Free tool / AI writing tells
Scan your text for AI writing tells
Paste an essay, a blog draft, an email — anything up to 3,000 words — and see the specific style patterns that make writing read as machine-written: slop vocabulary, cliché phrases, em-dash pileups, sentences that are all the same length, and rule-of-three lists. Each one is highlighted where it sits. Free, runs entirely in your browser, nothing you paste is uploaded, no signup.
Up to 3,000 words · runs in your browser · nothing is uploaded · ⌘/Ctrl + Enter to scan
What it found
Every match is marked in the text below, and each pattern gets its own count. There is no overall number on purpose: a count you can check beats a figure you have to take on faith.
Slop vocabulary
Words that chat models reach for far more often than people do — delve, tapestry, testament, pivotal, seamless.
Why this pattern reads as machine-written: these words are rare in everyday writing but sit at the top of a language model's default vocabulary, so a cluster of them is the first thing many readers notice.
Cliché phrase density
Stock phrases and templates — "in today's fast-paced world", "it's important to note", "it's not just X, it's Y".
Why this pattern reads as machine-written: models fall back on the same connective filler in every draft; a person has usually cut it by the second pass.
Em-dash density
How often the text uses an em dash (—) per 1,000 words. Edited prose usually uses them sparingly; a pileup is one of the most commonly reported tells in recent chat output.
Why this pattern reads as machine-written: a run of em dashes is the punctuation habit readers have learned to associate with chat output, even when they couldn't say why.
Sentence-length uniformity
How much sentence length varies. People write in bursts — a long sentence, then a short one. Nothing is highlighted for this one; it is a property of the whole text.
Why this pattern reads as machine-written: model output tends toward sentences of similar length, which gives it an even, metronomic rhythm.
Triadic structure (the rule of three)
Three-item lists: "faster, cheaper, and more reliable". One is fine; one in every paragraph is a template.
Why this pattern reads as machine-written: three-item lists are a default shape in model output, so they stack up faster than a person would write them.
Also checked: repeated paragraph openers
Paragraphs that start the same way — "Additionally", "Furthermore", "Moreover".
Why this pattern reads as machine-written: templated openers are how a model keeps a structure going; people vary how they enter a paragraph.
Want the patterns gone, not just marked? The same text goes straight into the ToHuman editor — no second paste.
Fix this with ToHumanFree plan: 2,500 words a month, no credit card.
What counts as an AI writing tell
An AI writing tell is a style habit that shows up far more often in text produced by a language model than in text a person drafted and edited: a particular vocabulary, a stock phrase, a punctuation habit, a rhythm. Tells are about how the text reads, not where it came from — a person can pick up every one of them, and a model can be prompted out of most of them. That is why this page counts and highlights patterns instead of guessing at origin: the useful question is not "who wrote this?" but "which habits make it read that way, and where are they?"
The 5 tells this scanner looks for
1. Slop vocabulary
A set of roughly a hundred words that chat models use many times more often than people do — delve, tapestry, testament, pivotal, seamless, leverage, robust, multifaceted, navigate, unlock. Each occurrence is highlighted and the count is shown per 1,000 words, so a single robust in a long report is not treated the same as six in a paragraph.
Before: "Let's delve into the rich tapestry of options and unlock the potential of your workflow."
After: "Here are your options, and how to get more out of them."
2. Cliché phrase density
Stock phrases and sentence templates: "in today's fast-paced world", "it's important to note", "at the end of the day", "navigate the complexities", "a testament to", and shapes like "it's not just X, it's Y" or "not only X but also Y". About two hundred phrases and four templates are matched; the count is again shown per 1,000 words.
Before: "In today's fast-paced world, it's important to note that time management is not just a skill, it's a necessity."
After: "You have less time than you think, so plan the week before it starts."
3. Em-dash density
How many em dashes (—) the text uses per 1,000 words, counting a double hyphen or a spaced en dash used the same way. Edited prose usually uses them sparingly; a pileup is one of the most commonly reported tells in recent chat output. The dash itself is good punctuation — the pattern is the pileup.
Before: "The plan is simple — ship early — get feedback — then iterate — quickly."
After: "The plan is simple: ship early, get feedback, then iterate quickly."
4. Sentence-length uniformity
How much sentence length varies across the text, measured as the spread of sentence lengths relative to their average. People write in bursts: a long sentence, then a short one. Model output tends to settle on sentences of similar length, which reads as even and metronomic. This needs at least five sentences to say anything, and nothing is highlighted for it — it is a property of the whole text, not of one spot.
Before: "The team met on Monday to plan the launch. They reviewed the timeline and the budget in detail. Everyone agreed the schedule was realistic. The launch was set for the first week of March."
After: "The team met Monday. They went through the timeline and the budget line by line, and by the end everyone agreed the schedule would hold. Launch: first week of March."
5. Triadic structure (the rule of three)
Three-item lists — "fast, secure, and reliable" — counted per paragraph. One is ordinary rhetoric. One in every paragraph, each with items of the same shape, is the template showing through.
Before: "Our platform is fast, secure, and reliable. It helps teams plan, build, and ship. You'll save time, money, and effort."
After: "Our platform is fast and it doesn't go down. Teams use it to plan and ship; most say it saves them about a day a week."
It also checks one more thing when the text has three or more paragraphs: whether several of them open with the same word or pair of words ("Additionally, …", "Furthermore, …", "Moreover, …"). Templated openers are how a model keeps a structure going; people vary how they enter a paragraph.
What this scanner does not do
It does not tell you whether a text was written by a person or by a model, and it does not try to. There is no overall number, no percentage and no yes-or-no answer on this page — only counts of specific patterns, each one highlighted so you can see exactly what triggered it. It is not an AI detector and it should not be used as one. Tools that claim to tell the two apart get it wrong often enough to have cost real people grades and jobs; the numbers are in our write-up on AI detection false positives. A text with several tells can be entirely human-written (plenty of people write "delve"), and a text with none can be model output that was prompted well. Read the highlights as editing notes, nothing more.
Nothing you paste is sent to ToHuman or to anyone else: the scan runs in your browser. The only thing recorded is an anonymous count of how many words were scanned and how many patterns were present, so we know the tool is being used.
Fix the tells it found
If you want the patterns gone rather than marked, the same text can go straight into ToHuman: the button carries it over to the editor (or to signup first, then the editor) so you don't paste twice. ToHuman rewrites the sentence-level habits — the vocabulary, the filler, the rhythm — while keeping what the text says, and you can run the result back through this page to see what changed.
Free plan: 2,500 words a month, no credit card.
Built by the ToHuman team.