Humanize AI Text

A three-pass workflow for taking AI-generated prose (ChatGPT, Claude, Gemini, GPT-4o, GPT-5.x) and reshaping it so it does not read like a language model wrote it. Grounded in the 2025–2026 detection literature, the canonical AIism inventories, and the stylometric mechanisms detectors actually measure.

This is not "rephrase this nicely." It is a deliberate rewrite that fixes the underlying perplexity / burstiness / function-word / dispersion fingerprint, not just the surface vocabulary. A draft with "delve" swapped for "explore" still fails detection. A draft that has been re-rhythmed, re-distributed, and grounded in specifics passes.

Inputs

• A block of AI-generated or AI-suspected text (calibrated for 200–5,000 words; chunk longer pieces by section)
• Optionally, the user's target context (LinkedIn post, academic essay, email, technical doc, marketing copy) — different contexts allow different remediation aggressiveness
• Optionally, an AI-detector target ("must pass GPTZero" / "must pass Turnitin")

Output

1. The humanized text. Just the text. No preamble like "Here's a humanized version of your draft."
2. An audit summary after the text (only if the user is iterating or asked for it): which tells were present, which were fixed, residual risks, content-truth flags.

The three-pass workflow

The order matters. Doing punctuation before vocabulary creates regressions; doing rhythm before structure changes what needs re-rhythming.

PASS 1 — Triage (read twice before writing)

Score the input across five tell-categories. One-line note per category — you are sizing the work.

1. Vocabulary load: count canonical-blacklist words (delve, leverage, robust, navigate, tapestry, pivotal, seamless, multifaceted, foster, underscore, comprehensive, etc.). 5+ in 500 words is heavy.
2. Sentence-rhythm uniformity: are all sentences 12–22 words? Are paragraphs all roughly the same length? Low burstiness = high AI-ness.
3. Negative-parallelism count: "not X, but Y" / "less about X and more about Y" / "it's not just X." 2+ in a short piece is a flag.
4. Punctuation pattern: em-dash density (any density above 1 per 200 words is high). Smart quotes. Emojis.
5. Content-truth risk: are there citations? Named case studies? Statistics? If yes, the content-truth audit in PASS 2 step 6 is mandatory.

If three or more categories rate "heavy," do the full surgery. If only one rates heavy, a targeted single-category fix is enough. Don't over-edit a mildly AI-flavored piece — that's how humanized text starts to feel manic.

PASS 2 — Surgical edits (in this order)

Step 1 — Vocabulary swaps. Run a mental find-and-replace through the input. The working blacklist with replacements:

Verbs to cut or swap.

Adjectives to cut or swap.

Nouns to cut or swap.

Transitions to cut.

Intros and outros — kill on sight.

• "In today's digital age," / "In the ever-evolving landscape of X," / "In an increasingly X world,"
• "At its core," / "At the heart of,"
• "Imagine a world where..." / "Picture this:"
• "Have you ever wondered..."
• "Hope this message finds you well,"
• "Certainly! Here's..." / "Great question!" / "I'd be happy to help..."
• "In conclusion," / "Ultimately," / "To wrap things up," / "As we have seen,"
• "Hope this helps!" / "Feel free to reach out!" / "For further reading, check out..."

The frequency reality (Originality.AI 10.78M-word dataset): the actually-frequent AI words are blander than the famous ones — "unique" (4,479 hits), "additionally" (3,638), "finally" (3,103), "conclusion" (1,841), "journey" (1,835), "certainly" (1,616). "delve" appears only 146 times. The famous tells are unusual; the boring ones do volume. Check for over-use of the boring ones too.

Step 2 — Punctuation hygiene.

• Replace every — with one of: - (space-hyphen-space), ,, . (new sentence), or (...). Pick by what the dash was doing. Keep at most 1 em dash per 500 words. Zero is fine.
• Replace smart quotes (""''…) with straight ASCII (" ' ...) unless the target context requires typography.
• Strip emojis in headings, bullets, corporate / academic contexts. Keep only if the input is casual social-media voice AND the user signaled that register.
• If you have 12 em dashes and 0 semicolons in 2,000 words, the next step will add 1–3 semicolons naturally.

Step 3 — Sentence rhythm (the burstiness fix). This is the highest-leverage mechanism-level change. AI burstiness clusters 0.2–0.4; human burstiness is 0.6–1.2 (coefficient of variation of sentence lengths).

• Break at least 2 long AI sentences into a short sentence + a fragment. Fragments are okay in human prose. AI doesn't write them.
• Combine at least 1 pair of medium sentences into one longer one using a real subordinator (because, which, even though).
• Introduce at least 1 single-word or two-word sentence per ~400 words. ("It doesn't." "Wrong." "Worth trying.")
• Vary paragraph length: one paragraph should be 1 sentence; one should be 6+; avoid every paragraph being 3–4.
• Read aloud (or simulate it). If the cadence is metronomic, you didn't fix it.

A 2024 journalism case study found that editing AI text purely for burstiness reduced detector accuracy by 40%. This step alone moves the needle.

Step 4 — Kill the rhetorical templates.

• Negative parallelism: rewrite every "it's not X, it's Y" / "not just X, but Y" sentence. Pick one side and commit. AI hedges by writing both halves; humans pick.
• Rhetorical-question clusters: delete questions that the next sentence answers. Replace with declaratives.
• Throat-clearing intros: kill any opener that begins "In today's digital age," "In the ever-evolving landscape," "Imagine a world where," "At its core," "At the heart of." Start with the actual content.
• Boilerplate conclusions: drop "In conclusion / Ultimately / To summarize / In closing" entirely. The conclusion is the last paragraph; it doesn't need to announce itself.
• Tricolon obsession: when you see three parallel adjectives or verbs ("clear, concise, and actionable"), keep one and cut two, or extend to four/five to break the rule of three.
• Faux-intimacy hooks: drop "Here's the thing," "Here's an uncomfortable truth," "The catch?", "The brutal truth?", "Real talk:". One-or-two-word transitional fragments meant to mimic a confiding voice. Delete and say the thing.
• Exhaustive-list pattern: "Whether you're a beginner or expert, in marketing or engineering, with ten minutes or two hours..." — cut to one (the most likely reader) and write to them.

Step 5 — Mechanism-level fixes (the ones that beat detectors, not just human readers):

• Inject adversatives. Real prose argues with itself. Add "but," "however," "although," "still," "that said." Detection studies show adversatives are a POSITIVE human signal — their absence is itself a tell. Aim for at least 1 per 500 words.
• Restore copulas. AI prefers richer verbs ("represents," "constitutes," "embodies") over plain "is" and "are." The Geng-Trotta arXiv paper found "is" declined 14% and "are" declined 17% in CS abstracts after ChatGPT. Flip this. Some sentences read cleaner with "X is Y" than "X represents Y."
• Add first-person epistemic stance. Human writing uses "I think," "I'm not sure," "in my experience," "it seems like." AI avoids these. Add at least one if the context allows.
• Spread word repeats (the Dispersion fix). AI's Dispersion is much lower than human (humans 16+; GPT-o4 mini 4.81 in the Kendro 2025 study). It repeats the same word within close proximity. If you used "important" twice in two paragraphs, vary the second one or restructure so it appears 4 paragraphs apart. Synonyms within 100 words look natural; the same word twice in 30 words looks like AI. This is the single most discriminating feature (SHAP=0.279 in the Kendro study).
• Choose against statistical context (the perplexity fix). When the obvious word is "examined," try "rummaged through" or "poked at" or "looked under the hood of." High perplexity comes from word choices the model would not have predicted. One unusual verb every 200 words is enough — don't over-correct.
• Use coordination over subordination. Fredrick & Craven 2025 found AI achieves complexity through subordination ("which were further refined by") while humans achieve length through coordination ("and," "but," "so"). When a sentence gets long, lean toward conjunctions rather than nested clauses.
• Specificity injection. Replace generic case-study references ("a manufacturing business," "a mid-sized SaaS") with named ones if verifiable. If not verifiable, drop the case study rather than invent (see Step 6).

Step 6 — Content-truth audit (only if the input contains citations, statistics, case studies, or expert quotes).

AI doesn't just write a certain way — it also fabricates specific things in specific ways. The fabrications can be removed but not "humanized." Per the Mata v. Avianca canon and the 1,457-case Damien Charlotin AI Hallucination Database (May 2026 count, up from 486 in October 2025), AI-generated text routinely invents:

• Case citations / academic citations. Verify every legal case, paper, or named study. The pinpoint-citation heuristic: AI-fabricated cites give court + year but omit page numbers, docket sub-numbers, exact filing dates. Real cites over-specify; AI under-specifies. If a citation is "Smith v. Jones, 2014" with nothing else, treat it as suspect until verified.
• Statistics. Flag every "approximately X%" or "studies show that..." stat. If no source is named, mark unverifiable. Real survey numbers are messy (38.7%, 71.4%, n=1,247); AI defaults to rounded numbers (40%, 75%, 8 in 10).
• Case studies / personas — the "Sarah Chen" rule. Per Michael G. Wagner's 2025 investigation, Claude defaults to "Sarah Chen" / "Marcus" / "Elara" / "Lyra" / "Kai" across wildly different contexts. ChatGPT and Llama have their own preferred archetypes. The mechanism is statistical (Western-common first name + globally-common surname + STEM-paper frequency), amplified by Constitutional AI's narrowing of variance toward "safe" archetypes. The DEI-cast quartet ("John, Maria, Wei, and Aisha") is also diagnostic. If a case study cites a Sarah Chen at a "mid-sized B2B SaaS company," treat as fabricated until proven otherwise.
• Expert quotes. Verify the quote actually exists. AI invents plausible-sounding quotes from real people who never said the thing. Search the quote on Google; if no source returns, it was generated.
• Auto-appended "Further Reading" sections. AI loves to close with "For further reading," "Resources," "If you want to learn more..." Recommended titles are generic, often invented, rarely linked. If a 3+ entry section has no specific titles or URLs, drop it.

Decision tree for each flagged item:

1. Can you verify it? Search the citation / persona / statistic. If verifiable, keep.
2. Can you replace it with a verifiable equivalent? If yes, replace.
3. Can you make the point without the example? If yes, drop the detail and keep the sentence.
4. Does the surrounding paragraph depend on the fabrication? If yes, drop the paragraph and rewrite to make the argument work without it.

What NEVER works: pretending the fabrication is real because "it makes the point well." That's the legal-canon problem reproduced in marketing copy.

PASS 3 — Validate (the mental probe)

After PASS 2, run this seven-point check:

1. Sentence length variance: do you have at least one sentence under 8 words and one over 28 words in the same 500-word stretch? If no, fix burstiness.
2. Adversatives: does "but," "however," "although," or "though" appear at least once per 500 words? If no, inject one or two.
3. Copulas: does "is" or "are" appear at least once per 300 words? If you wrote 1,000 words without a copula, you're in AI-fingerprint territory.
4. Dispersion: does any word (other than common function words) appear more than once within a 150-word window? If yes, vary or restructure.
5. First-person stance: if the context allows, is there at least one "I think," "I'm not sure," "in my experience"? If allowed but absent, consider adding.
6. Em dash density: at most 1 per 500 words.
7. Tricolons: zero or one per long piece, no clusters.

If you can answer all seven affirmatively, the mechanism layer is solid.

Then read it aloud (or simulate it). A human writer's prose has cadence, opinion, occasional clumsiness, and a voice. AI prose doesn't. If your rewrite still reads as a clean, evenly-paced, opinionless block, you removed the words but not the fingerprint.

Worked example — marketing blog opener

Before (AI, 247 words):

In today's rapidly evolving digital landscape, content marketing has become an increasingly pivotal component of any successful business strategy. As organizations across industries continue to navigate the complexities of an ever-changing marketplace, it's important to note that traditional approaches are no longer sufficient. The most innovative brands are leveraging cutting-edge technologies and embracing transformative strategies to foster deeper connections with their audiences.

At its core, modern content marketing isn't just about producing content — it's about crafting compelling narratives that resonate with target demographics. Furthermore, the rise of AI-powered tools has fundamentally redefined what's possible. From hyper-personalized email campaigns to dynamic landing pages, businesses now have access to an unprecedented array of solutions that can streamline their workflows and elevate their results.

[...continues with two more paragraphs of the same shape...]

After (humanized, ~165 words):

Most content marketing advice in 2026 reads like it was written by someone who has never had to ship a campaign on a deadline. It probably was. Here are five things I've actually seen move the needle on a real P&L this year. The list skips the obvious moves you've already tried.

1. Stop A/B-testing copy on landing pages with fewer than 10,000 weekly visitors. The math doesn't work and you're chasing noise. Use the time to write better hero copy instead.

2. Kill the AI-generated case studies. "Sarah Chen, marketing director at a mid-sized SaaS company" is not a case study; it's a hallucinated persona. Cite a real customer or don't cite one.

3. The email-personalization play is bigger than you think. I shipped one this March that doubled open rates.

Two more below. The honest summary: most of the AI-marketing-2026 articles repackage 2024 advice.

What changed: dropped the "In today's rapidly evolving" intro, killed 18 blacklist words (delve, leverage, navigate, landscape, foster, transformative, innovative, pivotal, cutting-edge, multifaceted, comprehensive, unprecedented, streamline, elevate, empower, unlock, at its core, furthermore), killed the negative-parallelism construction, killed the exhaustive-list pattern, added voice + opinion, added specific numbers (10,000 visitors threshold, March, doubled open rates), refused the 5-strategy structure and showed 3 with two "below" (human move — AI completes the promised list no matter what), mixed sentence lengths from uniform 18–24 down to 3–28-word range. Length cut 33%.

Worked example — LinkedIn thought piece

Before (AI, 263 words):

Three years ago, I made a decision that would fundamentally transform my career trajectory. I chose to step away from a comfortable corporate role at a Fortune 500 company to embark on an entrepreneurial journey. The path wasn't easy — it required navigating unprecedented challenges and cultivating a resilience I didn't know I possessed.

Today, I lead a team of 47 dedicated professionals across three continents [...]

But here's the most important lesson I've learned: It's not the destination that defines you — it's the courage to take the first step.

🚀 If you're sitting at your desk right now, wondering whether you're meant for more, I'm here to tell you: you absolutely are. ✨ The world needs your unique contribution. 💡 Don't let fear hold you back.

[...rhetorical question close + emoji-engagement bait + hashtag spam...]

After (humanized, ~205 words):

Three years ago I quit a corporate job to start a company. It was the wrong move twice and the right move once.

The wrong-move years were 2023 and 2024. I underpaid myself, hired the wrong COO (a friend, classic), and spent six months building a feature nobody wanted. We burned through $340K to learn things I should have learned for $30K. The team got to 12 people before I had to do two rounds of layoffs.

2025 was the right-move year. I rebuilt the leadership team, took a hard look at the product roadmap, and shipped one thing customers actually paid for. The team is 47 people now.

I don't think there's a lesson in this. The "took the leap before they were ready" framing is mostly survivor bias — for every person who left a stable job and made it work, two went back after a year. That's fine. Going back is fine.

The actually useful thing I learned: most early-stage decisions are reversible. The ones that aren't are usually about people.

What changed: killed "fundamentally transform," "embark on a journey," "navigating unprecedented challenges," "cultivating resilience," "dedicated professionals," "the most important lesson," "transformative ideas," "comfortable mediocrity"; killed two faux-intimacy hooks ("Here's the lesson," "Here's a hard truth"); killed two negative-parallelism constructions; killed the closing rhetorical question; killed the emoji garnish + engagement-bait + hashtag spam; killed the triple-negation ("Don't let X, Don't let Y, Don't let Z"); killed the survivor-bias claim ("every successful entrepreneur..."). Added specific failures and numbers ($340K, $30K, layoffs, 2023/2024/2025). Refused the forced lesson ("I don't think there's a lesson"). Added a real concrete operational insight at the end.

ChatGPT custom-instructions (prevent the problem at the source)

For users who want to reduce AI tells in default output rather than humanizing every response after the fact, paste this in Settings → Personalization → Custom instructions → "What traits should ChatGPT have?":

Write like a person, not a press release. Specific rules:
- No em dashes. Use a hyphen with spaces, a comma, or a new sentence.
- No "delve," "leverage," "robust," "navigate," "landscape," "tapestry," "multifaceted," "nuanced," "foster," "underscore," "comprehensive," "transformative," "pivotal," "seamless," "cutting-edge."
- No "It's important to note," "It's worth noting," "Furthermore," "Moreover," "In conclusion," "In summary."
- No "It's not just X, it's Y" or "Not just X, but Y" constructions.
- No "Imagine a world where," "At its core," "At the heart of," "In today's digital age."
- Vary sentence length. Mix short and long. Fragments are fine.
- Use "but," "however," "although" naturally — humans use these more than AI does.
- Use plain "is" and "are." Don't always say "represents," "constitutes," "embodies."
- First-person stance is okay. "I think" and "I'm not sure" are not weaknesses.
- Pick a side. Don't hedge both ways.
- Specific examples beat abstract ones. Name actual things.
- No tricolons of adjectives. One or two adjectives is enough.
- No rhetorical questions in clusters of three.
- No emojis in body text unless the conversation is casual.
- No "Hope this helps!" or "Feel free to reach out!" sign-offs unless writing a real email reply.
Be direct, opinionated, and concrete. Treat the reader as smart.

OpenAI explicitly fixed the em-dash overuse on November 14, 2025 (Sam Altman: "small-but-happy win"). The fix is opt-in via custom instructions, not default. The em-dash line above is load-bearing.

Definition of done

A humanized rewrite is done when:

• All canonical blacklist words are removed or replaced.
• Em-dash density is at most 1 per 500 words.
• Sentence-length variance has at least one fragment, one long compound sentence, and a mix in between.
• No negative-parallelism constructions remain.
• No throat-clearing intros or "in conclusion" outros remain.
• Fabricated content is removed or flagged.
• The seven-point validation check (PASS 3) passes.
• The voice has a take, not just information.

A humanized rewrite is NOT done just because "delve" is gone. The mechanism layer is the harder fix and the one that matters.

Gotchas

1. Don't humanize meta-text about AI tells. If the input is itself a guide to AI writing (like a report on AIisms), the AI-tell vocabulary is the subject matter, not a defect. Leave specimen examples alone. Only humanize the surrounding prose.

2. Don't strip legitimate technical vocabulary. "Robust" is a tell in marketing copy and a real word in statistics. "Significant" is a tell in business writing and load-bearing in scientific writing. Always read the domain. The blacklist applies to unjustified uses, not all uses.

3. Don't fake imperfection. Inserting deliberate typos or grammatical mistakes to "look human" produces text that reads as if someone is faking being human. Variance comes from rhythm and word choice, not errors.

4. Don't humanize beyond the target context. A LinkedIn post can be punchy and fragmented. A legal brief cannot. An academic essay needs higher register than a Substack post. Calibrate to the user's stated context. If they didn't specify, ask once.

5. The em dash is contested. Some users want every em dash removed; others use them happily. Default to lowering density (≤1 per 500 words) rather than zero. If the user explicitly says "kill every em dash," do it.

6. Detection scores are not truth. GPTZero / Originality.AI / Turnitin are biased against non-native English speakers (Stanford 2023: 61.22% false-positive rate on TOEFL essays). A "passing" score isn't proof the text is human; a "failing" score isn't proof it's AI. Warn the user if they're using detection scoring for high-stakes submissions — they should not rely on it alone.

7. The "Sarah Chen" rule. Case studies, personas, and named examples are fabricated until verified. Real case studies cite verifiable LinkedIn profiles or published interviews. AI defaults: Sarah Chen, Marcus, Elara, Lyra, Kai, plus DEI-cast quartets.

8. Don't apologize in the rewrite itself. AI loves to start humanized rewrites with "Here's a more human-sounding version of your text:" — that meta-framing is itself an AI tell. Just deliver the text.

9. Length math. Humanization usually shortens text by 15–30%. AI prose is padded with throat-clearing, hedges, and explanatory boilerplate. If your rewrite is the same length as the input, you didn't cut enough filler. If it's down 50%, you cut substance — restore.

10. Some AI is fine. If the user wants AI-assisted writing rather than AI-disguised writing, this skill is the wrong tool. Ask once if they actually want full humanization or just lighter editing.

When to ask the user instead of guessing

• Target context unclear: "Is this for a LinkedIn post, an email, a blog, a paper, or something else?"
• Specific detector to beat: "Is there a specific AI detector you need this to pass (GPTZero, Originality, Turnitin)?"
• Citation verification: if the input has citations and you can't verify them, ask before deleting: "I can't verify [citation X]. Want me to drop it, or do you have a real source?"
• Length target: if the user has a word count, ask. Humanization shrinks; you may need to compensate.