Wikipedia
Drafts and new articles
Editors tag suspect pages with the AI-generated template and focus on hallucinated facts and fake sources more than on style.
SPOTTING AI TEXT
Signs of AI writing: 24 tells, the data, and why none is proof
A reader's field guide to text that might come from a chatbot. Every sign is grouped by level, rated by how much it tells you, and tied to the research or the Wikipedia editors who documented it.
Scan my text free Updated
chatgpt.com
Today
Write a two line bio for our bakery's website.
Message ChatGPT
MacBook Pro
9:41
ClaudeCleanLocal scanPASTED FROM CHATGPT
SCANNING
…AI tells
nestled in the heart ofbrochure phrase1
testament tosignificance puffery1
fosteringtrailing -ing clause1
One tell proves nothing. A cluster of them is what readers notice.
9:41
ClaudeCleanLocal scanPASTED FROM CHATGPT
SCANNING
…AI tells
nestled in the heart ofbrochure phrase1
testament tosignificance puffery1
fosteringtrailing -ing clause1
One tell proves nothing. A cluster of them is what readers notice.
Short answer
The clearest signs of AI writing are chatbot leftovers: lines like “Certainly! Here is”, unfilled placeholders, citation tokens such as oaicite, and links ending in utm_source=chatgpt.com. Fabricated references come next. Clusters of AI-favored words (delve, underscore, tapestry), trailing -ing commentary and formula endings are moderate signs. Single words, em dashes and a formal tone prove very little, and both people and detectors misjudge human text often.
THE CATALOGUE
What are the signs of AI writing?
Twenty-four signs in six levels, from single words up to fake references. Try the paragraph first, then read the full list. Each example is a line we wrote to show the pattern.
TRY IT FIRST
Spot the tells in this paragraph
A short bio we wrote the way a chatbot often does. Click any dotted phrase to see which sign it is and how much it proves.
garden-bio.docx, example we wrote
Certainly! Here is a short bio for the garden page. Nestled in the heart of Riverside, the Elm Street Garden stands as a testament to community spirit, fostering connection and pride. It's not just a garden. It's a living classroom. Our workshops are inclusive, hands-on, and transformative. Experts argue that shared gardens improve wellbeing. turn0search0 **Key takeaway:** every plot counts. Founded in [YEAR], the garden welcomes everyone. Visit elmstreetgarden.org/?utm_source=chatgpt.com
SIGNS FOUND
0 of 11
Click a dotted phrase to inspect it.
- Strong
- Moderate
- Weak alone
01
Word choice
#1AI vocabulary in clusters
ModerateAdditionally, this intricate interplay underscores a pivotal shift in the evolving landscape.
Words on Wikipedia's watch list such as delve, underscore, tapestry and pivotal. One is noise; many, in text written after 2022, is a strong tell.
#2Significance puffery
ModerateThe festival stands as a testament to the town's enduring legacy.
Small subjects inflated into symbols: a testament, a pivotal moment, setting the stage for, an indelible mark.
#3Brochure language
ModerateNestled in the heart of the valley, the village boasts breathtaking natural beauty.
Travel-ad wording in a text that should be neutral: nestled, vibrant, renowned, boasts, a diverse array.
#4Unnamed authorities
ModerateExperts argue that the policy has drawn significant attention from observers.
Industry reports, observers and some critics who are never named or linked. Check whether a real source follows.
02
Sentence patterns
#5Trailing -ing commentary
ModerateThe team shipped the update in March, highlighting its commitment to users.
A participle bolted onto a fact to add meaning. PNAS researchers measured these clauses at 2 to 5 times the human rate in instruction-tuned models.
#6Negative parallelism
Weak aloneIt's not just a tool, but a new way of working.
Found in 6% of ChatGPT chats in July 2025 in the Washington Post's analysis. Speechwriters love it too.
#7Rule of three
Weak aloneThe app is fast, reliable, and secure.
Triplets of adjectives or short phrases, used by reflex. Human writers use them constantly, so they only count in a crowd.
#8Avoiding plain is and has
Weak aloneThe library serves as a hub and features a reading room.
Serves as, stands as, represents, boasts and features where is or has would do. Wikipedia counts plain phrasing as a human sign.
03
Tone and content
#9Chat leftovers
StrongCertainly! Here is a revised version of your paragraph. I hope this helps!
Lines addressed to the person who asked the chatbot, pasted along with the answer. Few innocent explanations exist.
#10Knowledge and source hedges
StrongWhile specific details are limited in available sources, as of my last knowledge update...
The model explaining what it does not know. Knowledge-cutoff lines are older (2022 to 2023) but still turn up.
#11Placeholder text
StrongFounded in [YEAR] by [Founder Name], the company... Published 2025-XX-XX.
Template slots nobody filled in, like [Describe the specific section] or PASTE_URL_HERE.
#12Formula endings
ModerateDespite these challenges, the town's future looks bright.
A closing paragraph about challenges, legacy or future outlook, attached to topics that never needed one.
04
Formatting
#13Markdown in plain text
Moderate**Benefits:** faster load times. ## Overview
Asterisks and pound signs that a chat window turns into bold and headings, left raw in an email, form or wiki.
#14Bold-header bullet lists
Weak aloneSpeed: pages load faster. / Cost: plans start lower. / Support: help is always on.
Every bullet opens with a bolded label and a colon. Common in slide decks too.
#15Title Case and scattered bold
Weak aloneKey Takeaways And Next Steps
Headings in Title Case and bolded phrases mid-paragraph. Many style guides call for both, so treat it lightly.
#16Emoji bullets
Moderate✅ Faster onboarding ✅ Better retention 🚀 Ready to scale
By July 2025, 70% of ChatGPT messages in the Post's sample had an emoji, and the green check appeared 11 times more often than in human writing.
05
Punctuation and characters
#17Em dashes
Weak aloneThe plan was simple — ship fast, learn faster.
A famous tell that is fading: OpenAI made ChatGPT follow no-dash instructions in November 2025. More in why AI uses em dashes.
#18Curly quotes
Weak aloneShe called it “a turning point” in the report.
Wikipedia notes ChatGPT and DeepSeek typically use curly quotes while Gemini and Claude typically do not. Word and macOS add them automatically.
#19Hidden Unicode characters
Weak aloneA zero-width space between two words, or a no-break space after a comma.
Copying from chat apps, web pages and word processors can carry them. See the invisible character detector guide; they are not a watermark either, as this comparison explains.
#20Chatbot citation tokens
Strong:contentReference[oaicite:16]{index=16} turn0search0 [cite: 1] [web:1]
Internal markup from ChatGPT, Gemini and Perplexity that leaks when an answer is copied. Wikipedia editors log each new format.
06
Citations and links
#21utm_source=chatgpt.com
Strongexample.org/report?utm_source=chatgpt.com
ChatGPT tags links it cites with this tracking parameter, so the URL records where it came from.
#22Broken identifiers
StrongISBN 978-1-23-456789-4 (checksum fails); a DOI that opens an unrelated paper.
Invented references. One public database lists over 2,000 legal decisions involving AI-hallucinated content.
#23Sources that do not say it
ModerateA book cited with no page number for a very specific claim.
Real sources attached to claims they never make. Only opening the source catches this.
#24Bulk dead links
ModerateFive of eight reference links return 404 and have no archived copy.
Links that never existed leave no trace in the Internet Archive. Old links that rotted usually do.
Most of these patterns come from Signs of AI writing, the guide Wikipedia editors keep for spotting undisclosed chatbot text. The strength ratings are ours, based on how often each sign shows up in human writing too. If you are the one writing and want to remove these habits, the Claude humanizer guide covers the editing side.
WEIGH THE EVIDENCE
How strong is each sign of AI writing?
A chat leftover and an em dash are both called signs, but they are not close in weight. Start at the top of the ladder.
STRENGTH OF EVIDENCE
Not every sign is worth the same
Pick a rung. Higher rungs are harder to explain any other way; lower rungs show up in human writing every day.
Strong
- “Certainly! Here is…” or “I hope this helps”
- [Insert date] style placeholders
- oaicite, turn0search0, [cite: 1]
- ?utm_source=chatgpt.com in a link
- “As of my last knowledge update…”
Leftovers from a chat window. There are very few innocent reasons for a person to type these into a finished article or essay.
What to do: Treat as high confidence that a chatbot touched the text. Still ask how it was used: drafting, translating and copyediting are different things.
| Sign type | What it looks like | Human false positives | Best check |
|---|---|---|---|
| Chat leftovers and tokensAct first | Certainly!, [YEAR], oaicite, utm_source=chatgpt.com | Very low | Search the text for the strings |
| Fabricated referencesAct first | Dead DOIs, failing ISBNs, quotes no one said | Low (people mistype, rarely invent) | Open every source |
| Vocabulary clusters | Many watch-list words in one passage | Medium | Count per paragraph, compare with older work |
| Sentence patterns | -ing tails, not just X but Y, lists of three | High alone | Only count alongside stronger signs |
| Formatting | Raw asterisks, emoji bullets, bold labels | Medium to high | Ask where the text was drafted |
| Punctuation and tone | Em dashes, curly quotes, perfect grammar | Very high | Ignore unless other signs agree |
Swipe to compare →
Highlighted rows are the signs we would act on first. Risk levels are our summary of Wikipedia's guidance and the false positive studies cited further down.
CHECK WHAT YOU CANNOT SEE
Some signs are invisible. Show them.
Paste a suspect text and ClaudeClean lists hidden characters like zero-width spaces and no-break spaces, scanned in your browser. It is not an AI detector and gives no verdict. It shows you what is there.
- 300 rewrite credits a month
- 20,000 characters per request
- .txt, .docx and .pdf files
AI WORDS LIST
Which words does AI overuse? What the data says
Lists of AI-sounding words are everywhere. These are the ones backed by measurements on millions of documents.
DATA · 15.1 MILLION ABSTRACTS
How far a few words jumped after ChatGPT
Kobak and colleagues compared 2024 word use in PubMed abstracts with what the pre-2023 trend predicted. A ratio of 1x means no change.
EXCESS RATIO, 2024 (OBSERVED ÷ EXPECTED)
- “delves”28.0x28.0x expected
- “underscores”13.8x13.8x expected
- “showcasing”10.7x10.7x expected
0x10x20x30x
dashed line = 1x, the expected level
WORDS WITH EXCESS USE THAT YEAR
up to 190
454
20212024
The pandemic changed science vocabulary through topic words. The 2024 jump came from style words, and it was more than twice as wide.
| Measure | Value |
|---|---|
| delves, 2024 ratio | 28 |
| underscores, 2024 ratio | 13.8 |
| showcasing, 2024 ratio | 10.7 |
| Excess words in 2021 | 190 |
| Excess words in 2024 | 454 |
100x+
How much more often GPT-4o used tapestry, palpable and camaraderie than humans
WIKIPEDIA WATCH LIST, GROUPED
- Verbs
- delve, underscore, highlight, showcase, foster, garner, bolster, emphasize, enhance, align with
- Adjectives
- pivotal, crucial, intricate, meticulous, robust, vibrant, enduring, valuable, key
- Nouns
- tapestry, testament, interplay, landscape (abstract), deep dive
- Openers
- Additionally, at the start of a sentence or paragraph
Kobak and colleagues counted 454 words used far more than expected in 2024 abstracts, against up to 190 in the COVID year of 2021. Most of the 2024 words were style words rather than topic words: 66% verbs and 14% adjectives, per the paper.
Word favorites move. The Washington Post found delve in about 1 in 1,000 ChatGPT chats by July 2025, while core rose fivefold and modern showed up in more than 8% of messages. A list built in 2024 already misses the newer habits.
Why these words? A COLING 2025 study found its results consistent with human feedback training playing a part. Nobody has shown it is the only cause.
HOW THE TELLS CHANGED
AI writing patterns from 2022 to 2026
Signs age. The ones that gave away 2023 text are not the ones that give away text today.
Nov 30, 2022
ChatGPT launches
Wikipedia uses this date as a cutoff: edits made before it can be ruled out as chatbot text.2022 to 2024
Cutoff disclaimers and lectures
Lines like “As of my last knowledge update” and “it's important to note” were common. Wikipedia now lists both as historical signs.Apr 2023
Detectors meet non-native writers
Stanford researchers show seven detectors flag 61% of TOEFL essays as AI on average (Liang et al.).Feb 2025
Citation tokens leak into articles
Around this time, editors start logging ChatGPT markup like oaicite and turn0search0 in pasted text. DeepSeek and Perplexity formats were logged later in 2025.Jul 2025
Delve, measured
Science Advances publishes the excess vocabulary study of 15.1 million abstracts (Kobak et al.). In ChatGPT itself, delve has already faded.Nov 2025
The em dash gets a switch
OpenAI makes ChatGPT follow no-em-dash instructions (Ars Technica), and TechCrunch calls Wikipedia's list the best guide to spotting AI writing.Mar 2026
Wikipedia bans LLM-written articles
The WP:LLM guideline prohibits using LLMs to generate or rewrite article content, with exceptions for translation and basic copyediting (WikiProject AI Cleanup).Jul 2026
Model habits split
The Economist reports that only Claude used em dashes more than professional writers, while ChatGPT used them less.Sep 2026
Em dashes become a historical sign
Wikipedia's guide dates em dash overuse to 2022 through September 2026, as newer models increasingly avoid the character.
WHERE YOU WILL SEE THEM
Signs of AI writing in the wild
The same signs look different on an encyclopedia page, in a feed and in a classroom.
WIKIPEDIA · AI CLEANUP
How an editor reads a suspicious draft
A made-up draft about a made-up bakery. Watch the order: vocabulary first, then the leftovers that are much harder to explain away, then the tag.
en.wikipedia.org
W
AI CleanupDraft:Harbor Lane Bakery
From Wikipedia, the free encyclopedia
This article may incorporate text from a large language model, which is prohibited in Wikipedia articles.
Harbor Lane Bakery is a family-owned bakery nestled in the heart of a small harbor town. Founded in 1998, it boasts a rich tradition of sourdough baking and stands as a testament to the town's enduring food heritage, highlighting the pivotal role of local shops in a vibrant community. :contentReference[oaicite:2]{index=2}
Challenges and Legacy
Despite its success, the bakery faces several challenges, including rising flour costs and limited parking.
References
- “Harbor Lane turns 25.” Harbor Gazette. harbor-gazette.example/2023/bakery?utm_source=chatgpt.com
- Smith, J. Coastal Baking Traditions. ISBN 978-1-23-456789-4
1/4 · Reading the draft
MacBook Pro
en.wikipedia.org
W
AI CleanupDraft:Harbor Lane Bakery
From Wikipedia, the free encyclopedia
This article may incorporate text from a large language model, which is prohibited in Wikipedia articles.
Harbor Lane Bakery is a family-owned bakery nestled in the heart of a small harbor town. Founded in 1998, it boasts a rich tradition of sourdough baking and stands as a testament to the town's enduring food heritage, highlighting the pivotal role of local shops in a vibrant community. :contentReference[oaicite:2]{index=2}
Challenges and Legacy
Despite its success, the bakery faces several challenges, including rising flour costs and limited parking.
References
- “Harbor Lane turns 25.” Harbor Gazette. harbor-gazette.example/2023/bakery?utm_source=chatgpt.com
- Smith, J. Coastal Baking Traditions. ISBN 978-1-23-456789-4
1/4 · Reading the draft
SOCIAL FEED
Five signs in one work post
A generic post we wrote in the style that fills professional feeds. The pins show up one by one; the legend says how much each one is worth.
9:41
Search
A. Example
Product lead · 2h
One year of leading a product team. A few reflections.
Leadership is not just about managing tasks, but about empowering people.
Here are 3 lessons I learned:
- ✅ **Listen first:** the best ideas came from the team.
- ✅ **Say it twice:** clarity beats cleverness.
- ✅ **Celebrate wins:** small ones count.
The result? A team that is more aligned, more engaged, and more resilient.
What lessons have you learned? Let me know in the comments!
Like Comment Repost Send
- 1
Emoji bullets
In the Washington Post's sample of ChatGPT messages, 70% had at least one emoji by July 2025, and ✅ showed up 11 times more often than in human writing.
- 2
“Not just X, but Y”
The same analysis found this frame in 6% of July 2025 chats. People use it too, so it counts only in company.
- 3
Raw **asterisks**
Bold headers typed as Markdown. The chat window renders them; a post box or email often shows the stars.
- 4
Rule of three
Three parallel items where one specific example would do. Weak alone, common in speeches and ads.
- 5
Chat-style closer
“Let me know” and offers of more help are on Wikipedia's list of collaborative phrases left over from a chat reply.
Count them together. Any one of these appears in plenty of human posts; all five in eight lines is what makes a reader stop.
Work
Posts, emails and cover letters
Formatting signs (emoji bullets, bold labels, raw asterisks) and chat closers stand out most here, since nobody checks references.
School
Essays and assignments
Fake citations matter most. If you are the student under suspicion, read how to prove you didn't use AI and what Turnitin can and cannot detect.
Reviews
Product and place reviews
Brochure language with no specifics (no dish named, no date, no price) is the giveaway in reviews, more than any single word.
THE LIMITS
Why none of these signs is proof
People who use chatbots a lot get good at spotting them. Everyone else, and most software, makes a lot of mistakes.
61%
Average rate at which seven detectors labeled non-native English essays as AI
97.8%
Of those human essays flagged by at least one of the seven detectors
72%
Of human-written texts that untrained educator raters labeled as AI
1 of 300
Articles misjudged by the majority vote of five experienced LLM users
Worth weighing
- Chat leftovers. Greeting lines, placeholders, citation tokens and chatgpt.com tracking links.
- Fabricated sources. Identifiers that fail, quotes nobody said, pages that do not exist.
- Dense clusters. Several levels of signs in one short passage, in text written after 2022.
- A sudden shift. Vocabulary or structure that does not match the writer's earlier work.
Not evidence on its own
- Perfect grammar. Careful writers and spell checkers produce it every day.
- A formal or bland tone. Wikipedia lists robotic-sounding prose as an ineffective indicator.
- Transition words. However, moreover and additionally appear in decades of human writing.
- A detector score. Wikipedia says the tools beat chance but have non-trivial error rates.
Simple English is not a sign
The writers most likely to be wrongly flagged are those writing in a second language, because their word choices are more predictable. In the Liang study, the same essays were flagged far less after ChatGPT was used to make their vocabulary more varied. Treat any sign as a reason to ask, never as a finding.
QUICK ANSWERS
Common questions about spotting AI text
Short answers to the questions people ask right after reading a list like this.
Do AI detectors work?
Partly. Wikipedia says tools like GPTZero and Pangram “perform better than random chance” but tells editors not to rely on them alone. Their errors fall hardest on non-native writers.
Would a watermark settle the question?
Only for the company that set it. Text watermarks such as Google's SynthID sit in word choices and are read with the provider's key, so readers cannot see them. The AI text watermarking guide covers who can check what, and this page covers Claude.
Is delve still a sign of AI?
Less than it was. It jumped 28-fold in 2024 science abstracts, then ChatGPT largely dropped it by mid-2025. Today it mostly marks text from 2023 and 2024 models, or people imitating them.
Can AI text show none of these signs?
Yes. A careful prompt plus a human edit can remove most of them, and newer models already avoid several. The signs catch lazy pasting far more reliably than deliberate use.
They can be disguised, but it will be hard to do away with them completely.
THE READER'S CHECKLIST
How to tell if something is written by AI, step by step
Eight checks, cheapest and strongest first. Stop when you have enough to ask a fair question.
- 01
Check the date first
Text published before November 30, 2022, when ChatGPT launched, is very unlikely to be chatbot output. Version history and archive snapshots help. - 02
Look for chat leftovers
Search for Certainly, I hope this helps, as of my last, oaicite, turn0search, [cite and utm_source=chatgpt.com. One hit outweighs a page of style signs. - 03
Open the references
Click every link, look up every DOI and test every ISBN. Fabricated or mismatched sources are the most serious finding on this page. - 04
Count clusters, not words
A single delve means nothing. Ten watch-list words, three -ing tails and a challenges paragraph in 300 words is a pattern. - 05
Look for first-hand detail
Names, numbers, dates, places and small specifics only the writer would know are hard for a model to invent convincingly. - 06
Compare with earlier work
A sudden shift in vocabulary, structure or polish from the same writer is a stronger signal than any list. - 07
Ask the writer about it
People can explain why they made a choice, where a fact came from and what they left out. Keep the question neutral. - 08
Weigh it, do not convict
Formal tone and simple vocabulary describe many second-language writers. Never act on a detector score or a style hunch alone.
ON THE OTHER SIDE
If you wrote it and it reads like AI
This page is for readers. Writers who see their own habits here have a different job.
Style
Edit the habits out
The Claude humanizer guide has before and after rewrites and a ten-pass edit checklist.
Dashes
Decide on em dashes
Read why AI uses em dashes, then replace them case by case.
Accused
Keep your evidence
Drafts, version history and notes do more than any rewrite. See how to prove you didn't use AI.
CLAUDECLEAN
See the characters, then clean up your own draft.
Paste text to find hidden Unicode in your browser. For your own writing, an optional rewrite with a non-Claude model removes common AI habits and shows a diff. No detector result is promised.
01
> source text
Paste the text
Up to 2,000 characters without an account, or a .txt, .docx or .pdf file on a plan. ClaudeClean doesn’t store what you paste.
02
scan report
3 zero-width1 soft hyphencleanable signals2See what is hidden
Zero-width spaces, no-break spaces, soft hyphens and lookalike letters are listed and cleaned locally, before anything leaves your browser.
03
before → after
clearusefulanswerfocusedhelpfulresponse
Copied ✓Rewrite your own draft
Optional: up to three OpenAI passes that cut em dashes, hedges and summary closers, with every change shown in a diff for you to review.
PRICING
Clean the text free. Rewrite more with Pro.
The scan and hidden-character cleanup are free. Pro adds room for rewrites and files.
PRO
CLAUDECLEAN PRO
ClaudeClean Pro
$5/ month
Cancel anytime.
- 300 credits every month (≈300,000 characters of rewriting)
- 20,000 characters per request
- Clean .txt, .docx and .pdf files
- Light / medium / strong intensity
- Dashboard with history
Checking account…
Free: 10 credits a month with an account, or try 2,000 characters right here without one. Strong rewrites cost 2× credits.
SOURCES
Research behind this guide to signs of AI writing
Numbers on this page come from these. The example sentences and strength ratings are ours.
Wikipedia
Signs of AI writing
The editor-maintained catalogue of words, patterns, markup and citation problems, plus its list of ineffective indicators.
View sourceScience Advances
Delving into LLM-assisted writing in biomedical publications through excess vocabulary
Kobak et al., July 2025: 15.1 million PubMed abstracts, 13.5% lower bound for 2024, delves at 28 times expected.
View sourcePNAS
Do LLMs write like humans?
Reinhart et al., February 2025: participial clauses at 2 to 5 times the human rate, tapestry in 23% of GPT-4o outputs.
View sourcePatterns
GPT detectors are biased against non-native English writers
Liang et al., 2023: 61.22% average false positive rate on 91 TOEFL essays across seven detectors.
View sourceACL 2025
People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text
Russell, Karpinska and Iyyer: expert majority vote misjudged 1 of 300 articles, beating most detectors.
View sourceWashington Post
What 328,744 ChatGPT messages reveal about its style
November 2025 analysis of gpt-4o chats from May 2024 to July 2025: emoji in 70%, not just X but Y in 6%, delve fading.
View sourceAlso cited: the educator study by Cheng et al. (19% overall accuracy), Juzek and Ward on why ChatGPT says delve, TechCrunch on Wikipedia's guide, Ars Technica on ChatGPT and em dashes, The Economist on model differences, WikiProject AI Cleanup, the AI-generated template and the AI Hallucination Cases database. The Washington Post analysis is also available via The Boston Globe. Want to check a text for hidden characters? Open the cleaner.
FAQ
Signs of AI writing FAQ.
The strongest are chatbot leftovers (Certainly! Here is..., placeholder brackets, citation tokens like oaicite, links ending in utm_source=chatgpt.com) and fabricated references. Next come dense clusters of AI-favored words, trailing -ing commentary, puffery and formula endings. Em dashes, curly quotes and formal tone prove little.
Check the date, search for chat leftovers, verify the sources, then look for clusters of signs rather than single words. Compare with the writer's earlier work and ask them about their choices. No method, human or software, gives certainty.
Studies of 2024 text found big jumps in delves (28 times the expected rate in PubMed abstracts), underscores and showcasing, and GPT-4o used tapestry over 100 times more than humans. Word favorites shift between model versions, so lists go stale.
Nobody knows for sure. A COLING 2025 study found its tests consistent with human feedback training playing a role. The word has also faded: by July 2025 it appeared in about 1 in 1,000 ChatGPT chats, per a Washington Post analysis.
Weakly at best. Writers used em dashes long before chatbots, OpenAI made ChatGPT follow no-dash instructions in November 2025, and Wikipedia's guide now files em dash overuse under historical signs.
They do better than chance but make errors. A Stanford-led study found seven detectors flagged 61% of essays by non-native English writers as AI on average. Wikipedia tells its editors not to rely on detectors alone.
It is the most complete public catalogue, built by editors who review AI-generated drafts, and TechCrunch called it the best guide available. Wikipedia itself warns that text with these signs is not always AI-generated.
Yes. Formal tone, perfect grammar and common transitions all appear in human writing, and in one study untrained raters labeled 72% of human texts as AI. Version history and drafts are the best defense.