the signs of ai writing that hold up are patterns that repeat across a text. a long dash where a comma would do. words like "delve", "leverage" and "seamlessly". adverbs that clear their throat before the point. sentences that deny one thing to reveal another. stock openers, and a question at the end that begs for comments.
one of these proves nothing. people used every one of them before chatbots existed. four or five in a short post usually mean the draft went out unedited, by a person or a model. the useful question is whether anyone edited it.
| group | examples | patterns |
|---|---|---|
| punctuation | the em-dash, the semicolon | 2 |
| vocabulary | "delve", "tapestry", "leverage" | 28 |
| filler adverbs that clear the throat | "essentially", "fundamentally", "arguably" | 9 |
| contrast shapes | the negate-then-reveal, the negate-then-reveal, compact, "it's not x. it's y." | 12 |
| stock openers | "in today's [anything]", "thrilled to announce", "let's dive in" | 12 |
| closers that beg for engagement | "let that sink in", "read that again", "thoughts?" | 9 |
every group below is generated from the same pattern table our generator runs on its own drafts, so each item is something our code catches today. the full table, with the regex for each entry, lives on the blacklist. see all 112 on the blacklist.
punctuation: the em dash and the semicolon
the em dash is the most cited tell, because chat models use it where most people type a comma, a colon or a full stop. the semicolon is a quieter one: in a social post it reads as written for print, and we rewrite it for that reason.
what to do instead: our generator swaps each long dash for a hyphen or a comma and ends the sentence at each semicolon. in your own drafts, read the sentence out loud and use the mark you would hear.
more on this: why ai uses em dashes
vocabulary: the words ai overuses
a short list of words turns up in model output far more than in normal speech: verbs that promise movement, adjectives that promise scale, nouns borrowed from biology and textiles. none of them is wrong. a paragraph with three of them reads like a press release.
- "delve"
- "tapestry"
- "leverage"
- "journey"
- "landscape"
- "robust"
- "seamlessly"
- "foster"
- "game-changer"
- "unlock"
- "elevate"
- "paradigm"
- "synergy"
- "holistic"
- "multifaceted"
- "nuanced"
- "pivotal"
- "groundbreaking"
- "navigate"
- "resonate"
- "empower"
- "transformative"
- "ecosystem"
- "stakeholder"
- "actionable"
- "impactful"
- "scalable"
- "innovative"
what to do instead: say the plain word. "leverage" is "use". "navigate" usually means you read an email. we flag these and leave them in place, because irony is legal and a regex cannot tell a joke from a sales deck.
more on this: corporate jargon examples and plain replacements
filler adverbs that clear the throat
models open sentences with an adverb that grades the sentence before you read it. it signals importance or contrast and adds no information. delete one and the sentence keeps its meaning, which is how you know it was filler.
- "essentially"
- "fundamentally"
- "arguably"
- "interestingly"
- "undoubtedly"
- "importantly"
- "significantly"
- "moreover"
- "furthermore"
what to do instead: delete it. if the sentence then reads as weaker, the problem was the sentence.
contrast shapes: denying one thing to reveal another
the sentence shape a model falls into most is the denial followed by the reveal: "it's not x. it's y." it sounds like insight and skips the argument. close cousins: "less about x, more about y", the three-fragment staccato, and the then-versus-now couplet.
- the negate-then-reveal
- the negate-then-reveal, compact
- "it's not x. it's y."
- "it is not x. it is y."
- "x isn't y. x is z."
- "this isn't x. it's y."
- "you're not x. you're y."
- "not just about x"
- "less about x, more about y"
- "it was never about x"
- the staccato "no x. no y."
- the then-vs-now couplet
what to do instead: state the claim you meant and give the reason. a regex cannot rewrite an argument, so our generator sends these sentences to a repair pass that rewrites them, and the detector highlights them in your drafts.
more on this: edits that stop ai slop before it ships
stock openers
the first line of a model draft often announces the post instead of starting it. "in today's fast-paced world", "let's dive in", "here's the thing". each one was a fresh move once. now they mark a template.
- "in today's [anything]"
- "thrilled to announce"
- "let's dive in"
- "without further ado"
- "here's the thing:"
- "let's be real"
- "unpopular opinion:"
- "picture this:"
- "imagine this:"
- "here's what nobody tells you"
- "it's important to note"
- "it's safe to say"
what to do instead: cut the first line and start at the second. it is usually the real opening.
closers that beg for engagement
the last line of a model-written social post is often a question or a command aimed at the comment box: "thoughts?", "agree?", "let that sink in". the rocket emoji belongs here too. readers have seen thousands of these and scroll past them.
- "let that sink in"
- "read that again"
- "thoughts?"
- "agree?"
- "what do you think?"
- "curious to hear your thoughts"
- "let me know in the comments"
- "drop it in the comments"
- the rocket emoji
what to do instead: end on your last real sentence. if you want replies, say something specific enough to disagree with.
more on this: engagement bait examples
how to check a text for ai tells
- paste the text into the slop detector. it runs this table in your browser and uploads nothing.
- count the groups it hits, and ignore single hits. one em dash is a style. hits in four groups are a pattern.
- read the highlighted sentences out loud. the ones you would never say to a person are the ones to rewrite.
open the slop detector or read what ai slop is for the bigger picture.
questions about ai writing tells
what are the most common signs of ai writing?
em dashes where a comma would do, a cluster of words like "delve", "leverage" and "seamlessly", throat-clearing adverbs such as "moreover" and "essentially", sentences that deny one thing to reveal another, stock openers, and a closing question aimed at the comments. the signal is several of these together.
are em dashes a sign of ai writing?
on their own, no. writers used em dashes long before chatbots. in social posts they became a tell because chat models put them where people typing on a phone rarely do. an em dash next to two or three other patterns from this page is a stronger sign than the dash alone.
what words does ai overuse?
the vocabulary list on this page is the set our generator flags: words such as "delve", "tapestry", "leverage", "journey", "landscape", "robust", "seamlessly", "foster", "unlock", "elevate", "synergy", "holistic", "navigate", "empower" and "ecosystem". each has a plain word that says the same thing.
can you tell for sure if a text was written by ai?
no. these patterns show that a draft went out unedited. they cannot show who or what wrote it. people write every one of them, and a model told to avoid them produces far fewer. treat any tool that claims certainty with suspicion, including ours.
how do i check my own post for ai tells?
paste it into the free slop detector on this site. it highlights every pattern from this page, grades the post from a to f and runs in your browser without uploading the text.
want drafts that skip all of this?
the ai shitpost generator runs this table on every draft before you see it.
