Signs of AI Writing: A 10-Point Checklist for Your Draft (2026)

The most common signs of AI writing are stock words (crucial, pivotal, showcasing), grand claims about "significance", "not just X, but Y" sentences, lists of three, vague attributions like "experts say", bold and emoji used as decoration, and one even, upbeat tone with no real opinion. No single sign proves anything, and humans write this way too. Use the checklist below to edit your own AI-assisted draft, not to judge someone else's.

Where this list comes from

Wikipedia's volunteer WikiProject AI Cleanup maintains a field guide to the habits of chatbot text, built from real examples. It is the most detailed public list, and its own caveats shape how to use it:

• The signs are observations, not rules. Not all text with these features is AI-generated, because chatbots learned from human writing, including blogs and editorials.
• The guide warns against fixing only the surface signs, because the deeper problems (accuracy, sourcing, empty claims) are what matter.
• The vocabulary moves. "Delve" was famously overused in 2023 and early 2024, then dropped off sharply in 2025, so any list of "AI words" is dated within a model or two.

So treat what follows as a checklist for quality. If a sentence trips it, the sentence is probably vague, and vague is worth fixing whoever wrote it.

A 10-point checklist of AI writing signs

Points 1, 2, 5 and 7 are the ones that carry the least information. That is why cutting them improves the writing even if nobody ever asks who wrote it.

A worked before-and-after

This example was written for this post, with made-up numbers. It is an illustration, not output from any tool, and no detector was run on it.

The AI draft:

In today's fast-paced digital landscape, effective communication is crucial for team success. Our new weekly review isn't just a meeting, it's a game-changer that fosters collaboration, boosts accountability, and drives innovation. By leveraging clear agendas, teams can unlock their full potential.

After an edit:

Our Monday review used to run 50 minutes and end with no owners. Now everyone brings one number and one blocker, and we stop at 25 minutes. Fewer things slip, because every action item has a name next to it.

What changed, in checklist terms:

1. The opener (point 1 and 2) went. It said nothing a reader didn't already know.
2. "Isn't just a meeting, it's a game-changer" (point 5) became two plain facts.
3. Three abstract benefits (point 6) became one concrete change.
4. "Leveraging clear agendas to unlock potential" (point 1) became what the team actually does.

Notice what an editor cannot invent: the 50 minutes, the owners, the "one number and one blocker". Those details have to come from you. A tool can reshape the sentences, but it cannot know how your team's meeting works, which is the reason to add those specifics yourself.

Are em dashes a sign of AI writing?

They were a strong tell, and they are a weaker one now. Wikipedia's guide lists em dash overuse as an indicator for roughly 2022 to September 2026: chatbots used them where a person would use a comma, colon or parentheses. It also says OpenAI changed GPT-5.1 to follow em dash instructions, and that the newest models are being tuned to avoid them. The guide cites a July 2026 analysis in The Economist finding that, among current models, only Claude used em dashes more than professional writers.

The practical read: one em dash proves nothing, because careful human writers use them. The sentence around it is what to check. If it is a punchy "this, not that" line, apply point 5. And swapping every dash for a comma while leaving the rest alone fixes nothing.

Can people reliably spot AI writing?

Not by default. A 2025 study by Russell, Karpinska and Iyyer found that annotators who often use chatbots for writing were very good at it: a majority vote of five of them misclassified only 1 of 300 articles. They leaned on specific vocabulary cues and on subtler qualities like originality and clarity. Wikipedia's guide, which cites that study alongside others, says readers who rarely use chatbots did only slightly better than chance.

Two lessons follow. If you edit a lot of AI text, your instincts are probably decent, but they are still instincts. And if you are a teacher, editor or manager, a style hunch is not a finding. Non-native English writers are flagged more often by detectors (see Liang et al.), and formal or formulaic human writing trips the same signs. If you are on the receiving end of a suspicion, here is how to find out which detector your school uses and what its score means.

Doing the edit with Unflag

Do this on your own draft, where AI help is allowed. Check your school's or employer's AI policy first, and disclose AI help where the policy asks for it.

1. Rewrite the draft. Open Rewrite, pick Improve Writing (or Simplify or Make Shorter if it is long-winded, or the Email or LinkedIn format for a post or message), and choose a tone such as Direct or Casual. It rebuilds sentences from the meaning rather than swapping synonyms, and it keeps the language you wrote in.
2. Humanize AI when the whole draft reads like a model wrote it. Humanize AI works on the full text with one of 18 tones and needs at least 50 words on the web. It tries to keep figures such as prices and percentages, but check every number yourself. The difference between a paraphraser and a humanizer explains which to start with.
3. Walk the checklist on the result and add the specifics only you know. This is the step that makes it yours.
4. Fact Check anything with claims in it. Fact Check is a Plus feature: it lists each claim with a verdict and sources. See how to fact-check an AI answer.
5. Optionally, run the AI Detector as a self-check. It shows an AI score, an AI-generated vs human-generated split and a verdict, and it does not predict what any other detector will say. Here is how detectors like Pangram work.

On the free plan you can run up to 600 characters per input, with a few free uses each day, and you sign in first on the web. Unflag Plus raises the limit to 10,000 characters and adds Fact Check. As of October 2026, Plus is $8.99 a month or $45.99 a year on the web; the apps are billed weekly or yearly through the stores. On a phone, get the app for iPhone or Android.

Will removing these signs get my text past an AI detector?

Nobody can promise that, and you should distrust any tool that does. Detectors use different methods, update often, and flag human writing too. Unflag's own site says it cannot guarantee that any output will pass a third-party detector. The better goal is a draft that is specific, accurate and reads like you, which is also the one that holds up when someone reads it closely.

When this checklist is not the right tool

If you only need typos fixed, a grammar checker is enough. If you are judging someone else's work, the signs are a reason to ask a question, not to reach a verdict: ask for drafts or version history and talk to them. And if a text is full of invented facts or fake citations, no amount of style editing helps. Check the claims first.

How we put this together

We read Wikipedia's "Signs of AI writing" guide (October 2026), the Russell, Karpinska and Iyyer study and the Liang et al. paper on detector bias, and checked Unflag's behavior against its own tools and limits. The example paragraphs were written for this post and have no detector scores attached.

The short version: the signs are habits, not proof. Cut the filler, replace the vague with the specific, check every claim, and keep a copy of your drafts.

Frequently asked questions

What are the most common signs of AI writing?

Stock vocabulary such as crucial, pivotal and showcasing, grand claims about significance, 'not just X, but Y' sentences, lists of three, vague attributions like 'experts say', bold text and emoji in headings, and one even, upbeat tone from start to finish. Chatbot leftovers such as 'I hope this helps' or stray Markdown asterisks are the clearest tell.

Are em dashes a sign of AI writing?

They were a common tell, but a weak one now. Wikipedia's guide says em dash overuse marked AI text from about 2022 to September 2026, and that newer models are being tuned to avoid them. Plenty of people write with em dashes too. Treat one as a prompt to read closely, never as proof.

Can you tell if something was written by AI just by reading it?

Sometimes, but not reliably. A 2025 study found five annotators who often use chatbots for writing misclassified only 1 of 300 articles by majority vote. Wikipedia's guide adds that readers who rarely use them did only slightly better than chance. Style alone is a hint, not evidence, so never accuse someone on it.

How do I make an AI draft sound like me?

Read it aloud, mark the stock phrases and inflated claims, and replace each with something specific only you know: a number, a name, a decision, an opinion. Cut anything that restates the previous sentence. Then vary sentence length. A rewrite tool can speed up the first pass, but your details make the draft yours.

Will fixing these signs make my text pass an AI detector?

No tool or edit can guarantee that. Detectors disagree with each other, change their models and sometimes flag human writing. Unflag's own site says it cannot guarantee any output will pass a third-party detector. Edit for readers, check the facts, and follow your school or employer's AI policy, including any disclosure rule.

Is Unflag free to use for editing a draft?

There is a free plan: Rewrite and Humanize AI take up to 600 characters per input, with a few free uses a day, and you sign in on the web. Unflag Plus raises the limit to 10,000 characters and adds Fact Check. As of October 2026, Plus is $8.99 a month or $45.99 a year on the web.