Is this AI-generated?
Paste any text for a named verdict plus the sentences that give it away.
50 free scans a day · 2,500 words per scan
What this check is
A verdict about a piece of text, not a finding about a person
Paste the passage. You get one of three named verdicts — Human-Written, AI-Assisted, or AI-Generated — plus sentence-level shading showing which lines carried the signal. 50 free scans a day, up to 2,500 words each, no account, no card. Around fifty words is the floor; confidence rises with length.
The percentage on the dial backs the verdict beside it — how human the text reads for Human-Written, how likely it is that a model wrote it for an AI verdict. The verdict comes from thresholds calibrated per length band, not round cut-offs, and text that is part human and part machine is genuinely mixed rather than undecided. That mixed ground is where most real documents now live, which is why it gets its own verdict instead of being rounded up or down.
You can also read for the signs yourself: even rhythm, claims that never quite get asserted, decorative vocabulary, a conclusion that adds nothing, missing specifics. Every one of those is defeated by a careful editor, and several of them fire on perfectly human writing — formal prose, non-native English, anything written to a template. That is the gap a detector is for, and also why a detector is not enough on its own.
Can AI detectors be wrong? Yes, in both directions. False flags cluster on non-native English, heavily formal writing, and very short passages. Missed detections cluster on heavily edited or deliberately rewritten AI text. Read the sentence-level highlights before you act on the number. Do not use this page to accuse someone. A report is evidence to weigh alongside draft history, the writer's usual voice, and whether they can talk about what they wrote.
The signs above are ordered by how well each one survives a person editing the draft. Vocabulary is the first thing a thesaurus pass changes and the weakest tell. Rhythm is the last thing a careful editor notices in their own work, which is why evenness is the tell that still shows up after a synonym pass. Missing specifics — names, dates, numbers, first-hand detail — are the other strong signal, and they are also the first thing a regenerate can invent or drop, so check them even when the cadence looks human.
If you already have a guess about the source model, the ChatGPT, Claude, and Gemini detector pages go into that model's scaffolding. If you want the product tour rather than a question-shaped page, that is the AI detector. If you are a teacher deciding what to do after a flag, that is the educator page. This URL exists because Search Console showed people asking "is this AI generated" — a different question than "AI detector" — and a pure tool page gives a question-shaped query nothing to match.
By eye
Five signs
you can read yourself
Worth knowing even with a scanner open — a detector gives you a number, but these are what let you argue with it. They are ordered by how well each one survives a person editing the draft.
The rhythm is too even
Machine paragraphs tend to run to similar lengths, and the sentences inside them to similar shapes. Human writing is lumpy — a short sentence lands after a long one because the writer wanted it to. Evenness is the tell that survives the most editing, because it is the hardest one to notice in your own draft.
Nothing is ever quite asserted
Every claim arrives with its own counterweight attached. "While this is significant, it is important to consider…" A person with a point of view eventually commits to it; a model trained to be balanced often never does.
Decorative vocabulary
Words doing ornamental rather than semantic work — delve, tapestry, realm, testament, pivotal, underscore. Any one of them is fine. Several in a page, in text that otherwise reads plainly, is not.
The conclusion adds nothing
A final paragraph that restates the introduction in different words. Human endings usually arrive somewhere the opening did not predict, because the writer worked something out along the way.
Specifics are missing
No names, no dates, no numbers, no anecdote that could only come from one person. Models generalise well and remember nothing, so machine text is often fluent and simultaneously about nobody in particular.
…and why that isn’t enough
Every sign above is defeated by a careful editor, and several of them fire on perfectly human writing — formal prose, non-native English, anything written to a template. That is the gap a detector is for.
Reading the result
What the answer actually means
The percentage on the dial backs the verdict beside it — how human the text reads for a Human-Written verdict, how likely it is that a model wrote it for an AI one. The verdict comes from thresholds calibrated per length band, not round cut-offs, and text that is part human and part machine is genuinely mixed rather than undecided.
That mixed ground is where most real documents now live, which is why it gets its own verdict here instead of being rounded up or down. A drafted-by-ChatGPT-then-edited-by-you essay is not a failure of the detector to make up its mind; it is an accurate description of the document.
And no detector is right every time. Short passages, non-native English and heavily rewritten text are all genuinely hard. Read the sentence-level highlights before you act on the number — the reasoning is the part you can check.
By model
Know which one you're looking for?
Each model leaves a different mark. If you have a guess about where the text came from, these go into the specific tells.
Who this check is for
A question-shaped query, a tool and a method
People land here because they are holding a specific piece of text, not because they wanted a product tour. The scanner answers that text. The signs below answer the next time you do not have a scanner open.
You wrote it and want a second read
Paste your own draft. See which sentences still look machine-shaped. That is useful even if you never show the report to anyone else. A thesaurus pass will usually still show up. A genuine structural rewrite may not. Either result is information about the draft, not a character judgement.
You have to review someone else's writing
A named verdict plus sentence shading is something you can discuss. It is not something you should treat as the sole basis for a grade, a rejection, or an accusation. Ask how the flagged lines were written. Ask for version history. The conversation is the product; the report is the map.
The passage is short or heavily edited
Those are the hard cases. Treat the result with more caution, read the highlights, and look for missing specifics by eye. Below about fifty words the engine will not score at all — not because it is being difficult, but because a couple of sentences do not carry enough rhythm to judge.
Not: a courtroom exhibit
No detector, including this one, is reliable enough to carry an accusation by itself. The consequences of a false positive fall on a person. We do not publish a headline accuracy figure. We do not publish pass rates against other scanners. Where tools disagree, the text is genuinely ambiguous.
After you scan
The report is a map of attention, not a verdict about a person
Point at two or three shaded lines and ask how they were written. That is a discussable observation. Reading the percentage aloud is not. On a document judged human or mixed, you get both colours, so you can see which passages read machine-written and which do not. On a document judged AI-authored, every line is shaded — strongly where it gives itself away, softly where it does not — and never painted as human, because a humanizer is built so that no individual sentence looks wrong.
Then look for the signs that do not need a scanner: even paragraph mass, claims that never quite get asserted, a conclusion that adds nothing, missing names and dates. If those signs and the shading agree, you have a stronger case for a conversation. If they disagree, the text is genuinely ambiguous — which is also what it means when two detectors split. We do not publish a headline accuracy figure. We do not publish pass rates against GPTZero, Turnitin, or Originality.ai.
Text is scored in memory and discarded when the response returns. 50 scans a day, 2,500 words each, about fifty words before the engine will score. Lists, code, and slide bullets are a weak sample even when they are long. Continuous prose is the sample this page is for. If you do not have the right to paste the document, do not paste it.
If the result on your own writing looks wrong, it may simply be wrong. Formal prose, technical writing, and fluent English written as a second language all sit close to the decision boundary. Tell us; confident misses are the most useful mail we get. If the result on a machine draft looks clean, that is the hard case — usually a structural rewrite rather than a synonym pass — and a reason to read the words, not a reason to trust a marketing claim about undetectable output.
FAQ
Common questions
How to check, how much text you need, and what a result does and does not entitle you to conclude.
How can I tell if something was written by AI?
Read for rhythm before vocabulary. Machine text tends toward even paragraph lengths, evenly shaped sentences, claims that always arrive with a balancing qualification, and a conclusion that restates the opening. Missing specifics — names, dates, numbers, first-hand detail — is the other strong signal. A detector adds a statistical read on top of those impressions, which is useful precisely because the impressions are hard to trust on a single short passage.
Is there a free way to check if text is AI generated?
Yes — this page. Paste the text and scan it: no account, no card, 50 scans a day at up to 2,500 words each. You get a verdict plus sentence-level shading showing which passages carried the signal.
How much text do I need?
Around fifty words is the floor, and confidence rises with length. Short passages are genuinely hard for any detector — there is not enough material for the statistical patterns to separate from noise — so treat a result on a couple of sentences with real caution.
Can AI detectors be wrong?
Yes, in both directions, and anyone who tells you otherwise is selling something. False flags cluster on non-native English, heavily formal writing, and very short passages; missed detections cluster on heavily edited or deliberately rewritten AI text. This is why the report shows you the sentences behind the call rather than only a number — the reasoning is the part you can actually check.
What if the text was written by AI and then edited by a person?
That is the most common real case and it has its own verdict here: AI-Assisted. It means machine signals are present without the document reading as machine output end to end. It is not a hedge between the other two answers — it is a description of what most drafts now actually are.
Can I use this to accuse someone of using AI?
Please don't treat it as proof on its own. A detector produces evidence to weigh alongside everything else you know — draft history, the writer's usual voice, whether they can talk about what they wrote. No detector, including this one, is reliable enough to carry an accusation by itself, and the consequences of a false positive fall on a person.
If you are a teacher, the educator page is the process between a flag and a decision: triage privately, compare with earlier writing, ask for version history, talk, act only on convergence. A percentage read aloud as an opening move is the wrong start. If you are an editor or a hiring manager, the same caution holds — especially on short passages, which are a weak sample.
What do the three verdicts mean on a real document?
Human-Written means the text reads as composed by a person. AI-Generated means strong machine signals throughout. AI-Assisted means a model was involved without the document reading as machine output end to end — the most common real case, and the one a single percentage handles worst. The percentage backs the named verdict; it does not stand on its own. Do not re-derive a verdict by cutting a number at a round threshold; length calibration would be thrown away.