AI tells: a writer's checklist of the signs, with examples, and how to avoid them in 2026

Updated 23 October 2026. This checklist now also answers whether people can tell you use AI: the three levels at which readers notice, when they care and when they do not, and why AI-assisted posts land differently from AI-drafted ones.

AI tells are the habits that make text read as machine-written: stock words like delve and leverage, “not just this, but that” contrasts, lists of three, evenly sized paragraphs, generic closing questions and, in older AI output, a heavy sprinkle of em dashes. No single tell proves a person used AI; readers react to the cluster. To avoid them, ban the stock words, vary your rhythm, add one detail only you know, and run a two-minute scan before you publish. Whether readers can actually tell, and when it matters, is covered further down.

This is the draft-time checklist. The read-it-in-the-wild version, for spotting AI writing in other people’s posts, is 9 signs of AI writing. All before-and-after examples below are written for this article to show each pattern and its fix; none are quotes from real writers.

The AI tells list at a glance

Nine common AI tells, what each looks like, and the fix (examples written for this article).

What Wikipedia’s “Signs of AI writing” guide says

The most cited public list of AI tells is Wikipedia:Signs of AI writing, an advice page kept by WikiProject AI Cleanup, a volunteer group of Wikipedia editors. It is worth reading because it is careful about what the tells can and cannot prove. Five points from it (read 22 October 2026) change how you should use any checklist, including this one:

• It is descriptive, not a rulebook. The page says its list “consists of observations, not rules”, and that not all text with these signs is AI-generated. It is also not an official Wikipedia policy.

• The favorite words change with each model generation. It notes that “delve” was famously overused by ChatGPT in 2023 and early 2024 and became less frequent later. For models from mid-2026 on, it lists words such as quietly, steady, matters, practical and prioritize, plus more full stops and commas and shorter sentences.

• Em dash overuse is now filed under historical indicators, dated 2022 to September 2026. Cutting stray em dashes is still good hygiene, but the em dash is no longer the giveaway it was.

• Structure tells last longer than word tells. It describes models overusing the rule of three and parallel constructions built on “not”, “but” or “however”, the same shapes as tells 3 and 4 below.

• Some popular “tells” do not work. It lists perfect grammar, formal or “fancy” prose, mixed registers and bland prose as unreliable signs on their own, because plenty of people write that way.

The practical lesson: a word list goes stale every few months, but the structural habits (even rhythm, symmetric contrasts, tidy triplets, generic closes, nothing specific) stay. Build your checklist around those.

Tell 1: The stock vocabulary cluster

What it is: three or more words from the AI-favored cluster in one post. The long-running core: leverage as a verb, delve, unlock, navigate, harness, foster, elevate, embark, robust, seamless, comprehensive, holistic, plus openers like “in today’s fast-paced world” and bridges like moreover and furthermore. Newer models add quieter words, as above.

How to avoid it: ban the cluster and use the plain word. Leverage becomes use. Delve becomes look at. Unlock becomes enable. Navigate becomes handle. Foster becomes build. Elevate becomes improve. Robust becomes strong. Seamless becomes smooth. The full substitution table is in the words AI overuses.

Before: “We leveraged the new framework to navigate the comprehensive onboarding redesign and unlock a more seamless user experience.” After: “We used the new framework to rebuild onboarding. It now takes two screens instead of five.”

Tell 2: The symmetric two-part hook

What it is: “Most people think one thing. The reality is another.” And its variants: “It’s not about the tactic, it’s about the system.” “Forget the hacks, focus on the basics.” General models reach for this shape when asked for something engaging.

How to avoid it: open with a specific observation, a named context, a concrete confession or a plain claim. Before: “Most people think founder content is about credentials. The reality is it’s about voice.” After: “I read 200 founder posts last week. The ones I remembered each had one detail only that founder could know.” For more openers that work, see how to write hooks.

Tell 3: The “not just this, but that” frame

What it is: “It is not just about marketing, it is about culture.” Fine once. A fingerprint when it repeats. Models use it to add apparent depth to a flat point.

How to avoid it: use it at most once per post, and only when it narrows to a sharper claim. Before: “This is not just about marketing, it is about culture. It is not just about culture, it is about leadership.” After: “This is about who owns the work: whether leadership treats marketing as a result of the culture or as the thing that shapes it.”

Tell 4: The rule of three

What it is: everything comes in threes. Three adjectives, three short phrases, three bullet points. Models use triplets to make thin analysis look complete.

How to avoid it: write the number of items you actually have. If you have two good reasons, give two. If one adjective does the job, use one. Before: “The update is faster, smarter, and more intuitive.” After: “The update loads in half the time. That’s the whole change.”

Tell 5: The beige bullet middle

What it is: four or five evenly weighted bullets that could appear in any post on any topic: “Consistency is the multiplier. Quality beats quantity. Trust the process.”

How to avoid it: every bullet must be specific to your work. The test: could a stranger reading the bullet alone tell which writer it came from? If not, cut or rewrite it. Before: “Show up every day. Trust the process.” After: “In my third week of daily posting, replies dropped by half. They came back in week five, and I still do not know why.”

Tell 6: The generic closing question

What it is: “What’s your take?”, “Drop your thoughts below”, “Save this for later”. The sign is the genericness, not the call to action itself.

How to avoid it: end on the last sentence of your argument, ask a specific question you actually want answered, or just stop on a strong line. Before: “What’s your take? Let me know below!” After: “If you have kept a daily posting habit past week three, I would like to know what changed for you.”

Tell 7: Even paragraph and sentence rhythm

What it is: every paragraph the same length, every sentence the same length, every paragraph the same shape. Human writing is uneven by default.

How to avoid it: vary length on purpose. A one-line paragraph. Then a longer one that builds, qualifies and lands. Then a fragment. Readers pick up the shape of a page before they read a word.

Tell 8: Voice-flat coherence

What it is: writing that is fluent, on topic, consistent and forgettable. Nothing marks it as yours. This is the tell behind what people now call AI slop, and the hardest to self-diagnose.

How to avoid it: put at least one sentence in every post that only you could have written: a named example from your work, a phrase people know you for, an opinion a general model would soften. Before: “Effective marketing requires a deep understanding of customer needs.” After: “Our best sales insight always came on the third call with the same prospect, never the first.”

Tell 9: Em dash density

What it is: two or more em dashes in a short paragraph. This was the best-known tell of 2024 and 2025, and it still turns up in plenty of AI-edited text, but the Wikipedia guide now treats it as a historical indicator because newer models lean on full stops and commas instead.

How to avoid it: if dashes are not part of your natural style, replace them with a full stop, a colon, a comma or brackets, whichever matches the job the dash was doing. If they are part of your style, keep the ones that do real work. Plenty of good writers have always used them, which is why the dash alone was never proof.

AI tells in creative writing and fiction

The same checklist carries over to stories, fanfiction and creative essays, with a few genre-specific shapes worth watching for in your own drafts:

• Characters who explain their feelings outright instead of showing them through action or dialogue.

• Every scene ending on a reflective summary line that tells the reader what the scene meant.

• Description in tidy triplets: “the air was cold, still, and heavy.”

• Dialogue where every character speaks in the same polished register.

• Even scene and paragraph lengths, with no sudden short beats.

None of these proves anything on its own; good human writers do each of them sometimes. The fix is the same as for posts: specifics, uneven rhythm, and a voice that refuses the safe default.

AI tells to avoid: the two-minute pre-publish scan

• Vocabulary scan: zero words from your banned list, including the newer ones.

• Hook check: not the two-part contrast, not a credentials opener, not a framework count with generic items.

• “Not just this, but that” count: zero or one.

• Triplet check: any list of three that should really be two or four.

• Bullet check: every bullet specific to your work.

• Close check: no generic question; end on a point, a specific question or nothing.

• Rhythm scan: visible variety in paragraph and sentence length.

• Voice check: point to one sentence only you could have written. If you cannot, add one.

• Dash check: em dashes only where they are genuinely your style.

How to check a draft for AI tells with a tool

If you want a second pair of eyes, our free AI detector highlights the words and patterns in your text that read as AI and gives a human fix for each, rather than a single “percent AI” score. The humanize AI text tool rewrites flagged lines toward a natural voice. Scoring detectors such as GPTZero and Originality.ai return a probability instead; our comparison of the best AI detectors in 2026 covers how they differ and why every one of them misfires on some human writing.

For the audience side (who notices these tells and whether they care), see the next section. For the workflow this checklist plugs into, see the hybrid human-AI writing workflow, and for a line-by-line edit of an AI draft, how to make AI writing sound human.

Can people tell you use AI?

Some can, more half-notice, and many never do. That makes “can people tell if you use AI” three questions in one: can a reader consciously label a post as AI-written, can they sense that an account reads generic without naming why, and do they feel a voice getting flatter over months? The answers differ, and so does the damage.

Three levels of noticing

Nobody can measure the size of each group with any rigor (platforms do not survey this, and self-reported detection is unreliable), so this is directional:

1. Explicit detectors. A small group reads closely and has read enough AI text to label it on sight: the stock words, the two-part hook, the beige bullets. They tend to be writers, editors, marketers and heavy platform users, which is to say a high-value part of your audience.

2. Implicit pattern-matchers. A larger group never labels a post “AI”. They file the account as interchangeable rather than recognizable, then engage less, follow less and forget faster. They will not tell you why.

3. The unaware. Readers who take each post on its own merits. This is the group behind the claim that “most people can’t tell”. The claim is right about them and wrong when stretched to everyone.

The first two groups overlap heavily with the people who subscribe, buy, refer you and amplify your work. The unaware give you impressions; the others give you the business. So even if most readers cannot tell, the readers who matter most often can.

Post-level tells vs timeline-level flattening

Readers pick up two different signals. The first is the tells in the checklist above, visible in a single post. The second is harder to name: an account’s voice going flat across dozens of posts, so readers can no longer predict what you would say about something new. A screenshot or forwarded link is judged on the single post. A follower who sees you for months is judged on the timeline, and that is where most of the relationship lives.

When readers care, and when they do not

Noticing and caring are different. Readers tend to care when:

• Your voice is the reason they follow you, so a flatter voice reads as a quiet withdrawal.

• The post claims expertise or judgment, so generic output reads as a substitute for your take.

• You disclose AI use but the posts still read generic, which looks like a permission slip rather than honesty.

They tend not to care when:

• The post reads like you, however it was drafted.

• The content is functional reference, such as a narrow how-to, where readers came for the information.

• The disclosure describes the artifact, such as “an AI-drafted summary of my talk”, rather than licensing everything you post.

AI-assisted vs AI-drafted

Readers treat these differently without naming the difference. AI-assisted means you do the thinking and the editing, with AI suggesting or drafting inside your judgment, so the post carries your voice. AI-drafted means the model does the thinking and you polish the surface, so the post carries the model’s default voice.

Most people slide from the first to the second gradually. You start by editing AI drafts heavily. Months later, under publishing pressure, the edit pass gets lighter, the workflow feels the same, and the voice starts to flatten before you notice.

What this means for how you use AI

• Voice training beats disclosure. Drafts that start from your own posts remove the flattening at the source; a disclosure line only speaks to the small explicit group. See how to train AI on your voice and why all AI-written tweets sound the same.

• Audit timelines, not just posts. Once a quarter, read 20 to 30 recent posts in a row and ask whether they read as one specific person.

• Keep the edit pass real. Every so often, write a post entirely by hand. It recalibrates your ear, so you notice when drafts have drifted.

Why there is no honest detection percentage

Figures like “60 percent of readers can spot AI content” usually trace to surveys that cannot be compared, samples full of the people most likely to notice, or articles citing each other. Small public tests do not settle it either. In Hootsuite’s experiment, three human tweets were paired with ChatGPT versions on one account: the AI posts had a higher engagement rate in two of three pairs, while the human link post drove more profile visits, link clicks and bookmarks. Three pairs on one account proves little either way. Machine detectors have the same problem from the other side; see AI detection tools tested in 2026. The honest answer is the three levels above, not a single number.

The one-line answer

How do you avoid AI tells? Treat the list as habits, not a word filter: ban the stock vocabulary and update it as models change, refuse the two-part hook and the generic close, use the number of items you actually have, make every bullet specific, vary your rhythm, and put one sentence in every post that only you could write. Then run the two-minute scan. Can people tell you use AI? Some can, more sense it without naming it, and those are often your most valuable readers, so keep AI in an assisting role and your edit pass real.

If you want a tool that starts from your own patterns instead of the average, audenAI, the brain inside VoiceMoat, learns from your own posts across the elements of writing style, avoids several tells on this list by default (such as the “it’s not this, it’s that” hook and the bait-question close), keeps a list of words it must never use, and checks your drafts for Style and Stance. VoiceMoat is a Personal Brand OS for LinkedIn and Twitter. audenAI suggests. You decide. To test whether any AI tool really writes like you, use the 7-test check.