How to Edit AI Out of Your Writing

You are allowed to use AI to write. You are not allowed to publish what comes out of it without editing. Those are two different things, and the people who collapse them are the ones who get caught. This post is for the people who use Claude, ChatGPT, Gemini, or any other model to draft posts, emails, proposals, blog drafts, or whole books, and who want their published work to read like them and not like the model. The patterns below are the things to strip. The case studies are the reason it matters now and not in some abstract future.

Why does this matter right now?

Three things changed in 2025 and 2026 that made AI-detection move from background concern to real career risk.

• LinkedIn started limiting reach for AI-detected posts. In late 2025, LinkedIn’s VP of Global Editorial Laura Lorenzetti said publicly that the platform was rolling out measures to “detect and limit the reach of automated and AI-generated comments.” LinkedIn’s own line: it is fine to use AI to help you write, but your posts need to represent your voice. Third-party LinkedIn-marketing blogs put the reach impact at ~30% lower distribution for flagged posts, though LinkedIn itself has not published a specific number.
• A book contract was canceled within 24 hours of an AI score. In March 2026, Hachette pulled the planned US release of Shy Girl by Mia Ballard after a New York Times story citing a 78% AI-generated score from a detection vendor. The book had already gone through Hachette’s Orbit imprint editorial process. The story is more complicated than the headline. The author has since said an editor she hired on the self-published Kindle version added AI text without her knowledge, and the analysis was reportedly run on a pirated copy and not the Hachette-edited manuscript. Either way, the contract was gone in a day. Slate’s deep read is the best single account.
• Schools and employers started using detectors that they later admitted do not work well. Turnitin sells a 1% false-positive rate as a vendor claim. A Washington Post replication found ~50% false positives on a small sample. Australian Catholic University ran around 6,000 misconduct cases in 2024, the bulk of them AI-related, with a large share dismissed after closer review. Temple University and several others abandoned the Turnitin AI tool after internal testing. The bigger problem is bias. A 2023 Stanford paper led by Weixin Liang in Patterns tested seven AI detectors on 91 TOEFL essays written by non-native English speakers and found the detectors misclassified more than half as AI-generated, while flagging native-English essays at near-zero rates. Tighter syntax and simpler vocabulary read as “AI-like” to the detector.

The takeaway is not that you should stop using AI. It is that whatever you publish needs to not pattern-match on what every detector now looks for, and what every editor, reviewer, and platform algorithm has learned to spot. The bar is editing. Not detection avoidance. Just editing.

What happened to Shy Girl, and what does it teach you?

The short version. Shy Girl was self-published on Kindle Unlimited by Mia Ballard in early 2025. A Reddit user claiming to be a book editor flagged AI-style markers in the prose: repetitive adjective-before-every-noun, similes always in threes, list-of-three description patterns. Hachette’s Orbit imprint then signed Ballard and scheduled a US/UK publication. In March 2026, the New York Times ran a story alleging the book was largely AI-generated, citing a Pangram detection score of 78.4%. Hachette canceled the US release within 24 hours and discontinued the UK release. Ballard told the NYT that an acquaintance she had hired to edit the self-published Kindle version had used AI without her knowledge, and that she would pursue legal action against the editor.

The whistleblower, a publishing consultant named Thad McIlroy, later wrote his own account of how he broke the story for The Walrus. Two details from that account are worth knowing if you want to think about this clearly. First, the Pangram score was reportedly run on a pirated copy of the Kindle Unlimited version that McIlroy obtained from a file-sharing site, not the Hachette-edited manuscript. Second, McIlroy did not contact Ballard before pushing the NYT to publish. The case is being held up as proof that AI-detection works. It is also proof that the system in front of a $250,000 advance can collapse in a day, regardless of whether the underlying analysis was sound.

Three weeks earlier, in May 2025, three romance authors (K.C. Crowne, Rania Faris, Lena McDonald) were caught with literal ChatGPT prompts left inside their published books. McDonald’s Darkhollow Academy: Year 2 contained the line: “I’ve rewritten the passage to align more with J. Bree’s style, which features more tension, gritty undertones, and raw emotional subtext.” McDonald’s defense was that as a full-time teacher and mom, she could not afford a professional editor. 404 Media has the full story.

The lesson from both: even if you intend to use AI as a research and drafting aid, the published output cannot read like a model wrote it. That is the editing standard. The rest of this guide is the editing standard.

Are AI detectors actually accurate?

No, not in the way the vendors claim, and not in a way you should bet your career on (yours or someone else’s). Vendor-quoted numbers run 98 to 99 percent accuracy, but every independent benchmark surfaces meaningful false-positive rates. Originality.ai shows around 5% false positive in one independent test; GPTZero around 12%; Copyleaks varies. Walter Writes ran a 2026 reliability survey across five detectors and concluded that no single tool should be used as the sole evidence in a high-stakes decision. Their writeup is here.

The two more important facts. Turnitin itself acknowledged that documents in the 0-20% AI-flagged range have a higher false-positive rate, and now flags those with an asterisk. And per the same Stanford research, non-native English speakers are flagged at much higher rates because the detectors read simpler vocabulary and tighter syntax as AI-like.

So the real risk is not that you wrote the post in five minutes with Claude and forgot to edit it. The real risk is that you wrote it yourself in a clean, plain style, and a detector flags it anyway. The defense against both is the same. Make the writing sound like a person made specific choices, not like a model running on autopilot.

What are the AI writing tells you should strip?

The most authoritative single inventory of these patterns is the Wikipedia community guide called Signs of AI writing, built by editors who spent two years triaging AI-generated submissions. Empirical work backs it up. Kobak et al. studied 14 million academic abstracts and found 11 specific words (delve, intricate, underscores, showcasing, pivotal, and others) jumped 5 to 25 times their pre-ChatGPT frequency starting in 2023. A 2025 follow-up paper, “Human-LLM Coevolution: Evidence from Academic Writing,” found something quieter and more useful: once “delve” became publicly mocked as an AI tell, its frequency in academic writing actually dropped. Writers were editing it out. Patterns that escaped public attention (like the rise of “significant”) kept climbing.

What follows is the practical inventory. Categories first, then specific tells with example sentences and a rewrite.

Category A: Vocabulary tells

Category B: Sentence-frame tells

• “In today’s fast-paced world…” Cut it. Open with the specific point.
• “In today’s digital age…” Same. Cut.
• “It’s important to note that…” Cut. Just say the thing.
• “It’s worth noting that…” Cut.
• “When it comes to X…” “When it comes to pricing, transparency matters” becomes “Pricing should be transparent.”
• “Whether you’re X, Y, or Z…” “Whether you’re a beginner, intermediate, or expert…” becomes “Even if you’ve never opened the tool.”
• “From X to Y, here’s everything…” Collapse to one specific lead.
• “Not just X, Y as well.” Once per article is fine; AI uses it every paragraph.
• “In conclusion…” / “To wrap up…” / “In summary…” Cut. End on a sentence that matters.
• Closing rhetorical question: “So, are you ready to transform your business?” Cut.

Category C: Rhythm and structure tells

These are the tells AI-trained detectors look for that most human editors miss.

• The Rule of Three, applied everywhere. Every list has exactly three items. Every sentence has three adjectives. Real human writing varies. Force yourself to make some lists two items, some four.
• Symmetrical paragraph length. Every paragraph runs three sentences of similar length. Mix in one-line punches and longer beats.
• Stacked subordinate-clause openers. “While many tools exist, choosing one is hard. While each promises results, few deliver.” Vary openings.
• Present-participle dangle. “…creating a seamless experience, fostering engagement, and driving results.” Cut the trailing trio.
• Heading parallelism. Every H2 starts with the same grammatical form. Vary: a question, a fragment, a directive.
• The “Not X. Y.” punchy two-sentence rhythm, used in every section. Once per article, fine.

Category D: Formatting and punctuation tells

• Em-dash overuse. Wikipedia’s signs guide notes that LLMs use em-dashes more than nonprofessional human writers, especially where commas, parentheses, or periods would feel natural. Replace ~80% of em-dashes with one of those. (OpenAI’s GPT-5.1 was tuned to suppress em-dashes specifically because of this association.) Target: at most one em-dash per 500 words.
• The bullet + bold-header + colon + description pattern. “Speed: It’s fast. Cost: It’s cheap.” Wikipedia calls this exact pattern a hallmark of AI prose. Use prose lists instead, or strip the bolding.
• “Key Takeaways” callout in every section. Fine on a long pillar. Deadly when it appears every 400 words.
• Title Case In Every Heading. Default AI behavior. Sentence case has more personality.
• Excessive bolding inside paragraphs. Every key noun bolded looks like an output, not a written piece.

Category E: Tone tells

• Promotional puffery. “Renowned,” “groundbreaking,” “cutting-edge,” “innovative,” “world-class.” Cut all of them.
• Vague attribution. “Experts say” / “studies show” / “industry reports indicate.” Name the expert. Link the study. If you cannot, cut the sentence.
• The “stands as a testament” formula. “Her work stands as a testament to dedication” becomes “She worked on it for nine years.”
• Hedged optimism close. “Despite challenges, the future looks bright.” AI loves this ending. Skip it.
• “In an era of…” openers. Cut.
• The “genuine” or “authentic” insertion. “A genuine connection” or “truly authentic experience” is almost always padding.
• “Honestly…” or “Honest truth…” Cut. If you have to say it’s honest, the rest of the writing isn’t doing its job.

How do you actually filter these patterns from your writing?

Two ways, used together. The first is a prompt you give your AI tool when the draft is done. The second is a manual five-pass checklist that takes about ten minutes for a 2,000-word post.

The pre-publish prompt (copy this into Claude, ChatGPT, or Gemini)

The manual five-pass checklist (about 10 minutes for a 2,000-word post)

1. Ctrl-F sweep. Search the document for every word in the vocabulary table above. Cut or replace each one.
2. Em-dash audit. Find every em-dash. Count them. Target less than one per 500 words.
3. Sentence-opener audit. Read down the left margin of the document. If most sentences start with subordinate clauses or present participles, recast half of them.
4. Three-count. Find every list of exactly three items. Change one to two or four.
5. Read it aloud. If you wouldn’t say a sentence to a friend, rewrite it. This single pass catches more AI prose than any tool.

What is the complete catalog of AI writing tells?

The most-cited reference work on this is Wikipedia's “Signs of AI writing” guide, maintained by WikiProject AI Cleanup. The guide was built from observations of thousands of instances of AI-generated text appearing in Wikipedia edits and is the de-facto authority that researchers, journalists, and platform-moderation teams reference. The 29-pattern list below is organized in four buckets the way the WikiProject volunteers organized it: content patterns, language patterns, style patterns, and communication patterns. Each entry includes a short Before / After example so you can see the move and the fix.

If you write a lot, the most efficient way to run this list across your drafts is the open-source Humanizer skill for Claude Code and OpenCode (MIT-licensed, 20K+ stars on GitHub, maintained by Siqi Chen). It encodes the full 29-pattern list as a single skill you can invoke with /humanizer in either tool. Voice-calibration is built in: paste a few paragraphs of your own writing and the skill matches your sentence rhythm rather than producing generic “clean” output. The patterns below are also worth knowing manually so you can spot them when you read your own drafts.

Content patterns

• 1. Significance inflation. AI inflates ordinary facts with weight they do not earn. Before: "marking a pivotal moment in the evolution of…" After: "was established in 1989 to collect regional statistics."
• 2. Notability name-dropping. Long chains of outlet citations as a credibility shortcut. Before: "cited in NYT, BBC, FT, and The Hindu." After: "In a 2024 NYT interview, she argued…" (one specific reference does more work).
• 3. Superficial -ing analyses. Strings of “-ing” verbs substituted for real analysis. Before: “showcasing the platform's flexibility while underscoring its transformative impact and reflecting industry trends.” After: “The platform handles both. Adoption tripled in six months.”
• 4. Promotional language. Travel-brochure phrasing in non-promotional contexts. Before: "nestled within the breathtaking region." After: "is a town in the Gondar region."
• 5. Vague attributions. Claims attributed to “experts” / “studies” with no source. Before: "Experts believe it plays a crucial role." After: "According to a 2019 survey by Pew Research…"
• 6. Formulaic challenges. The flattening of specific obstacles into a single rhetorical move. Before: “Despite challenges in the industry, AI continues to thrive.” After: “Adoption stalled in healthcare and legal because of privacy review cycles, but it's up 40% in marketing.”

Language patterns

• 7. AI vocabulary. Words AI reaches for at much higher rates than human writers. Before: “Additionally, it serves as a testament to the evolving landscape of innovation.” After: “It also signals that the field has changed.” Common offenders: actually, additionally, testament, landscape, showcasing, underscores.
• 8. Copula avoidance. AI dodges “is” and “has” with substitutes that sound more important than they are. Before: "serves as a foundation," "features extensive coverage," "boasts strong performance." After: "is a foundation," "has extensive coverage," "has strong performance."
• 9. Negative parallelisms and tailing negations. "It's not just X, it's Y" structures and "…, no guessing" tails. Before: "It's not just a tool, it's a transformation." After: "It changes how this gets done."
• 10. Rule of three. Forced triplets used as a stylistic crutch. Before: "innovation, inspiration, and insights." After: Use the natural number of items, even if that is two or four.
• 11. Synonym cycling. AI cycles through synonyms to avoid repetition even when repetition is clearer. Before: "protagonist / main character / central figure / hero." After: "protagonist" (repeat when clearest).
• 12. False ranges. Pseudo-comprehensive ranges that bracket the topic. Before: "from the Big Bang to dark matter." After: List the topics directly.
• 13. Passive voice and subjectless fragments. Subjectless fragments and unnecessary passives that hide the actor. Before: "Mistakes were made and lessons were learned." After: "We shipped the bug. Here's what we learned."

Style patterns

• 14. Em-dash overuse. AI uses em-dashes at much higher rates than human writers. Before: "institutions—not the people—yet this continues—." After: Prefer commas or periods. One em-dash per page is fine; one per paragraph is the tell.
• 15. Boldface overuse. Random terms bolded for emphasis the writer cannot justify. Before: "OKRs, KPIs, BMC." After: "OKRs, KPIs, BMC."
• 16. Inline-header lists. Bullet items that bold-restate the bullet name. Before: "Performance: Performance improved." After: Convert to prose or drop the redundant header.
• 17. Title Case Headings. Title-casing every word in a heading. Before: "Strategic Negotiations And Partnerships." After: "Strategic negotiations and partnerships."
• 18. Emojis. Decorative emojis at the start of headers or bullets in formal contexts. Before: “🚀 Launch Phase · 💡 Key Insight · ✅ Done.” After: “Launch / Insight / Done.” Remove unless the publication actually uses emojis as part of its house style.
• 19. Curly quotes. Smart-quote inconsistency in the same draft. Before: She said “the project” in one paragraph and “the project” in the next. After: Pick one quote style and apply it consistently throughout.
• 26. Hyphenated word pairs. Stacked compound modifiers signaling business-vocabulary affect. Before: "a cross-functional, data-driven, client-facing approach to value creation." After: "a team that uses customer data to plan its work." Drop the hyphens or commit to one strong modifier.
• 27. Persuasive authority tropes. Authority-borrowing phrases that delay the actual claim. Before: "At its core, what matters is that ultimately, the truth is engagement." After: "Engagement is what matters here." Also retire: fundamentally, in essence, the bottom line is.
• 28. Signposting announcements. Announcements before the actual content arrives. Before: "Let's dive in! Here's what you need to know about Claude Code." After: "Claude Code is Anthropic's terminal CLI for software work. The install command is one line."
• 29. Fragmented headers. Bare-noun headers followed by one sentence the heading already implied. Before: ## Performance followed by "Speed matters." After: ## Why Anytype feels slow on first launch followed by "It downloads 200MB of conflict-free history before anything is usable." If the section is one line, fold it into the parent prose.

Communication patterns

• 20. Chatbot artifacts. Closer phrasing AI defaults to in any reply context. Before: "I hope this helps! Let me know if you need anything else." After: Delete the closer. The reply is the answer; the offer to help more is implied by being available.
• 21. Cutoff disclaimers. The training-cutoff hedge dressed up as humility. Before: "While specific details are limited in available sources, it could potentially be argued…" After: Check a current source and name it: "Anthropic's pricing page on 2026-05-23 lists Pro at $20/month."
• 22. Sycophantic tone. Opener phrasing that performs enthusiasm AI is trained to provide. Before: "Great question! You're absolutely right that…" After: Start with the answer. Compliments mid-conversation read as filler.

Filler and hedging

• 23. Filler phrases. Multi-word filler phrases where one word does the same work. Before: "In order to access the feature, due to the fact that it requires authentication…" After: "To access the feature, sign in first." Also retire: at this point in time (now), in spite of the fact (although), with regard to (about).
• 24. Excessive hedging. Hedge-word stacks signaling unwillingness to commit. Before: "It could potentially possibly help with that." After: "It might help." Or just commit: "It helps in two cases (X and Y)."
• 25. Generic conclusions. Closing platitudes that fill space without saying anything. Before: "The future looks bright as exciting times lie ahead in this rapidly evolving space." After: Replace with one specific next step ("The next Anthropic release is Q3 2026") or delete the conclusion entirely.

What is “AI-fingerprinting” your writing?

The concept is simple. Pick three to five specific moves that are yours and use them deliberately. A phrase. A punctuation habit. A willingness to start sentences with “And” or “But.” A specific kind of analogy you reach for. An AI imitator can match your topic. It cannot match your fingerprint without being trained on a huge corpus of your work.

Examples of fingerprint moves that show up in respected human writers:

• Mary Karr’s habit of dropping a single-word sentence after a long descriptive run.
• Helen Garner’s use of the word “anyway” mid-paragraph.
• David Sedaris’s tendency to undercut his own sentences with a parenthetical.
• Patti Smith’s run-on lyrical openings followed by a clipped declarative.

You probably already have at least one fingerprint move. Notice it. Use it more often. That single decision will do more to mark your writing as yours than any anti-AI checklist.

What does this mean for LinkedIn, college, and hiring filters?

• LinkedIn. Disclose AI use in the first two sentences if you used it heavily. LinkedIn’s own guidance is that AI-assisted drafting is fine if the post represents your voice. The reach throttling appears to target obvious patterns. Edit before you post.
• College. Your school’s policy varies wildly. Find the actual policy in the syllabus or student handbook before deciding what to disclose. Universities like Australian Catholic dismissed a large share of AI-misconduct cases on closer review, but the initial accusation still cost students months. The safest default is to disclose what you used and how.
• Job applications. A 2025 TopResume survey of hiring managers found 52% said AI-assisted writing in cover letters and resumes was fine. The dealbreaker was submitting unedited AI output. The edit is what matters.
• Academic publishing. Most journals now require AI-use disclosure. They do not yet run automatic detection on every submission. The risk is post-publication discovery and retraction.

Frequently asked questions

If I use this prompt, will AI detectors still flag my work?

Possibly, but at lower rates. The detectors mostly look for the same patterns this guide strips. What they cannot do is verify intent. Even a perfectly human-written piece can score 70% AI-generated if the style happens to match the patterns the model was trained on. Your defense is editorial: make the writing concrete, specific, and tied to your own voice.

Is it okay to use AI at all?

Yes. The view at Beginners in AI is that AI is a tool that augments human work, not a substitute for human judgment, voice, or relationships. Use it for research, drafting, idea generation, and editing assistance. Then put your own voice on the output. The case studies in this post are about people who skipped that last step, not about people who used AI in the first place.

What about non-native English speakers who get flagged?

Several studies found that AI detectors flag non-native English speakers at significantly higher rates than native speakers. The fix is the same as for everyone else: add specific personal context, vary sentence rhythm, name sources, use concrete details. If you are a non-native speaker who has been flagged, ask your institution to review the detector’s accuracy data and to apply human judgment. Several universities have walked back their detector use after seeing the disparity.

Should I just rewrite by hand and skip AI?

You can, but you do not have to. Most professional writers we know use AI for research, outline, and first draft (see our best Claude prompts for drafting templates), then put two to three full editing passes on top. That workflow produces better writing than either alone. The mistake is publishing the AI’s first draft.

Does this guide apply to fiction?

Yes, more so. The Mia Ballard case shows what happens when an editor pastes AI prose into a novel without re-voicing it. Fiction has a higher bar than nonfiction because reader trust depends on hearing the author’s voice. Strip every tell. Add your fingerprint. Read the chapter aloud.

Sources and where to go deeper

• Wikipedia: Signs of AI writing (the WikiProject AI Cleanup authoritative guide)
• Humanizer skill on GitHub (open-source Claude Code / OpenCode skill encoding the 29-pattern list)
• Wikipedia: Signs of AI Writing (community-curated inventory of AI tells)
• Slate: Shy Girl by Mia Ballard, the canceled book
• The Walrus: How the whistleblower broke the Shy Girl story
• 404 Media: Authors leaving AI prompts in their novels
• Entrepreneur: LinkedIn limiting AI-content reach (Lorenzetti statement)
• Social Media Today: LinkedIn’s AI content stance
• K-12 Dive: Turnitin’s higher false-positive rate
• USD Law Library: AI detector false-positive research
• Kobak et al., “Delving into ChatGPT usage in academic writing” (arXiv)
• “Human-LLM Coevolution: Evidence from Academic Writing” (arXiv)
• Lit Hub: A handy guide to spotting AI writing

You may also like

• Best AI Prompts for Cover Letters — the editing pass applied to actual cover-letter drafts
• Best AI Prompts for LinkedIn — where AI tells get you down-ranked silently
• Best AI Prompts for Blog Writing — the editing pass for long-form content
• Prompt to Workflow: The AI Ladder — when to turn the editing pass into a reusable skill
• AI Prompt Library — 200+ tested prompts in 11 categories