Signs of AI Writing: How to Spot and Fix Them
AI Writing Tips

Signs of AI Writing: How to Spot and Fix Them

All articlesWriteHumanly TeamNovember 2, 20268 min read

A plain-English guide to the signs of AI writing, from not-X-but-Y contrasts to chatbot leftovers, with before and after fixes and the limits of spotting AI.

If you read a lot of chatbot output, you can usually feel it before you can name it. The draft is tidy and confident, and it says less than its length suggests. This guide names the patterns behind that feeling, with a short made-up before and after for each, so you can catch them in your own drafts.

Our main source is Wikipedia's Signs of AI writing, a field guide kept by the volunteer editors of WikiProject AI Cleanup. Their page is written for encyclopedia articles, so what follows is our own plain-English summary for everyday writing. We also draw on blader/humanizer, an MIT-licensed editing checklist built on that guide, with more than 63,000 stars on GitHub. If you'd rather skip the reading, paste a draft into our free Signs of AI Writing checker and it will mark these patterns for you.

Why AI drafts share the same habits

A language model writes whatever is most likely to come next. As the humanizer checklist puts it, the model's default is the choice that suits the widest range of readers, while a human writes for one reader about one thing. Every pattern below is a version of that safe, general choice.

Structural tells

Start with these. The checklist puts them first because word habits change with each model release while sentence and paragraph shapes last much longer. A find-and-replace pass can't fix them.

1. The contrast nobody asked for

Wikipedia calls this a negative parallelism. The sentence knocks down a claim no one made, so the real point sounds bigger by comparison. Watch for "not just X, but Y," "it's not X, it's Y," and the reversed form, "Y rather than X."

Before: "This isn't just a budgeting app. It's a whole new way to think about money."

After: "The app sorts your spending into categories each week and flags any bill that went up."

Keep a contrast when your reader actually believes the thing you're correcting. Otherwise, state the point directly.

2. One-line closers

A short line tacked onto a paragraph that repeats it with more drama, such as "Let that sink in" or "That's the real win." It asks the reader to pause on nothing new.

Before: "We cut page load time from four seconds to one. Bounce rate fell by a third. That's the real win."

After: "We cut page load time from four seconds to one, and bounce rate fell by a third."

Cut the closer. If it adds a fact or a consequence the paragraph didn't already show, fold it into a normal sentence.

3. Staged run-ups

Phrases that announce a point instead of making it: "Let's dive in," "Here's the thing," "Here's what you need to know," or a lone "Honestly?" at the start of a paragraph. They add a beat of suspense and no information.

Before: "So, what makes a cover letter work? Let's dive in. Here's the thing: hiring managers skim."

After: "Hiring managers skim, so put your strongest qualification in the first two lines."

4. Arguing with no one

This one rebuts an objection the reader never raised. Typical openers are "To be clear," "Don't get me wrong," and "I'm not saying X, but."

Before: "Don't get me wrong, meetings have their place. To be clear, I'm not against them. But most could be an email."

After: "Most of our weekly meetings could be an email."

5. Forced groups of three

Wikipedia's guide calls this the rule of three. Adjectives, phrases and examples arrive in threes as a rhythm, whether or not the writer has three things to say.

Before: "Our team is passionate, dedicated, and innovative, bringing creativity, expertise, and vision to every project."

After: "Our team has built 40 client websites since 2022, most of them for local restaurants."

The fix needs a fact only you have. If you don't have one, a shorter, plainer sentence beats a padded one.

6. Inflated significance

This is the first item in the content section of Wikipedia's guide. An ordinary fact gets tied to legacy, broader trends or a turning point. Look for "plays a crucial role," "a testament to," "the evolving landscape" and "marking a pivotal moment." A close relative is brochure language, where every town is "nestled" somewhere and every view is "breathtaking."

Before: "The bakery's new sourdough stands as a testament to the town's enduring culinary heritage."

After: "The bakery started selling sourdough in March. It sells out most Saturdays by ten."

7. Vague experts

The guide's term is vague attribution: "experts say," "studies show," "observers have noted," with no name and no link. It borrows authority the writer can't point to.

Before: "Experts agree that remote work boosts productivity."

After: "In our own team survey, 14 of 20 people said they get more done at home."

Name the source and link it. If you can't, drop the claim or present it as your own view.

8. Bold-label bullet lists

Wikipedia's style section flags what it calls inline-header lists, where every bullet opens with a bolded phrase and a colon. It also flags heavy bold in running prose and headings written in Title Case. One labeled list is fine. A page made of them reads like a slide deck.

Before: three bullets reading "Speed: Faster load times," "Security: Better protection" and "Support: Help when you need it."

After: "Pages load in about a second, logins use two-factor codes, and support answers email within a day."

9. Chatbot leftovers

Wikipedia files these under communication meant for the user: words addressed to the person who typed the prompt, pasted in by accident. Think "Certainly! Here's a revised version," "I hope this helps!" and "Let me know if you'd like a different tone," plus disclaimers like "as of my last update." The guide also lists unfilled placeholder text, such as a bracketed name or date the writer forgot to replace.

Before: "Certainly! Here is a polished version of your bio. Jane is a nurse with ten years of ICU experience. I hope this helps!"

After: "Jane is a nurse with ten years of ICU experience."

Leftovers come closest to a giveaway, since a person writing for their own readers has no reason to talk to a prompt.

Word-level tells

The guide treats a high density of what it calls AI vocabulary as a sign. The usual suspects include "delve," "tapestry," "intricate," "underscore," "showcase" and "vibrant." A related habit is dodging plain "is" and "are" with "serves as," "stands as" or "boasts."

Before: "The museum serves as a vibrant hub that showcases the intricate tapestry of local history."

After: "The museum covers local history, and it's busy most weekends."

Density is what counts. Any of these words can show up in careful human writing, and the favorites shift as models change. Our list of words ChatGPT overuses has plain swaps for the common ones, though swapping words won't help much if the sentence still makes one of the structural moves above.

Dashes

For a few years the long em dash was the best-known tell of all. Models used it where most people would use a comma, a colon or brackets, often with a space on each side. Our post on why AI uses em dashes covers the theories.

Before (we've typed the dashes as double hyphens): "The launch went well -- better than expected -- and the team -- finally -- got a weekend off."

After: "The launch went better than expected, and the team finally got a weekend off."

The picture is changing. Wikipedia's editors have moved dash overuse into their historical signs, dated 2022 to September 2026, noting reports that the newest models avoid the mark entirely. The page suggests that leaning almost only on commas and periods may be the newer pattern, though that line is tagged as needing a citation. If you just want stray dashes, curly quotes and hidden characters out of a draft, our free AI text cleaner does that in one paste.

What these signs can't tell you

Wikipedia's guide opens with its caveats. The signs describe common patterns and prove nothing, and because models learned from human writing, plenty of people write this way on their own. The guide also lists indicators that don't work, such as perfect grammar and bland or formal prose.

Judging by feel is also less reliable than it seems. In a 2025 ACL paper by Russell, Karpinska and Iyyer, annotators labeled 300 nonfiction articles as human or AI. Five people who often use chatbots for writing did very well: their majority vote got only one article wrong. The full paper also tested four people who rarely use them, and they did about as well as chance. They caught 56.7% of the AI articles but also flagged roughly half of the human ones, while rating their own confidence highly. Software isn't a safe fallback either: Wikipedia's guide warns that detectors such as GPTZero and Pangram have real error rates.

So never accuse someone of using AI from a checklist, a detector score or a hunch. Wikipedia counts being able to explain your own editing choices as a sign of human writing, so if you have a concern, ask the writer about the work.

How to use this list on your own drafts

Read the draft aloud and mark the structural tells first. Then rewrite each flagged paragraph around its main point instead of patching one phrase at a time, and check that every fact and number survived. Word swaps and punctuation come last. The Signs of AI Writing checker can do the marking for you.

If you started from an AI-assisted draft, our humanizer reworks it to read naturally in your voice while keeping the facts intact. Check your school's or employer's AI policy before you use any AI tool on work you'll hand in.

Frequently asked questions

What are the most common signs of AI writing?

Contrasts that knock down claims nobody made, dramatic one-line closers, staged run-ups, inflated importance, unnamed experts, bold-label lists and leftover chatbot lines. Clusters of words like "delve" are a weaker sign.

How can you tell if something was written by AI?

From the text alone, you usually can't be sure. A cluster of these patterns makes AI help more likely, but people write them too. Asking the writer about their choices tells you far more.

Are em dashes a sign of AI writing?

They were, but Wikipedia's guide now lists dash overuse as a historical sign because newer models reportedly avoid them. Plenty of human writers have always used dashes.

Will removing these patterns get my text past AI detectors?

Don't count on it. The humanizer project's own README says detectors still flag most of its output. Cut these patterns for the people reading your work instead.

Written by

WriteHumanly Team

The team behind WriteHumanly has spent thousands of hours studying how AI detectors actually score text, building tools used by students and professionals worldwide. We publish what we learn so other writers can make better decisions.

Make your draft sound like you wrote it

Paste your text and get a natural rewrite in seconds, with every fact kept. Free, no account needed.

Try WriteHumanly free

Related Articles