paragraph, or its first sentence. If the piece starts with "In today's…", "In an era where…", or a definition of the topic, the real opening is usually the second sentence.
2. Delete the last paragraph if it summarises. End on your strongest specific point or the practical next step. A piece doesn't need a "conclusion" to feel finished.
3. Break the rule of three. Find every list of three adjectives, examples, or benefits. Make it one, two, or four. Keep three only when there really are three.
4. Vary paragraph and sentence length. Merge two short paragraphs; split a long one; let a one-sentence paragraph stand. Inside paragraphs, shorten one sentence to under eight words and let one run long.
5. Move one conclusion to the top. AI drafts build up to the point. People often lead with it and then explain.

Pass 3: Word-level pass

Now the mechanical part. Use find (Ctrl+F / Cmd+F) so you don't rely on your eye.

AI vocabulary. The clearest evidence that these words are a machine fingerprint comes from Kobak and colleagues, who tracked word frequencies in over 15 million PubMed abstracts and found an abrupt jump in specific style words after ChatGPT's release; their Science Advances paper estimates at least 13.5 percent of 2024 abstracts were processed with an LLM, and the preprint reported "delves" appearing about 28 times more often than expected. Wikipedia's list notes the vocabulary shifts with each model generation: "delve," "tapestry" and "testament" peaked in 2023 to mid-2024, while "emphasizing," "enhance," "highlighting" and "showcasing" persisted into 2025. Check for both eras.

Hedges. "May," "can," "potentially," "it's important to note," "some argue." Keep a hedge only where you're genuinely unsure, and then say why. Everywhere else, commit.

Transitional crutches. Sentence-opening "However," "Additionally," and "In other words." Most sentences connect fine without a signpost.

Em dashes. Replace all but one or two with commas, periods, or parentheses. They aren't wrong, but they're now a tell.

Negative parallelism. "It's not just X, it's Y." "This isn't about A; it's about B." Say the Y part directly.

Participle tails. Sentences ending in ", highlighting the importance of…" or ", contributing to…" Cut from the comma onward. If the tail said something real, make it its own sentence.

Vague attributions. "Experts agree," "studies show." Name and link the source or delete the claim.

Our full list of signs of AI writing has examples of each.

Pass 4: Voice pass

Read it aloud, or have your phone read it to you. You're listening for the sound of a person.

• Contractions. "It is" to "it's," "do not" to "don't," wherever the register allows (emails, blog posts, most marketing; often not formal reports).
• First person where the genre allows it. "I'd skip this step if…" or "we tried this and…" Blogs and marketing almost always allow it; for academic work, check the assignment.
• Concrete verbs. "Is able to provide" to "gives." "Serves as a tool for" to "helps." Wikipedia's guide notes that AI avoids plain "is" and "has" in favour of "serves as," "boasts," and "features." Put the plain verbs back.
• One aside. A dry parenthetical, a small admission, a "this took me longer than it should have." One per piece, not one per paragraph.
• Your own habits. If you start sentences with "And" or use the occasional fragment, do it.

Pass 5: Check with a detector, then decide what to do with the score

Run the edited text through a checker. You can paste it into Rewritica's free AI checker to see roughly what a detector would see. Then:

• Low score, nothing highlighted. Done. Don't keep editing; over-edited text starts to sound odd.
• A few highlighted stretches. Nearly always generic sentences that survived Pass 1. Add a detail or cut them.
• Still high overall. The draft is still carrying the model's ideas and order. Go back to Pass 2, or retype the weakest paragraph from a blank line instead of editing in place.

Remember what the score is. Early detectors scored perplexity (how predictable each word is) and burstiness (how much that varies across the text); GPTZero says it moved to a deep-learning classifier in autumn 2023, and Turnitin trains a classifier on student writing versus model output. Neither is proof of anything, and both produce false positives. The passes above lower scores because they add the variety and specificity those classifiers associate with people, not because they trick anything. For the mechanics, see how AI detectors work; if you wrote something yourself and it still got flagged, read what to do when your essay is flagged as AI but you wrote it.

Before and after: one paragraph through all five passes

Say you run a small design studio, moved the team to a four-day week, and asked a chatbot for a paragraph about it for your blog. A typical draft:

In today's fast-paced business landscape, the four-day work week has emerged as a pivotal strategy for organizations seeking to enhance productivity, boost employee satisfaction, and foster a healthier work-life balance. By reducing the traditional five-day schedule, companies can unlock significant benefits — from lower burnout rates to improved talent retention. However, it's important to note that this transition isn't just about working less; it's about working smarter. Leaders must carefully navigate the challenges of implementation, including workload redistribution, client expectations, and cultural shifts. Ultimately, the four-day work week represents a testament to the evolving nature of work, underscoring the need for innovative approaches that prioritize both performance and well-being in an increasingly competitive environment.

Every tell is there: the stock opener, two lists of three, an em dash, "isn't just… it's," hedges, "navigate," "testament," a participle tail, a closing sentence that says nothing. And zero facts.

After the passes, written by the owner (an illustrative rewrite, not a real company):

We moved our nine-person studio to a four-day week in March 2025, and I expected the hard part to be fitting the work in. It wasn't. The hard part was Friday email. Two clients kept sending requests on Friday morning and getting annoyed at the Monday reply, so we wrote a one-line auto-responder and called each of them to explain. That fixed it in a week. Output didn't drop; billable hours per person went from about 26 a week to 25, which I'll take. What I'd warn other owners about: don't announce it as a perk. Announce it as a deal. Four days, same pay, same deadlines, and if deadlines slip we go back to five. Nobody has missed one yet.

What changed: team size, start date, the actual problem, the fix, two numbers, an opinion and a warning (Pass 1); opener and summary ending gone, sentences from three words to thirty, the point landing near the end as a payoff rather than a restatement (Pass 2); every swap-table word, hedge, em dash and "isn't just… it's" removed (Pass 3); first person, contractions, plain verbs and "which I'll take" as the one aside (Pass 4). There's nothing generic left for a detector to latch onto, and nothing a reader would skim.

Notice both versions are the same length. Humanizing is replacing the model's filler with your facts, not adding words.

How to humanize AI text for free

Nothing above costs money: a timer, find-and-replace, a read-aloud function (built into Word, your OS accessibility settings, and every phone), and a free detector check. If you do this often, keep a personal swap list and add to it when you catch the model using a new favourite.

Where a humanizer tool fits: it can do a decent version of Pass 3 and some of Pass 2 in seconds. Rewritica's humanizer, for example, strips the AI vocabulary, breaks up the rhythm and removes the structural tells, which saves you the mechanical 10 minutes. It can't do Pass 1. No tool knows your Friday-email problem existed. Spend the saved time on what only you know.

Edge cases

Essays and anything graded. If your course forbids AI drafting, humanizing an AI essay is still submitting AI work, and detectors are now tuned to spot edited output. If AI use is allowed with disclosure, edit it, cite the tool as required, and make sure the thesis, evidence and analysis are work you did and can defend in a conversation. The honest test: could you explain every paragraph to the professor without notes? If not, Pass 1 hasn't happened.

You're not a native English speaker. Formal, careful English with even sentence lengths and a limited vocabulary looks statistically like model output, which is why non-native writers get false positives. The passes still work, with one change: don't force contractions or idioms you wouldn't naturally use. Concentrate on Pass 1 and Pass 2. A plain short sentence in your own voice beats a borrowed idiom used slightly wrong.

Long documents. Do Pass 1 and 2 on the whole thing, then Pass 3 and 4 on the introduction, the first paragraph of each section, and the ending, which are what readers weigh most.

The model's facts. Humanizing doesn't verify anything. If the draft contains a statistic, a quote, or a citation, check it before you spend time rewriting the sentence around it. Models invent all three.

Frequently asked questions

How do I humanize AI text by myself without a tool?

Edit in passes rather than line by line: first add specifics only you know, then fix structure, then cut AI vocabulary and hedges, then restore contractions and first person. A 600-word draft takes about 25 minutes once you've done it twice.

Can I just ask ChatGPT or Claude to rewrite the text to sound human?

You can, and it helps as a first pass, but the model tends to swap one set of default phrasings for another and keeps the same even rhythm. Prompting gets you maybe a third of the way; the rest is your edits and your own details.

What words should I remove from AI-generated text?

Start with the high-frequency style words documented by Wikipedia and GPTZero: delve, crucial, pivotal, landscape, testament, underscore, showcase, foster, enhance, additionally, and stock openers like 'in today's fast-paced world.' Replace them with plainer words or cut the sentence.

Will manual humanizing stop AI detectors from flagging my text?

Usually it lowers the score a lot, because you're adding the variation and specificity detectors associate with human writing. It's not a guarantee, and detectors disagree with each other, so check with one, fix the flagged stretches, and keep your drafts and history as evidence.

Is it okay to humanize an AI-written essay for school?

Editing AI output so it reads better doesn't make the ideas yours. If your course bans AI drafting, humanizing the text is still a violation. If AI use is permitted with disclosure, edit it, disclose it, and make sure the argument and evidence are genuinely your own work.

Bottom line

Set a 25-minute timer and do the passes in order: meaning, structure, words, voice, check. Spend the most time on the first one, because adding what only you know is what makes text read as human; everything after is cleanup. Use a prompt or a humanizer tool for the mechanical part if it saves time, then put that time back into specifics. Run one detector check at the end, fix what it highlights, and stop. And if the draft is for a class, ask the honest question first: could you explain every paragraph without notes?