Readers can usually feel AI-written text before they can explain it: the paragraph is grammatically perfect and somehow empty, every sentence glides into the next, and nothing in it sounds like a person who actually holds an opinion. That feeling has a cause. Large language models have habits — favourite words, favourite sentence shapes, favourite ways of hedging — and when those habits pile up, prose starts to sound generated even when a human wrote it. Our free AI tell highlighter makes those habits visible: paste your writing and it scans for more than 70 such patterns, highlighting every hit with a plain-English explanation of why it reads as AI-ish and what to do instead. Nothing is uploaded; the whole scan runs on your device. This guide explains how to use it, what each pattern family catches, and — most importantly — how to revise so the result sounds like you, not like a template.
Step one: paste your draft. Drop in anything from a college essay to a marketing email. Aim for at least 20 words — very short texts give the pattern scan too little to work with, and the tool page says so plainly. If you have nothing handy, use the built-in sample to see the scan in action before trusting it with your own work. Step two: choose what to scan for. Four category toggles let you focus: overused transitions, clichés and buzzwords, hedgy qualifiers, and repetitive structures. Scan everything at once for a first pass, then toggle categories off to work through one family at a time. Step three: read the cards, not just the colours. Every match is highlighted in your text — amber for transitions, red for clichés, indigo for qualifiers, green for structures — and below the text you get counts, your top category, and a card per triggered pattern with its explanation and the exact phrases it caught. Hover any highlight to see its explanation inline. The revision step is where the value lives: work through the cards and decide what earns its place. Delete filler like "it's important to note", swap "utilize" for "use", break up em-dash chains. The goal is not zero highlights — it is deliberate, specific writing that sounds like you.
Overused transitions are the classic glue language models lean on to stitch ideas together: furthermore, moreover, additionally, firstly, lastly, in conclusion. One is fine; five in a page is a rhythm no human editor would leave standing. Clichés and buzzwords — delve, unlock, game-changer, cutting-edge, vibrant tapestry — appear in AI output far more often than in human drafts, because models reach for the most probable phrase, and the most probable phrase is usually the most tired one. Hedgy qualifiers — "it's worth noting that", "arguably", "tends to", "quite" — soften every claim the way a model hedging its bets does; a paragraph full of them reads as if nobody is willing to stand behind a single sentence. Structural habits are the subtlest family: em-dash chains, triple-item lists, and "not only X but also Y" constructions give prose that unmistakable generated rhythm even when every individual word is unremarkable. Each highlight links to a card explaining why the pattern reads as AI-ish, plus when it is perfectly fine to keep — because "moreover" is not a crime, and a triple list is sometimes exactly the right shape. The scan is a style editor, not a judge.
The tool page's own example is a 182-word product announcement draft. The scan finds 16 highlights: 7 transitions (moreover twice, furthermore, additionally, in conclusion, firstly, lastly), 4 clichés (delve, unlock, game-changer, cutting-edge), 3 qualifiers ("it's important to note", arguably, tends to) and 2 structural habits (a triple list, an em-dash chain). That is roughly one highlight every 11 words — a density where readers start hearing the "AI voice" even in human-written text. Revision takes ten minutes: moreover and furthermore become specific connectors ("because", "for example"), delve becomes "explore", the triple list becomes two crisp sentences. The result keeps 4 deliberate highlights — a 75% drop — and reads like a person with opinions instead of a model with templates. Notice what the revision did not do: it did not chase zero. The four remaining highlights are choices, not accidents — a transition that genuinely helps, a qualifier on a claim that genuinely deserves one. That is the standard to aim for in your own draft: not a clean scan, but a scan where every remaining highlight is one you can defend. Run the scan again after revising; the count, the top category and the cards all update, so you can see whether the density actually fell or whether you just moved the problem to another paragraph.
Treating a highlight as an accusation. A flagged phrase proves nothing about authorship — humans write "moreover" too, and models can avoid every pattern on the list. Use the tool on your own writing to improve it, not to accuse someone else's. Chasing zero highlights. Deleting every transition and qualifier leaves prose that is blunt, not human; the goal is deliberate writing, and some patterns earn their place. Scanning 15 words and concluding the tool is broken. Very short texts give the scan too little to work with; paste at least 20 words, and longer drafts give far better signal. Revising only the words, not the structures. Swapping "delve" for "explore" while keeping every triple list and em-dash chain leaves the generated rhythm intact — fix the sentence shapes too. Assuming a clean scan means human-written. The catalog is pattern-based, not model-based: it cannot tell you who wrote a text, and a careful writer — human or machine — can pass it. Forgetting the catalog is English-only. Paste English text; other languages are not covered, and the patterns would not transfer anyway, since every language has its own filler habits.
The honest limits are four, and the tool page states the first one in plain language: this is a style editor, not an AI detector. It flags patterns; it cannot tell you who wrote a text, and a highlighted phrase proves nothing about authorship. Second, the catalog is English-only, so non-English drafts are out of scope. Third, very short texts under about 20 words give weak signal — the scan needs enough prose for patterns to repeat before density means anything. Fourth, because the scan is pattern-based rather than model-based, it runs in seconds on any text with no account and no waiting, but it also cannot catch what is not in its catalog: a novel, fluent AI paragraph with none of the 70+ habits will scan clean. Within those limits, the tool does one job well: it shows you where your writing sounds generic, explains why in plain English, and leaves every revision decision to you. Writers use it before submitting essays, job applications and manuscripts; teachers use it to show students why a paragraph feels flat. The companion guides are our AI prompt guide, for writing prompts that produce less generic output in the first place, our word count guide, for the essay and application limits polished drafts most often have to meet, and our character counting guide, for fitting a revised draft to a platform limit.
Write prompts that produce less generic output in the first place.
The essay and application limits polished drafts most often have to meet.
Fit a revised draft to a platform limit when every character is audited.