# Free AI-writing detector ยท ScriptGrain > Paste any text and get an indicative, stylometric human-vs-AI reading, with the signals that give AI writing away. Free with a ScriptGrain account. Canonical: https://scriptgrain.com/ai-detect # Free AI-writing detector Paste any text and get an indicative, stylometric human-vs-AI reading, with the signals that give AI writing away. ## What it reads The detector reads the same style markers the voice profiler does: sentence-length variance, hedging, punctuation rhythm, and the stock phrasings models fall back on. Machine-written prose tends to even out sentence length, hedge symmetrically, and reach for the same connective tissue ("moreover", "it's worth noting", "in conclusion"): habits that are measurable, not vibes. ## What you get back A human-vs-AI confidence score plus the per-signal breakdown: which specific habits in the text read as machine-written and which read as human. That breakdown is the useful part: it tells you what to change, not just whether a detector somewhere might flag you. You can also check it, then humanise it: humanise rewrites the text so it reads as human-written, keeping its structure, tone and facts. It runs after the check, on paid plans, for 1 credit, and is free when the text already reads as human. For the wider picture on detector reliability, we keep a sourced data page of every published accuracy figure at [AI detector accuracy: every published number](https://scriptgrain.com/reference/ai-detector-accuracy). ## How the detector reads a text It reads style, not meaning. That's the first thing to get straight. A stylometric detector doesn't know whether your claims are true or whether a person sat and thought before typing. It measures habits. The same habits the voice profiler measures: sentence-length variance, hedging, punctuation rhythm, and the stock phrasings language models fall back on when they've got nothing specific to say. Then it returns a human-versus-AI confidence score, plus a breakdown of which habits read as machine-written. So the breakdown is the useful part. The number is a summary; the signals are the evidence. Here's the thing: two texts can land on the same score for completely different reasons. One might be flagged for flat sentence rhythm, every line about the same length, the same gentle cadence. Another might be flagged for a cluster of stock phrasings. The first is a rhythm problem you can hear and fix by reading it aloud. The second is a wording problem you can search for. A single figure hides that difference, and the difference is what tells you what to do next. Two details matter for trust. Counted facts are measured in code and passed to the judge, and any signal that quotes a phrase absent from the text, or contradicts the counts, is dropped. So if a signal says your draft leans on a particular phrasing, that phrasing is actually in your draft. And the same text gets the same reading every time. Run it twice, get the same answer. That sounds minor until you've used a tool that shifts every time you paste. It's free, too. Create a free account to see the result, no card. ## Why detectors disagree Paste one paragraph into three detectors and you can get three different answers. That isn't a bug in any one of them. It's what happens when different tools measure different things, weight them differently, and train on different samples of what "human" and "machine" look like. Every detector produces false positives. Every one. ScriptGrain keeps a sourced reference page of every published detector accuracy figure, because the honest position is that accuracy claims vary and you should see where each one comes from before leaning on any of them. Some of the disagreement is predictable, though, and it falls hardest on two kinds of writer. The first is the plain, well-edited writer. Think about what good editing does. It cuts the digressions, evens out the clumsy patches, removes the odd tic. You end up with clean, consistent prose, and clean, consistent prose is exactly what a rhythm-based signal reads as suspicious. A tidy paragraph of medium-length sentences can look statistically like machine output, because machines also produce tidy paragraphs of medium-length sentences. The irony is fairly rich. The more carefully you edit, the more your writing can resemble the thing the detector is hunting. The second is the writer working in a second language. Someone writing in their third or fourth language often relies on a smaller, safer set of phrasings, sticks to constructions they trust, and avoids risky idiom. That narrower range can read as low variance. It's careful writing, and it gets penalised for being careful. To be fair to detectors, none of this makes them useless. It means they're measuring a tendency, and tendencies have exceptions. Plenty of the exceptions are people who write well. ## What to do with a high AI reading Don't panic. Treat the score as a signal, never a verdict. That's the whole instruction, and the rest is how to follow it. So, step by step: 1. Read the signals, not the headline figure. Find out which habits were flagged. If it's sentence-length variance, look at your rhythm. If it's a stock phrasing, find it and ask whether you'd actually say it. 2. Check the flagged habit against what you know about your own writing. If you always write in even, measured sentences, the flag may simply be describing you accurately. That's a fact about your style, not an accusation. 3. Keep your drafts. Version history, earlier drafts, notes, the messy first pass with the half-finished sentences. That trail is real evidence of how a piece came together, and it's far more persuasive than any score. 4. Never let one number settle an argument. If someone else is reading a score about your work, point them to the signals and the drafts. A short example. Say a reader flags a paragraph of yours, and the breakdown shows flat sentence rhythm and two stock phrasings. You open the document, find the phrasings, and recognise one as a habit you picked up from years of report writing. You rewrite it the way you'd say it aloud. The rhythm issue you leave alone, because it's simply how you write when you're being precise. Now you've used the reading to understand your own text. That's the right use. What this isn't is a hunt for a trick to get past a checker. ScriptGrain's position is voice preservation. If your writing has drifted away from how you sound, the fix is to bring it back towards your own voice, and the detector is one way of noticing the drift. That's also what Humanise is for, on paid plans. It runs after the check, costs 1 credit, and is free when the text already reads as human. It keeps structure, tone and facts. ## Checking your own writing before it goes out A sensible pre-publication check is short. It shouldn't take longer than the edit itself. Run the detector on the finished draft, after your own edit and before anyone else sees it. Look at the signals first. A clean reading is reassuring, but a flagged habit is more useful, because it points at a specific sentence or pattern you can judge for yourself. Then read the flagged parts aloud. Your ear is a better instrument than any score here. If a line sounds like you, leave it. If it sounds like something you'd never say, change it. If you write to a measured voice, there's a second check worth having. A profile built from your published writing records 45 attributes across 8 layers: lexical, syntactic, tone and register, rhetorical, punctuation and format, function words, content patterns, and quirks and cadence. Every draft gets scored 0 to 1 against it, which is the voice match, and it comes back with per-feature deltas showing the draft's value against your profile's value. On-voice drafts typically land 0.85 to 0.95; partial matches 0.5 to 0.8; a clearly different voice under 0.45. A delta that shows your draft hedging more than you normally do, or running longer than your usual sentences, tells you precisely what to pull back. The two checks answer different questions. The detector asks whether the text reads as machine-written. The voice match asks whether it reads as you. You can pass the first and fail the second, and the second is the one your readers notice. So run both, keep your drafts, and read the signals. At the end of the day, the number is only ever a prompt to look closer at your own words. ## Questions ### Is the AI-writing detector free? Yes. Paste any text (an article, email, LinkedIn post) and the check runs free. Create a free account to see your result; no card is required, and joining our letter is optional. ### How does the detector work? It is stylometric: it reads the same style markers ScriptGrain's voice profiler measures (sentence-length variance, hedging, punctuation rhythm, and the stock phrasings language models fall back on) and returns a human-vs-AI confidence score with the exact signals that drove it. ### How accurate are AI detectors? No detector is perfect, and published accuracy figures vary widely between tools and text types. We have collected every published accuracy number, with sources and dates, at scriptgrain.com/reference/ai-detector-accuracy. Read that before treating any detector score as a verdict. ### Can the detector be wrong about my writing? Yes. Every detector produces false positives, especially on plain, well-edited prose. Treat the score as a signal, not a verdict: the per-signal breakdown shows you exactly which habits read as machine-written, which is more useful than the number itself. Related: the free [writing style analysis tool](https://scriptgrain.com/tools/writing-style-analysis) measures your own style, and a [ScriptGrain voice profile](https://scriptgrain.com/why) measures 45 attributes and writes with them.