Anti-AI Writing Style Guide

A reference for writing that avoids the telltale patterns of LLM output. Based on Wikipedia’s “Signs of AI writing” (WikiProject AI Cleanup), which catalogues patterns observed across thousands of AI-generated submissions.

The point of this guide is not that any single pattern below “proves” writing is AI-generated (LLMs were trained on human writing, so humans produce these patterns too). The point is that a cluster of these patterns is the strongest tell, and avoiding them produces tighter, less generic prose.

Core principle

If a sentence would fit inside a promotional brochure, a travel guide, or a generic LinkedIn post, rewrite it. Neutral, specific, declarative prose is the baseline. Let facts do the work; do not dress them up. This guide defines how I strip out AI-like patterns and where I deliberately add my own voice on top of that.

1. Language and tone

Puffery and grandiose language

Avoid writing that praises or elevates the subject beyond what the evidence supports. This is the single most common AI tell.

Watch-words and watch-phrases to cut:

  • stands as a testament (to)
  • plays a vital role / plays a crucial role / plays a pivotal role
  • underscores its importance / highlights the importance of
  • a testament to the power of [X]
  • rich cultural heritage / rich history / rich tapestry / rich cultural tapestry
  • breathtaking, stunning, must-see, must-visit
  • stunning natural beauty
  • enduring legacy / lasting legacy
  • nestled (in / within / between)
  • in the heart of
  • a beacon of / a hallmark of / a symbol of
  • renowned for, celebrated for, widely regarded as

Rewrite: replace evaluative adjectives with concrete nouns and numbers. “Nestled in the breathtaking region of X, Y stands as a vibrant town with a rich cultural heritage” becomes “Y is a town of 40,000 in X region.”

Editorializing

Do not add commentary the facts did not ask for. Red-flag structures:

  • A “Challenges” section tacked onto the end, opening with “Despite its [positive thing], [subject] faces challenges…”
  • A “Future Outlook” / “Looking Ahead” section that speculates or cheerleads.
  • Closings that affirm the subject’s lasting importance or recommend what it “should” do next.
  • Call-to-action language in contexts that are not advocacy.

If the source does not analyse the future, you do not analyse the future.

Vague attribution (weasel wording)

Avoid claims attributed to phantom authorities:

  • Industry reports suggest…
  • Observers have noted / observers have cited…
  • Some critics argue…
  • Experts say / many believe / it is widely acknowledged…
  • Has been described as…

Name the source. If you cannot name the source, cut the claim.

It is fine to say “In my projects…” or “In the deployments I’ve seen…” when describing concrete experience. Do not inflate this to “most deployments” or “industry standard” unless you can cite data.

Undue emphasis

Do not let one easily-Googled source dominate. If every third sentence references the same article, you are compiling, not writing.

Section-ending summaries

Do not close sections with “In summary,” “In conclusion,” “Overall,” or “Taken together.” The reader just read the section. Trust them.

Overused transition conjunctions

Sentence-starting transitions that signal LLM pacing:

  • Moreover,
  • Furthermore,
  • Additionally,
  • On the other hand,
  • That said,
  • It is worth noting that,
  • It is important to note that,

Most can be deleted outright. The next sentence rarely needs a usher.

2. Sentence-level tics

Negative parallelism

The most diagnostic single pattern. Forms:

  • “It’s not X, it’s Y.” — “It’s not a product launch. It’s a paradigm shift.”
  • “Not X but Y” / “Not only X but also Y”
  • “No X, no Y, just Z.” — “No UN mandate, no congressional approval, just raw military power.” / “No fever. No crisis. Just vibes and breakfast.”

If you notice yourself writing one, rewrite as a flat declarative.

Rule of three (triplets)

LLMs default to three-item lists to simulate thoroughness. Examples flagged as vacuous:

  • “innovative, transformative, and groundbreaking”
  • “social, political, and economic factors”
  • “keynote sessions, panel discussions, and networking opportunities”

If each item is not doing distinct work, collapse to one specific item or drop the list.

Grandiose adjective stacking

Avoid chains of abstract positive adjectives (“transformative, visionary, unparalleled”). Use one concrete adjective or none.

AI-frequent vocabulary

Words that appear far more often in LLM output than in ordinary writing. Use sparingly, and never as a default:

delve, intricate, tapestry, pivotal, underscore, landscape (figurative), foster, testament, enhance, crucial, essential, vital, robust, seamless, leverage, navigate (figurative), bustling, vibrant, myriad, plethora, realm, nuanced, holistic, paradigm, synergy, ecosystem (figurative).

None of these words are wrong in isolation. The problem is the density.

3. Style and formatting

Em dashes

LLMs overuse em dashes—especially as a “punched-up” sales rhythm—in nearly every paragraph. Human writing uses them, but less often and for specific effect. If a draft has an em dash every other paragraph, replace most with commas, periods, or parentheses.

Bolded list headers

The signature LLM list shape is:

  • Header: descriptive sentence fragment.
  • Header: descriptive sentence fragment.

Avoid this unless the document truly is a reference/spec. For prose, write prose.

Title case subheadings

Sentence case (“System design”) is less AI-flavoured than title case (“System Design”). Pick sentence case and stick with it.

Thematic breaks before headings

--- inserted above every heading is a Markdown-output artifact. Do not add decorative separators.

Excessive bold and italics

LLMs bold key terms in nearly every sentence. Bold at most one phrase per paragraph, and only where it genuinely helps scanning.

Emojis

A “checkmark / rocket / sparkles” cluster at the top of a section is a strong tell. Avoid emojis in serious writing.

Unnecessary lists

If an idea flows as two sentences of prose, write two sentences of prose. Bullets are not free structure; they impose a shape on ideas that may not want one.

4. Leftover artifacts (the obvious tells)

These should never survive a read-through, but they regularly do:

  • “As an AI language model…” / “As an AI, I…”
  • “I’m sorry, but…” / “I cannot…” (when the surrounding text is not a refusal)
  • “Certainly!” / “Great question!” / “Of course!” as an opener.
  • “Here is the revised version:” / “Below is the text you requested:”
  • “I hope this helps!” / “Let me know if you’d like me to adjust anything.”
  • References to the prompt itself: “As you requested…”, “Per your instructions…”
  • Meta-commentary about Wikipedia’s conventions or guidelines in the article body.
  • Fabricated citations — author/title/journal combinations that look plausible but do not exist. Always verify.

5. Positive rules (what to do instead)

  1. Prefer specific nouns and numbers over evaluative adjectives.
  2. Name your sources inline or in citations; never “observers.”
  3. One point per sentence. Let sentences end.
  4. Vary sentence length deliberately; avoid the steady mid-length rhythm LLMs default to.
  5. Cut the first and last sentence of each paragraph and see if the paragraph still works. Often it does.
  6. If a paragraph could describe any subject in its category with minor word swaps, it is generic. Rewrite with details only this subject has.
  7. Read it aloud. Sycophancy, puffery, and triplet-padding are audible.

6. Personal voice (how I want to sound)

These rules are about sounding like me, not like a neutral spec.

  1. I can speak from concrete experience

    It is fine to write in the first person when I’m talking about work I’ve actually done or systems I’ve seen: “In my POS projects, offline PIN is a resilience control first, a fraud control second,” or “In the acquirer stacks I’ve worked on, misconfigured CAPKs show up as ‘mystery declines’ long before anyone says ‘cryptography’.” I do not upscale that to “most deployments” or “the industry” unless I can point to data or named sources.

  2. Opinion is allowed, but it must ride on mechanics

    I can state preferences and judgments: “I’d rather pay the complexity cost of offline PIN than accept terminals that go blind during outages,” or “Treat AI-generated code as untrusted input, not as a correctness guarantee.” I avoid “X is the best” and instead tie the opinion to trade-offs, risk, or architecture.

  3. Plain, dry informality is okay

    Short, direct phrases like “this is where things get ugly,” “that model is incomplete,” or “the field will remind you it isn’t static” are fine when they clarify behavior. I avoid hype and cheerleading (“this is where the magic happens,” “game-changing,” “revolutionary”) and skip slang that will age badly.

  4. Metaphors must map to systems, not hype

    I can use simple system metaphors when they line up with how things actually work: “offline PIN is a local circuit breaker when the upstream grid is down,” “CAPKs are the trust anchors in the terminal,” or “AI leadership is a systems race, not just a chip race.” I avoid metaphors whose main job is to make the subject sound grand (“guardian angel for transactions,” “the beating heart of the ecosystem”).

  5. Rhythm over polish

    I aim for uneven rhythm on purpose: a few very short sentences for emphasis (“That interpretation misses what PIN actually does.” / “That is the harder question.”), then longer ones when the idea needs space. If a sentence turns into a slogan (“Financial trust is not just a feeling. It is an architecture…”), I keep it only if it names concrete mechanisms; otherwise, I flatten it into plain prose.

  6. Experience beats abstraction

    Wherever possible, I swap general claims for specific patterns, incidents, or failure modes: “Misconfigured CAPKs create ‘everything looks healthy but transactions still fail’ incidents,” “Reversals are instant at the message level, but issuer hold release may lag hours,” “Vibe programming makes accountability harder to trace in regulated stacks.” If a paragraph could describe any industry with a few word swaps, I rewrite until it could only describe payments, AI infrastructure, or the systems in front of me.

Quick self-check before shipping

  • Any sentence starting with Moreover / Furthermore / Additionally? Cut or rewrite.
  • Any “It’s not X, it’s Y” / “not only X but Y”? Rewrite flat.
  • Any “stands as a testament” / “plays a vital role” / “rich tapestry” / “nestled”? Delete.
  • Any trio of abstract adjectives? Pick one or none.
  • Any “In conclusion” / “In summary” / “Overall”? Delete.
  • Any bolded-header bulleted list where prose would do? Convert to prose.
  • Em dashes per page — are there more than two or three? Trim.
  • Any “As an AI” / “Certainly!” / “I hope this helps” residue? Scrub.
  • Any “Challenges” or “Future Outlook” section you added without a source? Delete.
  • Any claim attributed to “observers” / “experts” / “industry reports”? Name them or cut.
  • Any sentence that reads like a tagline (“X as the future of Y”, “redefining the way we do Z”)? Keep it only if it names a concrete mechanism or trade-off; otherwise, rewrite as a plain statement.

If the draft survives all these checks, it no longer reads like AI and still sounds like me.

Source

Wikipedia, Wikipedia:Signs of AI writing (WikiProject AI Cleanup), published August 2025. This document is a condensation and rephrasing of that page’s catalogue, aimed at writers who want to avoid the patterns rather than detect them.