Wikipedia’s AI Writing Guide Is the Best Prompt Engineering Resource Nobody’s Talking About

Wikipedia editors are quietly winning the fight against AI-generated content, and they’ve published a detailed playbook showing exactly how. Their guide, “Signs of AI Writing,” was built by WikiProject AI Cleanup over thousands of hours reviewing AI-submitted articles. It runs nearly 15,000 words. And while it was written for editors trying to catch undisclosed AI text, it’s become one of the most useful resources available for anyone trying to get AI to write more like a human.

Here’s the thing: if you understand what gives AI away, you can tell your AI exactly what not to do.

This article breaks down the key patterns Wikipedia identified, what they actually mean for your writing, and the exact prompts you can use to train Claude, ChatGPT, or Gemini to stop doing them.

The Important Caveat First

Before the list, Wikipedia is clear about one thing: no single sign proves something was AI-written. Humans use em dashes. Humans write in threes. Humans say “crucial” sometimes.

The problem is cumulation. When you see three or four of these patterns stacked in the same paragraph, the probability shifts. The guide is a pattern detector, not a checklist. Keep that in mind as you read this, and as you edit your own AI-assisted drafts.

The Patterns Wikipedia Identified

1. The Banned Word List (and It Changes by Model)

This is the one everyone knows about. Certain words show up in AI output at rates wildly above what human writers produce. Wikipedia’s editors have tracked how the list has evolved across model generations:

GPT-4 era (2023 to mid-2024): delve, intricate, tapestry, pivotal, underscore, meticulous, testament, vibrant, garner, bolstered, interplay, landscape.

GPT-4o era (mid-2024 to mid-2025): pivotal, fostering, crucial, align with, emphasizing, highlighting, showcasing, enhance, enduring, vibrant.

GPT-5 era (mid-2025 onward): emphasizing, enhance, highlighting, showcasing.

Grok is its own category. It leans toward pseudoscientific vocabulary: causal, empirical, correlate, and continues to overuse underscore as of 2026.

The practical takeaway is that “delve” becoming a meme in 2023 actually worked. OpenAI tuned it out. The patterns shift with each new model, but they never fully disappear. New tells emerge to replace old ones.

Also worth noting: Wikipedia is specific that a word being overused by AI does not mean its synonyms are overused too. The problem is the exact word, in the exact pattern, at an unnatural frequency.

2. The Em Dash Situation

This one has caused more arguments than almost anything else in writing circles. AI uses em dashes far more than most human writers do, and more specifically, it uses them in places where a comma or parenthesis would be more natural.

A sentence like “The startup raised $5 million — its first institutional round — from a UAE-based fund” is a reasonable use of em dashes. The problem is AI drops them into nearly every other sentence as a way of adding emphasis or inserting a subordinate clause. At that frequency, it reads as mechanical rather than stylistic.

Some AI companies have reportedly started suppressing em dash usage in newer models because the pattern became so widely known. But the underlying tendency remains: AI reaches for em dashes when a human writer would use simpler punctuation.

The fix is straightforward. Replace em dashes with commas or parentheses wherever the clause can handle it. Keep em dashes for genuinely emphatic breaks where nothing else works as well.

3. The “Not X, But Y” Structure

AI loves contrast-reframe sentences. “It’s not a product. It’s a movement.” “It’s not a setback. It’s an opportunity.” “This isn’t about speed. It’s about precision.”

Human writers use this construction occasionally, because it works well for emphasis. AI uses it constantly because it produces the appearance of depth without requiring any. The Wikipedia guide calls these “negative parallelisms,” and they are among the most recognizable patterns in AI-generated content precisely because they pile up quickly within a single piece.

If you find yourself reading a paragraph where two or three consecutive sentences follow this pattern, the writing almost certainly came from an AI with no editing afterward.

4. The Rule of Three

AI defaults to groupings of three almost everywhere: adjectives, benefits, examples, takeaways. “Innovative, transformative, and groundbreaking.” “The platform is fast, reliable, and secure.” Lists with exactly three bullets. Paragraphs with three supporting points.

This isn’t wrong on its own. The rule of three is a legitimate rhetorical device that predates AI by centuries. The problem is that AI applies it mechanically, in every paragraph, whether three items are the right number or not. Human writers sometimes give two examples. Sometimes five. The number matches the content, not a formatting preference.

When you notice that every list in a piece has exactly three items, you’re probably reading unedited AI output.

5. Formatting Overkill

AI has a compulsive relationship with structure. Bold lead-in phrases followed by colons and a sentence restating the bold text. Bullet points where a paragraph would read more naturally. Excessive bolding of terms throughout the body text as if every piece is a textbook.

The bolded-term-then-colon structure is described as a near-signature of ChatGPT output. It barely exists in natural human writing. Real editors bold text when it needs emphasis, not as a default formatting mode for every concept introduced.

AI also skips en dashes almost entirely, using hyphens for date ranges and scores where an en dash belongs. This is a technical tell that copy editors notice immediately.

6. Filler Transition Words at Unnatural Rates

“Furthermore.” “Moreover.” “Additionally.” “It is worth noting that.” “No discussion would be complete without.”

These words show up in AI text at three to five times the rate of human writing. They’re connective tissue that AI uses to stitch paragraphs together as it builds text sequentially. The problem is that human writers use transitions more sparingly, and they use them because the logic of the argument requires it, not as a default way to begin a new paragraph.

“It is important to note” and its variants are particularly telling. AI inserts editorial commentary about what matters without being asked, as if it cannot present information without also explaining to the reader that this information is significant.

7. The Treadmill Effect

This one is subtle but important. AI text often covers the same ground repeatedly without actually advancing the argument. A 500-word section might contain 100 words of new information and 400 words of restatement and padding. The writing moves, but it doesn’t go anywhere.

Human writing has a much higher information-per-word ratio. Good writers say a thing once, then move on. AI circles back, restates, qualifies, and elaborates in ways that look thorough but read as hollow once you’re aware of the pattern.

8. Hyperbole Without Evidence

Wikipedia editors noticed a consistent pattern of AI treating ordinary things as extraordinary. A regional statistics office described as representing “a pivotal moment in the evolution of regional statistics.” A minor product update framed as a paradigm shift. Adjective inflation where everything is “groundbreaking,” “comprehensive,” “holistic,” or “transformative” regardless of whether the subject warrants it.

Real writers back claims with evidence. AI reaches for adjectives when it doesn’t have specific facts to offer. If a paragraph is full of superlatives but light on numbers, quotes, or concrete examples, that’s the pattern in action.

What This Means for How You Use AI

Understanding these patterns gives you two practical advantages.

First, you can spot them in your own AI-assisted drafts before they go out. Run through the list mentally when reviewing any AI-generated content. Are there em dashes everywhere? Is every list exactly three items? Are there contrast-reframe sentences stacking up? Does the piece circle back and restate things it already said?

Second, you can instruct your AI to avoid them from the start, which is faster than fixing them afterward.

The Prompts: How to Train Your AI to Write More Like a Human

Here are prompts you can use with Claude, ChatGPT, or Gemini to address each of the main patterns. The key principle, as one writing coach noted, is not to try removing every tell at once. Too many restrictions produce stiff, generic output that sounds artificial in a different way. Focus on the patterns most relevant to your voice and content type.

Master System Prompt (Use This First)

This is a comprehensive instruction you can paste at the start of any conversation or save as a custom instruction. It addresses the most common patterns without overcrowding the AI’s output.

Write in a natural, human voice. Follow these rules strictly:

Avoid these words entirely: delve, tapestry, pivotal, meticulous, underscore, vibrant, foster, testament, enhance, crucial, intricate, landscape, bolstered, garner, showcasing, highlighting, emphasizing, realm, multifaceted, robust, seamless, holistic, comprehensive, transformative, groundbreaking, leverage, harness, empower, elevate, embark, navigate.

Do not use em dashes (—). Use commas or parentheses instead.

Do not use the “not X, but Y” sentence structure. Say what something is, not what it is not.

Do not group everything in threes. Lists can have two items or four or five. Match the number to the content.

Do not use filler transitions at the start of paragraphs: no “Furthermore,” “Moreover,” “Additionally,” “It is worth noting,” or “It is important to remember.”

Do not bold terms throughout body text as if every concept needs highlighting. Bold only when something genuinely requires emphasis.

Do not repeat ideas already stated. Make a point once and move forward.

Vary sentence length. Mix short sentences with longer ones. Do not make every sentence the same structure.

Back claims with specific facts, numbers, or examples. Avoid adjectives that are not earned by evidence.

How to Add This to Each Platform

Claude: Paste it at the start of your conversation before your actual request. For ongoing use, go to Settings, then User Preferences, and add it there so it applies to every new conversation automatically.

ChatGPT: Go to Settings, then Personalization, then Custom Instructions. Paste the prompt into the “How would you like ChatGPT to respond?” field. It will apply across all your conversations.

Gemini: Open the Gems section under the Gemini menu. Create a new Gem, give it a name like “Human Writer Mode,” and paste the instructions into the setup field. Select that Gem whenever you want these rules active.

Targeted Prompts for Specific Problems

If you only need to fix one or two patterns rather than all of them, these targeted prompts are faster to use.

To fix word choice:

Rewrite this text without using any of these words: delve, pivotal, tapestry, meticulous, underscore, foster, testament, enhance, crucial, vibrant, intricate, landscape, multifaceted, robust, seamless, holistic, transformative, groundbreaking. Replace them with simpler, more direct alternatives. Do not add new AI vocabulary to compensate.

To fix sentence structure:

Rewrite this text to vary sentence length throughout. Some sentences should be short, under ten words. Others can be longer. Remove any “not X, but Y” contrast structures. Make each sentence a different shape from the one before it where possible.

To fix formatting habits:

Rewrite this text as flowing prose. Remove all em dashes and replace them with commas or parentheses. Remove any bold text from the body paragraphs. If there are bullet point lists, convert at least half of them into prose sentences. Where lists remain, make sure they do not all have exactly three items.

To fix filler and padding:

Read through this text and remove every sentence that repeats an idea already stated. Cut all filler transition phrases: “Furthermore,” “Moreover,” “Additionally,” “It is worth noting,” “It is important to remember,” “No discussion would be complete without.” If a paragraph’s first sentence is removed and the paragraph still makes sense, remove it.

To fix hyperbole:

Review every adjective in this text. For each adjective like “groundbreaking,” “transformative,” “revolutionary,” “comprehensive,” or “holistic,” either replace it with a specific fact that earns the description, or delete it entirely.

The Review Prompt (Use This After Drafting)

Once you have a draft, use this to catch remaining patterns before publishing.

Read this text and identify any of the following: words from common AI vocabulary lists, em dashes used where a comma would work, “not X, but Y” sentence structures, lists that all have exactly three items, filler transition words starting paragraphs, bolded terms throughout body text, repeated ideas or padding, and adjectives without supporting evidence. List each instance with the line it appears in, then suggest a specific fix for each one.

A Note on Evolving Patterns

The Wikipedia guide is a living document, updated regularly as models change. What gave away GPT-4 is not what gives away GPT-5. The specific banned words shift with each generation. Em dash usage has reportedly decreased in newer models because the pattern became too well known.

This means the prompts above will need updating over time. The underlying philosophy, writing with specificity, varying structure, cutting repetition, and earning adjectives with evidence, will not change, because these are the qualities of good human writing regardless of era. The specific vocabulary tells will shift, but the structural and rhetorical patterns are more durable because they reflect how language models are trained, not just the quirks of one particular model version.

The most reliable long-term strategy is to develop an eye for the patterns yourself. Read your AI drafts critically. Ask whether each sentence is doing something new. Check whether the transitions are necessary or just there. Count how many times a list has exactly three items. That editorial eye, trained on Wikipedia’s framework, will serve you better than any single prompt.

The Bigger Picture

Wikipedia’s guide was built to protect encyclopedia quality, not to improve marketing copy or startup blog posts. But the patterns it documents are not platform-specific. They appear wherever AI generates text without careful human editing.

For founders, marketers, journalists, and content creators using AI as part of their workflow, the guide represents a practical quality control framework. If your content reads like Wikipedia’s examples of AI writing, it will register as inauthentic to readers even if they cannot name exactly why. Trust erodes subtly. Engagement drops. The audience moves on.

The goal is not to hide that you used AI. The goal is to produce writing worth reading, and right now, that still requires understanding what AI does by default and steering it in a different direction.

Wikipedia’s editors figured out what to look for. Now you know what to fix.

The original Wikipedia guide, “Signs of AI Writing,” is maintained by WikiProject AI Cleanup and is updated regularly as new model patterns emerge. It is publicly available at en.wikipedia.org.