AI writing tell checker
Paste text and see which patterns give it away: em dashes per 1,000 words, the “not just X, it’s Y” shape, model vocabulary, rule-of-three lists, and sentence rhythm that never varies. Runs in your browser — nothing is uploaded.
Is an em dash a sign of AI writing?
On its own, no. The em dash is ordinary punctuation, and good writers have used it for two centuries. What gives model output away is density: several per paragraph, where an editor would reach for one every few paragraphs. That is why this tool reports a rate per 1,000 words rather than a raw count.
A rough calibration: under 2 per 1,000 words is unremarkable, and above about 4 reads as machine-written to most editors. For a 500-word post, one or two em dashes are invisible and six start to look like a pattern.
Why does AI use so many em dashes?
Because the training data rewards them and the grammar is forgiving. Em dashes are dense in exactly the prose models saw most of — journalism, marketing copy, essays — and they let a sentence attach a qualifying clause without committing to a structure. Where a writer would have to choose between a comma, a colon, a parenthesis or a full stop, the dash works in all four positions. It is the path of least resistance to a fluent-sounding sentence, so it turns up constantly.
The same logic explains the other tells. The full breakdown of the em-dash tell covers the constructions that make it worse, and the wider list of AI writing signs covers the twelve patterns readers notice most.
What this checker looks for
- Em dash density — count and rate per 1,000 words.
- Antithesis constructions — “it’s not just X, it’s Y”, the most recognisable model sentence shape there is.
- Model vocabulary — delve, tapestry, realm, “it’s important to note”, “in today’s fast-paced world” and their relatives, drawn from our list of AI words to avoid.
- Rule-of-three lists — models produce triads far more often than people do. Real lists are lumpy.
- Uniform sentence rhythm — human prose varies sentence length by 50–80%; model prose clusters tightly around its mean.
- Assistant openers and bold-label bullets — “Certainly!”, “Great question!”, and formatted bullets in a message nobody asked to be formatted.
This is not an AI detector, and you should distrust the ones that claim to be
This tool counts patterns and tells you which it found. It cannot tell you what wrote a piece of text, and neither can anything else with acceptable reliability. Careful human writers trip every rule here; lightly edited model output passes all of them. Use the flags to make your own draft read better, not to accuse anyone of anything.
FAQ
Is an em dash a sign of AI writing?
On its own, no — the em dash is ordinary punctuation and good writers have always used it. The tell is density. Language models use several per paragraph where an editor would use one every few paragraphs. This checker reports em dashes per 1,000 words rather than a raw count, because the rate is the signal.
Why does AI use so many em dashes?
Because the training data rewards them. Em dashes appear heavily in polished web prose — journalism, marketing copy, essays — and they let a model attach a qualifying clause to almost any sentence without committing to a grammatical structure. It is the path of least resistance for a fluent-sounding sentence.
How many em dashes are too many?
Above roughly 4 per 1,000 words reads as machine-written to most editors; under 2 is unremarkable. For a 500-word post that means one or two em dashes are fine and six start to look like a pattern.
Is this an AI detector?
No, and be sceptical of anything claiming to be one. This counts specific stylistic patterns and tells you which it found. It cannot tell you who or what wrote a piece of text, and no tool reliably can — human writers trip these patterns and edited model output passes them.
Does my text get uploaded anywhere?
No. The analysis runs entirely in your browser as plain JavaScript. Nothing is sent to a server, stored, or logged — you can check by opening your network tab, or by disconnecting from the internet and running it again.
What should I do with the results?
Treat each flag as a question, not a verdict. Keep the em dash that carries real emphasis and convert the rest. Replace model vocabulary with the plainest word that means the same thing. Break one long sentence in half so the rhythm stops being uniform.
Want the fix, not just the diagnosis?
This page tells you what is wrong. The Comment Humanizer rewrites the draft in a real Reddit or X voice and strips the tells for you — free, no signup for the first few runs a day.
Finding the threads worth replying to is the other half. Prowlo serves your agent filtered Reddit and X conversations over MCP — read-only, so it finds the thread and you write the reply.