Humanize AI Text by Hand: A Demo

Why does AI writing sound robotic?

A large language model predicts the most statistically likely next word, and the most likely word is the most average one. Repeat that thousands of times and you get prose with no surprises in it: inflated significance, vague sourcing, hedged claims, and a closing paragraph that restates the middle. Readers may not name the patterns, but they feel them. We keep a working catalog of 29 specific tells, drawn from the same editing guide that powers our how to edit AI out of your writing checklist, and we run every post on this site against it before publishing.

The experiment: one draft, two versions, real scores

The search phrase “humanize AI text” gets tens of thousands of queries a month, and most tools answering it are paraphrasers that swap words around. I wanted numbers instead of promises. So I wrote a short piece on using AI for email the way a lazy chatbot session would produce it, measured it with our scanner, rewrote it by hand using the catalog, and measured again. Same scanner, same rules, both versions published here in full so you can check my work.

The before text: 14 red flags in 203 words

Fourteen hits across eight distinct patterns, and the regexes are the forgiving judge. A human editor spots more: the tired opener about our fast-paced world, the empty flattery of addressing every possible reader at once, and a final paragraph that says nothing three different ways.

The after text: zero flags, same length, more facts

Zero pattern hits, 198 words. Notice what the rewrite added, not just what it cut: a concrete time cost (two hours a day), named tools (Gemini in Gmail, Copilot in Outlook, Claude for long threads), a usable example instruction, and two safety rules with specifics. Fixing robotic text is mostly a research task disguised as an editing task. The vague sentences were vague because the draft contained no information.

Version	Words	Severity-1 hits	Distinct patterns
Specimen A (robotic draft)	203	14	8
Specimen B (hand rewrite)	198	0	0

Which AI tells matter most?

If you only fix seven things, fix these: chatbot artifacts (any sentence addressed to you rather than your reader), filler connectors that add length but no meaning, unsourced claims attributed to unnamed experts (name the source or cut the claim), significance inflation (delete the adjective, keep the fact), verbs that dodge the word “is” (things can just be things), announcements of what the text is about to do, and the generic conclusion (if the last paragraph summarizes, delete it). Our AI humanizer explainer covers why automated tools struggle with exactly these, and our AI detectors guide covers the other side of the arms race.

Can a humanizer tool do this for you?

Not the important part. A paraphraser can vary sentence rhythm and swap vocabulary, which is why its output sometimes slips past detectors. It cannot add the missing facts, and the missing facts are the actual problem. Worth knowing: OpenAI built its own AI-text detector and shut it down for low accuracy, and Stanford research found detectors flag non-native English speakers as AI at striking rates, a finding echoed in the underlying paper. Chasing detector scores is a game with broken referees. If you are tempted anyway because of a school or work policy, the answer is the policy conversation, not a tool. The goal that actually pays off is text worth reading, and the web is already drowning in the other kind, which even has a name now: AI slop.

How do you fix AI text yourself?

• Delete the frame. Cut any sentence talking to you instead of your reader, plus the intro that promises and the outro that summarizes.
• Replace every vague claim with a fact or delete it. “Analysts suggest X saves time” becomes a number, a named source, or nothing.
• Un-inflate. “Immense power” becomes what the thing actually does, in plain verbs.
• Read it aloud once. Anything you would not say to a colleague gets rewritten in the words you would say.
• Add one thing only you know. A price you checked, a result you got, an example from your own inbox. This is the step no tool can do, and it is the one readers notice.

Prompting better up front reduces the cleanup. Our blog writing prompts guide shows how to feed the model facts before it drafts, which prevents half these patterns from appearing at all.

Common questions about humanizing AI text

Do AI humanizer tools beat detectors?
Sometimes, and unpredictably in both directions. Detectors also flag human writing, especially from non-native speakers. Neither side of that fight is reliable enough to build on.

Is drafting with AI cheating?
Depends entirely on the rules where you are submitting the work. For your own newsletter, no. For a class with a no-AI policy, yes, and no rewriting trick changes that.

What is the fastest single fix?
Delete the last paragraph. AI drafts almost always end by restating themselves, and the piece almost always improves the moment that paragraph is gone.

How long did the rewrite take?
About twenty minutes, most of it spent looking up the facts the robotic version was papering over. The editing itself was maybe five.

Sources and method notes

• Stanford HAI: AI detectors biased against non-native English writers
• Liang et al., GPT detectors are biased against non-native English writers (arXiv)
• Grokipedia: AI slop

Read next

• How to Edit AI Out of Your Writing – the full 29-pattern checklist this experiment used.
• What Is an AI Humanizer? – what the tools do and where they fall short.
• AI Detectors, Explained – why the referees are unreliable.
• AI Content Detection – the wider arms race.
• Best AI Prompts for Blog Writing – prevent the tells before they happen.
• The AI Glossary – every term on this page, in plain English.