Wikipedia's signs of AI writing: checklist and checker
Wikipedia's editors keep the most detailed public list of how AI-generated text gives itself away. Here's a summary as of November 2026, what changed this year, and a checker that flags the text-level signs as you paste.
New in 2026
- Em dashes are now a "historical indicator". The section is titled "Overuse of em dashes (2022–September 2026)". It cites a July 2026 study that found only Claude still used them more than professional writers. Our own November test agrees for the ChatGPT app but not for the API: ChatGPT's default model used 0.75 em dashes per 1,000 words, but the GPT-5.5 to GPT-6.1 API models used 3.8 to 9.2 (human answers to the same questions: 0.21).
- Corrective disclaimers (mid-2026 on). Chatbots now add framings like "should be treated as speculative rather than definitive", "the available information does not establish…", or "should not be confused with…". The guide says these were almost unheard of before the GPT-5.6 era (July 2026).
- A GPT-6 vocabulary list. challenge, clearer, dependable, echoed, foster, leverage, matters, multifaceted, practical, prioritize, quietly, steady, universally, along with more periods and commas and short sentences.
The checklist
Content
- Undue emphasis on significance: "a pivotal moment", "an enduring legacy", "reflects broader trends".
- Undue emphasis on notability: listing media coverage or "online presence" to prove importance.
- Superficial analysis: "-ing" clauses that comment instead of inform ("…, highlighting its importance").
- Vague attributions: "experts argue", "observers have noted", with no one named.
- Outline-like conclusions: "Challenges and future prospects" paragraphs that say nothing specific.
Language
- AI vocabulary, by era. 2023–mid-2024: delve, tapestry, testament, intricate, pivotal, underscore, vibrant. Mid-2024–mid-2025: align with, enhance, fostering, showcasing. Mid-2025–mid-2026: emphasizing, highlighting. Mid-2026 on: see the GPT-6 list above.
- Avoiding "is" and "are": "serves as", "stands as", "boasts", "features".
- Negative parallelisms: "not just X, but also Y", "not X, but Y", "Y rather than X". Our page on "it's not X, it's Y" has counts by model.
- Rule of three: adjective, adjective, adjective; three short phrases for rhythm.
Style and communication
- Messages meant for the user: "Certainly!", "I hope this helps", "If you'd like, I can…".
- Hedging about sources: knowledge-cutoff notes, and in 2026 the corrective disclaimers above.
- Placeholders: "[Date]", "INSERT URL", unfilled templates.
- Markdown in plain text: **bold labels**, "###" headings, bullet stacks.
Citations (needs a human)
- Broken or invented links, invalid DOIs and ISBNs, DOIs that resolve to unrelated papers, book citations without page numbers,
utm_source=chatgpt.comin URLs.
What isn't a reliable sign
The guide warns that AI detectors are unreliable, that good grammar or a formal tone alone means nothing, and that any one sign is weak on its own. Look for clusters. That's how the checker above scores, too.
Prompt version
Paste this into ChatGPT, Claude or Gemini before your text:
Edit the text below so it does not show the signs in Wikipedia's "Signs of AI writing" guide:
- No inflated significance ("stands as a testament", "pivotal moment", "enduring legacy").
- Replace AI vocabulary (crucial, delve, fostering, showcase, underscore, vibrant, multifaceted, leverage, quietly) with plain words, or cut it.
- Use "is"/"are"/"has" instead of "serves as", "stands as", "boasts", "features".
- No negative parallelisms: "not just X, but Y", "it's not X, it's Y", "Y rather than X".
- No reflexive lists of three. No vague attributions ("experts say", "observers note").
- No disclaimers about how a claim "should be treated". No offers ("If you'd like, I can…").
- No bold run-in labels or bullet lists unless the content is a real list.
Keep every fact, name and number. Do not add new claims.
TEXT:
For agents (Claude Code, Codex, Cursor), the same rules come as a file: the free humanizer skill. To rewrite finished text with a before/after score, use the unslop tool.
Free: 3 AI rewrites a day (up to 300 words) and unlimited quick clean. Need more? A $5 pack of 50 rewrites never expires, or go Pro at $12/month · $79/year. See plans.
FAQ
What is Wikipedia's Signs of AI writing page?
A guide written by Wikipedia editors (WikiProject AI Cleanup) listing patterns they see in AI-generated edits: AI vocabulary, negative parallelisms, rule of three, promotional tone, chatbot correspondence, broken citations and more. It's a project page at en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing.
Are em dashes still a sign of AI writing on Wikipedia?
As of November 2026 the guide lists "Overuse of em dashes (2022–September 2026)" under historical indicators. The guide cites a July 2026 study that found only Claude still used them more than professional writers. Our own November test is more mixed: ChatGPT's default model rarely uses them, but GPT-5.5 to GPT-6.1 via the API still do.
What changed in the guide in 2026?
Three things stand out: a GPT-6-era vocabulary list (challenge, clearer, dependable, echoed, foster, leverage, matters, multifaceted, practical, prioritize, quietly, steady, universally), a new mid-2026 pattern of corrective disclaimers ("should be treated as X rather than Y"), and em dashes moving to the historical section.
Can I check my text against the list automatically?
Yes. The checker on this page highlights the signs that can be matched in text (vocabulary, negative parallelisms, rule of three, copula avoidance, chatbot offers, disclaimers, markdown) in your browser. Signs that need context, like fake citations, still need a human.
Is there a prompt or skill version?
Yes. The prompt on this page and the free humanizer skill (gounslop.com/humanizer-skill) turn the guide into instructions for ChatGPT, Claude, Claude Code and Codex.