Wikipedia's Signs of AI Writing: The Full List, Plus a Copy-Paste Prompt You Can Turn Into a Skill

Wikipedia's volunteer editors built the most complete public catalog of AI-writing tells available: em dashes, triplets, vague attribution, and about a dozen other patterns. This post walks through the list in plain language, gives you a copy-paste version formatted as reusable instructions, and shows you how to turn that into a custom skill or project in Claude, Gemini, Perplexity, or ChatGPT so you can run it against your own drafts automatically.

What You'll Learn

• What Wikipedia's "Signs of AI Writing" guide is and why it exists
• The full list of patterns it catalogs, in plain language
• A copy-paste prompt block you can use as-is
• Step-by-step instructions to turn that block into a Claude Skill, a Gemini Gem, a Perplexity Space, or a ChatGPT Custom GPT
• A second copy-paste block built from HumanizeAI's H.E.A.R.T. framework, for checking what a subtraction checklist can't: whether the piece actually adds anything
• A standalone version of H.E.A.R.T. formatted as a writing prompt, not just a checker
• What HumanizeAI's own fine-tuned model checks for beyond this list

The TL;DR

Wikipedia's "Signs of AI Writing" is a public guide maintained by WikiProject AI Cleanup, a group of volunteer editors who needed a shared reference for catching undisclosed AI-generated text before it gets published on the encyclopedia. It documents roughly a dozen recurring patterns: things like formulaic em dash use, three-item lists, vague attribution ("studies show"), and stock phrases like "it's important to note."

The guide itself is careful to say none of these signs prove AI authorship on their own. They're most useful as a combined signal, not a single tripwire. Below is the plain-language rundown, followed by a copy-paste version you can drop into any AI assistant to check your own writing against the same list.

The Breakdown

What is Wikipedia's Signs of AI Writing guide?

It's an internal editorial reference built by Wikipedia's volunteer community over thousands of real cases of AI-generated text showing up in encyclopedia articles. The goal was never to teach anyone how to write. It was to give editors a shared checklist for spotting undisclosed AI content, since Wikipedia requires that contributions reflect genuine editorial judgment, not an unedited model output.

The guide is public, actively maintained, and openly credited as a Wikipedia community project, which is part of why it has become the reference point for AI-detection discussion well beyond Wikipedia itself.

What are the actual signs, in plain language?

Here's the list, condensed from the guide:

1. Em dashes used formulaically. AI models use em dashes more often than typical human writing, and often in spots where a comma, colon, or parenthesis would read more naturally.
2. Rule-of-three lists. A reflex toward triplets: "innovative, transformative, and groundbreaking" instead of a single word or a differently sized list.
3. Vague attribution. "Studies show," "experts say," "observers have noted," with no name attached.
4. Overemphasis on significance. Everything "plays a vital role" or "serves as a testament," even when the underlying fact is minor.
5. False ranges. "Our services range from strategic planning to implementation support," a construction that sounds specific but says nothing concrete.
6. Section summaries that restate what was just said. "In summary," "in conclusion," "overall," used as filler rather than to add anything new.
7. Superficial tacked-on analysis. Phrases like "highlighting the shift" or "illustrating the impact" added after a plain fact, without actually explaining why it matters.
8. Negative parallelism. "It's not just about X, it's about Y," a contrast-reframe pattern that shows up constantly in AI output and rarely in natural speech.
9. Editorializing asides. "It's important to note," "no discussion would be complete without," inserted opinions about what matters that a neutral reference shouldn't be making.
10. Letter-style phrasing in non-letter content. "I hope this message finds you well," dropped into a blog post or article where it doesn't belong.
11. Leftover collaborative phrases. "I hope this helps!" or "let me know if you need anything else," a sign the AI output was copied in without review.
12. Formatting habits. Heavy use of boldface, excessive lists, or title-case headings applied inconsistently.

Does having one of these signs mean something was written by AI?

No, and this is the part most summaries of the guide leave out. Wikipedia's own editors are explicit that a single sign proves nothing. Human writers use em dashes. Human writers reach for triplets. The guide is meant to be read as a pattern-matching tool across several signs at once, in context, not as a single-strike accusation.

How can I check my own writing against this list?

The fastest way is to paste your draft into an AI assistant along with a clear instruction set built from the list above, and ask it to flag matches. Below is a ready-to-use version of that instruction set. Copy the whole block and paste it wherever you want to run the check.

COPY-PASTE BLOCK – Signs of AI Writing Checker

You are reviewing a piece of writing for patterns commonly associated with
unedited AI-generated text, based on Wikipedia's "Signs of AI Writing" guide
(WikiProject AI Cleanup). Read the text I provide and flag any instances of
the following, quoting the exact sentence or phrase for each one you find:

1. Em dashes used in place of commas, colons, or parentheses
2. Rule-of-three lists (three-item lists used as a rhetorical crutch)
3. Vague attribution with no named source ("studies show," "experts say")
4. Overemphasis on significance ("plays a vital role," "serves as a testament")
5. False ranges that sound specific but say nothing concrete ("ranges from X to Y")
6. Section summaries that just restate prior content ("in summary," "in conclusion," "overall")
7. Superficial analysis tacked onto a plain fact ("highlighting," "illustrating," "underscoring" used as filler)
8. Negative parallelism / contrast-reframe statements ("it's not just about X, it's about Y")
9. Editorializing asides ("it's important to note," "no discussion would be complete without")
10. Letter-style phrasing in non-letter content ("I hope this message finds you well")
11. Leftover collaborative phrases ("I hope this helps!", "let me know if you need anything else")
12. Inconsistent or excessive use of boldface, lists, or title-case headings

For each flagged instance, give me:
- The exact quote
- Which numbered pattern it matches
- A suggested rewrite that keeps the meaning but removes the pattern

Do not flag a pattern just because it could theoretically apply. Only flag
clear, specific matches in the actual text I provide. If nothing matches a
given pattern, say so and move on.

Here is the text to review:

[PASTE YOUR DRAFT HERE]

How do I turn this into a reusable skill or custom assistant?

Each major AI platform has its own version of "save these instructions so I don't have to paste them every time." Here's how to set it up in each one, using the copy-paste block above.

Claude (Skill or Project)

• For a Project: open Claude, create a new Project, and paste the copy-paste block into the Project's custom instructions field. Every chat inside that Project will apply the check automatically.
• For a Skill: if you have access to Claude's Skills feature, create a new skill, name it something like "AI Writing Tell Checker," and paste the block as the skill's instructions. Invoke it by name whenever you want to run a check.

Gemini (Gem)

• Open Gemini, go to "Gems," and select "New Gem."
• Give it a name (e.g. "AI Tell Checker") and paste the copy-paste block into the instructions field.
• Save it. It will appear in your Gems list for one-click reuse.

Perplexity (Space)

• Open Perplexity and create a new Space.
• In the Space's instructions or system prompt field, paste the copy-paste block.
• Use that Space any time you want to check a draft, and Perplexity will apply the same instructions to every thread inside it.

ChatGPT (Custom GPT or Custom Instructions)

• For a Custom GPT: go to "Explore GPTs" and select "Create." In the Instructions field, paste the copy-paste block, name your GPT, and save it for reuse.
• For lighter use: go to Settings, then Custom Instructions, and paste a shortened version of the block into the "How would you like ChatGPT to respond" field. This applies it to every conversation rather than a dedicated GPT.

All four platforms do the same basic thing: save the instruction block once, then reuse it without retyping. The output quality will vary slightly by model, since each one interprets the pattern list a little differently, which is itself a small demonstration of why a single detection checklist has limits.

What should I check for besides the tells?

Everything above catches what's wrong with a draft. It has nothing to say about what's missing. That's a different job, and it's the one HumanizeAI's H.E.A.R.T. framework is built for: Human First, Evidence Over Claims, Answer First, Real Voice, and Trust Signals Throughout.

Here's a second copy-paste block built from those five principles. Where the first block flags patterns to remove, this one flags gaps to fill: no direct answer up front, no named sources, no real example, no internal or external links. Run it against a draft after the tell checker, and you've covered both halves of the problem instead of just one.

COPY-PASTE BLOCK – H.E.A.R.T. Addition Checklist

You are reviewing a piece of writing for what it is missing, based on
HumanizeAI's H.E.A.R.T. framework (Human First, Evidence Over Claims,
Answer First, Real Voice, Trust Signals Throughout). Read the text I
provide and check it against each principle below. For any principle
that fails, tell me exactly what is missing and suggest a specific fix,
not a general note to "add more detail."

1. HUMAN FIRST -- Is there a clear, specific reader this was written for?
   Would that specific person find every paragraph useful, or does any
   section read like it was written for a search engine instead of a
   person?

2. EVIDENCE OVER CLAIMS -- Is every statistic, finding, or claim
   attributed to a named source with a date? Flag any "studies show,"
   "experts say," or unsourced statistic as a gap, not just a style issue.

3. ANSWER FIRST -- Does the piece (and does each major section) give its
   core answer in the first 150-200 words, or does it build up to the
   point instead of leading with it?

4. REAL VOICE -- Is there at least one specific, concrete, firsthand
   example, scenario, or observation that only someone who has actually
   done this could have written? Generic examples that could apply to
   any company do not count.

5. TRUST SIGNALS THROUGHOUT -- Is there a named author with real
   credentials, at least one outbound link to a credible external source,
   and at least one internal link to related content? Flag any missing
   piece.

For each principle, tell me: PASS or GAP, and if GAP, exactly what to add
and where. Do not suggest removing anything. This check is only about
what is missing, not what should be cut.

Here is the text to review:

[PASTE YOUR DRAFT HERE]

Set this block up the same way as the tell checker above: a Claude Project or Skill, a Gemini Gem, a Perplexity Space, or a ChatGPT Custom GPT. Many people build both as separate skills so they can run them back to back, subtraction first, then addition.

Can I use H.E.A.R.T. itself as a writing prompt, not just a checker?

Yes, and this is the more useful version for anyone drafting new content rather than auditing an existing draft. Instead of asking an AI assistant to review a finished piece, you can hand it the framework up front and have it write with H.E.A.R.T. built in from the first draft.

COPY-PASTE BLOCK -- H.E.A.R.T. Writing Prompt

When you write for me, follow HumanizeAI's H.E.A.R.T. framework:

HUMAN FIRST -- Write for [describe your specific reader here, e.g. "a
content marketing manager at a 50-person B2B SaaS company"], not for a
generic audience or a search engine. If a sentence would confuse or bore
that person, cut it.

EVIDENCE OVER CLAIMS -- Every statistic, finding, or claim needs a named
source, a date, and a real link. If you do not have a verifiable source
for a claim, either flag it clearly as [VERIFY: needs a source] or leave
it out. Never write "studies show" or "experts say" without a name
attached.

ANSWER FIRST -- Lead with the direct answer in the first 150-200 words.
Background and context come after the answer, not before it. Every
section should open with its most important point, not a warmup.

REAL VOICE -- Write like someone with real, specific experience doing
this, not like a summary of what other people have written about it.
Use contractions. Keep paragraphs to 2-4 sentences. No stock phrases like
"in today's fast-paced world," "it's important to note," or "in
conclusion."

TRUST SIGNALS -- Where relevant, note where a named source, a real
example, or a link would strengthen the piece, even if you cannot add
the link yourself.

Confirm you understand these five principles before we start, then apply
them to everything you write for me in this conversation.

Paste that into a fresh conversation, a Claude Project, or a Custom GPT before you start drafting, and it sets the ground rules before a single word gets written, rather than trying to fix them after the fact.

HumanizeAI Framework References

This checklist is a subtraction tool by design, and that's fine for what it is: a fast, free way to catch surface-level AI patterns in a draft. It maps to the "Real Voice" principle in HumanizeAI's H.E.A.R.T. framework, which calls for zero AI-tell constructions in published work. Where it stops is addition: it has no way to check whether a claim is sourced, whether an example is real, or whether the piece says anything new. That's the "Evidence Over Claims" half of H.E.A.R.T., and it's a different job than pattern-matching.

Founder Observation

I ran our own contrarian article on this exact guide through a version of this checklist before we published it. It caught four instances of negative parallelism, the "it's not X, it's Y" pattern, that I hadn't consciously noticed while writing, because the whole piece was built around a contrast argument. The checklist did exactly what it's supposed to do: it caught a real pattern. It also couldn't tell me whether the argument itself was any good. That's the distinction this whole list is missing when people treat it as a writing guide instead of a spell-checker. I really like the "it's not X, it's Y" pattern. I think it is really useful.

Research & Supporting Evidence

• Wikipedia:Signs of AI writing – the primary guide, maintained by WikiProject AI Cleanup, cataloging the patterns listed above with the explicit caveat that no single sign proves AI authorship.
• Jodie Cook, "The 10 Giveaway Signs Of AI Writing, Wikipedia Reveals," Forbes (September 2025) -- a plain-language breakdown of the guide's main categories with examples.
• Nathan Yau, "Signs of AI writing on Wikipedia," FlowingData (October 2025) -- covers the guide's own caveat that the list is not a ban on words or punctuation and that most human writers use these patterns too, just less often than raw AI output.

Key Takeaways

• Wikipedia's "Signs of AI Writing" guide catalogs about a dozen recurring patterns in raw AI-generated text, built for editors catching undisclosed AI content, not for teaching anyone to write.
• No single pattern proves a piece was AI-written. The guide is meant to be read in combination, not as a single tripwire.
• The copy-paste block above turns the list into reusable instructions you can run against any draft.
• The same block works as the instructions for a Claude Project or Skill, a Gemini Gem, a Perplexity Space, or a ChatGPT Custom GPT, with setup steps that take a few minutes on each platform.
• A pattern checklist can catch surface-level tells. It can't verify a claim, confirm a real example, or tell you whether the writing says anything worth reading. That's a separate job.
• A second copy-paste block, built from HumanizeAI's H.E.A.R.T. framework, checks for what the tell checker can't: missing sources, missing examples, missing direct answers, and missing trust signals.
• H.E.A.R.T. also works as a standalone writing prompt, set up before you draft rather than applied after, so new content starts with the addition principles built in.

FAQ

What is Wikipedia's Signs of AI Writing guide? It's a public reference maintained by Wikipedia's WikiProject AI Cleanup, cataloging recurring patterns found in undisclosed AI-generated text, such as formulaic em dash use, rule-of-three lists, and vague attribution. It was built to help Wikipedia editors catch AI content, not as a general writing style guide.

Can I use Wikipedia's AI writing guide to check my own content? Yes. The guide's patterns translate well outside Wikipedia. The copy-paste instruction block in this article lets you run the same checklist against any draft using Claude, Gemini, Perplexity, or ChatGPT.

How do I turn a prompt into a Claude Skill or Gemini Gem? In Claude, paste the instructions into a Project's custom instructions field, or into a new Skill if you have Skills access. In Gemini, create a new Gem under the Gems section and paste the instructions into its setup field. Both save the instructions for one-click reuse without retyping them each time.

Does removing every sign on this list guarantee my content sounds human? No. The list catches surface patterns, not substance. Content can have zero em dashes, zero triplets, and zero stock phrases and still be thin, unsourced, or generic. Removing tells is a different task from adding real evidence and experience to the writing.

Is Wikipedia's guide the same as an AI detector? No. It's a manual pattern reference, not an automated detection tool, and Wikipedia's editors are explicit that it doesn't prove AI authorship on its own. Automated AI detectors have their own accuracy problems separate from this guide.

What is the H.E.A.R.T. framework and how is it different from the tell checker? H.E.A.R.T. is HumanizeAI's writing framework: Human First, Evidence Over Claims, Answer First, Real Voice, and Trust Signals Throughout. Where the Wikipedia-based tell checker flags patterns to remove, H.E.A.R.T. flags gaps to fill, like missing sources, missing examples, or no direct answer up front. Used together, they cover both what to cut and what to add.

What HumanizeAI's Fine-Tuned Model Checks Beyond This List

The copy-paste block above is a free, general-purpose pattern check anyone can run in any AI assistant. Our own fine-tuned model goes further, and we want to be direct about what's live today versus what we're actively building, rather than blur the line.

Live today: our model runs the same category of pattern check as this guide, tuned specifically to HumanizeAI's brand-voice rules, alongside a stat-verification pass that checks every factual claim for a named source, a date, and a live link, flagging anything that can't be verified rather than letting it publish unresolved. It also checks for genuine firsthand experience signals in a piece, with a hard rule against fabricating an example just to pass the check.

In active development: a citability layer that asks a different question than any pattern checklist can answer, whether a piece would actually get cited by another writer or an AI system over what's already published on the same topic, plus a freshness layer that checks whether a piece is still accurate and current rather than quietly going stale after it publishes. Both are part of our current build, not shipped features yet, and we'll say so plainly here until they are.

Additional Resources

• How to Humanize AI Text: The Complete Guide for Marketers
• Wikipedia Wrote the Best Guide to Spotting AI Writing. Here's Why Following It Will Make Your Content Worse.
• Make AI Generated Text More Human
• Answer Engine Optimization Playbook

About the Author

Steve Palomares has spent 25+ years building software companies. Now owner of HumanizeAI, he writes about AI content strategy for marketing, AEO, GEO and growing software businesses with AI. Based in North Texas.