AI Detector

You paste a text or upload a document, the tool analyses it and shows two percentages: the probability that it was written by an artificial intelligence, and the probability that it was written by a human. The result comes with the style markers that were observed, so you know what the estimate rests on instead of receiving a bare number.

The analysis looks at how the text is written, not at what it is about. It examines the regularity of sentence lengths, the kind of logical connectors used, how constant the register stays, whether concrete examples are present or absent, and how anchored the writing is in the real world — dates, figures, lived experience. These are the points on which a generated text and a human one diverge most sharply.

Three hundred characters are needed at minimum. Below that, there is simply not enough material to observe a style: a sentence or two could have been written any way by anyone. The longer the text, the more solid the estimate, and the tool accepts up to twenty-five thousand characters at once.

• paste a text or upload a PDF, Word or text document
• three hundred characters minimum, twenty-five thousand maximum
• the analysis covers the style, never the subject matter
• the markers observed are shown with the result

The percentage indicates how much the style of the text resembles that of a generation. A score of 85% means the writing strongly shows the characteristics of AI-produced text — not that there is 85% certainty it was. The nuance matters, because it determines what you can do with the number.

Next to the score, a confidence level tells you what that estimate is worth on your particular text. Low confidence does not mean the tool did a poor job: it signals that this text is hard to call, usually because it is short, very technical, translated, or written in a careful academic style. In those cases the percentage must be read with a lot of reserve.

In practice three zones stand out. Below 40%, the text shows the irregularities of human writing. Between 40 and 70%, the signals are mixed and no conclusion is warranted. Above 70%, the style is clearly that of a generation — which remains a strong indication, never proof.

• the score measures a resemblance in style, not a certainty
• under 40%: irregularities typical of human writing
• 40 to 70%: mixed signals, no conclusion possible
• over 70%: a style clearly close to generated text

The detector is at its most reliable on texts of several paragraphs written freely, in prose: essays, articles, blog posts, cover letters, emails, reports. Those are the texts where a personal style has room to appear — or fail to appear. Five hundred words give a much steadier reading than three hundred characters.

It is less reliable on very short texts, on lists and tables, on texts that follow a rigid template such as legal notices or product descriptions, and on texts translated from another language. Translation in particular smooths out everything that makes writing personal, and the result looks generated even when a person wrote every word of the original.

Texts written by learners of English, or by people writing carefully in a language that is not their own, also tend to score higher than they should. Careful, correct, slightly stiff prose shares many surface features with generated text. The confidence level is there to flag exactly this situation, and it should be taken seriously when it reads low.

You do not have to copy and paste: the tool accepts PDF, Word (.docx) and plain text files up to eight megabytes. The text is extracted server side, then analysed exactly as if it had been pasted. A student essay, a report received by email or a downloaded article can be checked in two clicks.

One limit is worth knowing. A scanned PDF — a photographed or photocopied page saved as PDF — contains an image of the text, not the text itself. Nothing can be extracted from it, and the tool will tell you so. In that case, open the document, select the text if you can, and paste it directly.

Long documents are analysed up to twenty-five thousand characters, roughly ten pages. Beyond that, the text is truncated and the result says so. If you need to check a longer piece, analyse it section by section: that also gives you a more useful reading, since a document can mix human-written and generated passages.

Generated text has a recognisable texture once you know what to look for. Sentences tend to be of similar length, one after the other, with none of the very short or very long sentences that a person produces without thinking. Paragraphs are often of similar size too, and open with the same kind of transition: 'Moreover', 'Furthermore', 'In addition', 'It is important to note'.

The register never wavers. A human writer drifts — a touch more casual here, a sharper phrase there — while a generation holds one tone from the first line to the last. Generated text also favours fluent generalities over specifics: it will say that something 'plays a crucial role in various aspects' rather than name a place, a date or a figure.

Finally, generated text is rarely anchored. It has no anecdote that could only have happened to one person, no aside, no opinion that might be wrong, no reference to something the reader is assumed to know. It is correct, balanced and complete — and that completeness is itself a signal.

• very even sentence and paragraph lengths
• predictable transitions: 'moreover', 'furthermore', 'in conclusion'
• a register that never changes over the whole text
• generalities where a person would give a specific
• no anecdote, no aside, no opinion that could be wrong

Human writing is uneven. It has a four-word sentence next to a forty-word one. It digresses, then comes back. It uses a word twice in a row because the writer did not notice, or leaves a slightly awkward construction that a careful editor would have fixed. Those imperfections are not flaws in the eyes of a detector — they are evidence.

Human writing also takes positions. It says 'I think', 'this is wrong', 'nobody ever mentions'. It refers to specific things: a name, a street, a year, a price. It assumes the reader knows something and skips the explanation. And it changes register: an ironic aside in a serious paragraph, a formal sentence in a casual email.

The detector reports these markers with the same care as the generation markers, in the list under the score. A text that scores low should show several of them; a text that scores high should show few. Reading that list is often more instructive than the percentage itself.

A detector estimates the origin of a text from its style. It does not have access to the truth — nobody does, short of watching the text being written. That is why every serious study of AI detectors finds both false positives (human text flagged as generated) and false negatives (generated text passing as human), whatever the product and whatever its marketing says.

The models producing text also improve constantly, and each generation is harder to tell from human writing than the last. A detector calibrated on the output of last year's models will be less accurate on this year's. Ours analyses style markers that hold across models rather than the fingerprint of one particular system, which ages better, but it does not escape the general trend.

For all these reasons, the score on this page is an indication that can support a judgement, never a verdict that replaces one. Anyone deciding something important on the basis of a detector — a grade, a job, a publication — should treat the result as one piece of evidence among several.

A false positive is a text written by a person that the detector flags as generated. It is the most damaging error the tool can make, because it turns into an accusation against someone who did nothing wrong. Several groups of writers are disproportionately affected, and it is worth knowing who they are.

Non-native speakers come first. Writing carefully in a second language produces exactly the kind of correct, even, slightly formal prose that resembles generated text. Studies have shown detection tools flagging essays by non-native students far more often than those by native speakers. Students trained to write in a highly structured way, and professionals used to a formulaic house style, run into the same problem.

This is why the tool shows a confidence level and lists its observations instead of announcing a bare figure. A high score with low confidence on a formal, correct text is a signal to look closer, not a finding. If you are the author of a text that was flagged, the section further down explains what to do.

A high score on a student's work is a reason to look further, not a conclusion. Start by reading the markers the tool lists: do they describe a genuinely generated text, or a careful, formulaic essay? Compare the piece with the student's previous work. A sudden change of style, vocabulary or quality is more telling than any percentage.

The most reliable step is a conversation. Ask the student to explain a passage, to say where a particular idea came from, to summarise their argument without the text in front of them. Someone who wrote a piece can talk about it; someone who generated it usually cannot. Drafts, notes and document version history are also strong evidence, and students should be encouraged to keep them.

Many universities have revised their policies precisely because detectors were used as verdicts and produced unfair accusations. Using this tool as a first signal, followed by human judgement, is both fairer and more accurate than either alone.

• read the markers, not just the score
• compare with the student's earlier writing
• ask the student to explain or summarise the text
• look at drafts, notes and version history
• treat the score as a signal, never as a verdict

If you wrote a piece yourself and want to know how it will look to a detector, run it here first. A high score on your own writing usually means your prose is very regular: same sentence length, same paragraph shape, same transitions. That is not a fault in itself, but it is worth knowing before someone else runs the check.

The best protection is not to game a detector but to write in a way that is recognisably yours. Vary your sentence lengths. Use a concrete example where you would have written a generality. Take a position. Refer to something specific from your course, your reading or your experience. These are also, not coincidentally, the things that make writing better.

Keep your working files. Drafts, outlines, notes and the version history of your document are the most convincing evidence that a text is yours, far more convincing than a detector score in either direction. If you ever have to defend a piece, that is what you will need.

Editors and content managers use the detector to check submitted articles, guest posts and freelance work. The question is rarely 'was AI used' — it often is, legitimately — but 'was this text worked on by a person, or pasted straight from a generator'. The markers listed under the score answer that better than the percentage.

Recruiters run cover letters and written assignments through it, with the same caveat: a candidate who used a tool to polish a letter is not the same as one who generated the whole thing, and the detector cannot tell intent. Marketing teams check that copy written for them has a voice, and moderators of forums and review sites use it to spot batches of generated posts.

In every one of these settings, the tool works best as a filter that decides where to spend human attention, not as a machine that makes the decision.

People often ask whether the tool detects text from ChatGPT specifically. The analysis does not look for the fingerprint of any one product. It looks at style markers that generated text shares whatever the model behind it — uniformity, predictable structure, absence of anchoring. Text from ChatGPT, Gemini, Claude, Copilot, Mistral or a lesser-known model is analysed the same way.

This approach has a cost and a benefit. The cost is that it cannot say which model produced a text — no detector can do that reliably. The benefit is that it does not go blind every time a new model is released, as detectors trained on the output of one specific system tend to.

Newer models are, on the whole, harder to detect than older ones, and text that has been edited by a person after generation is harder still. A high score is therefore more informative than a low one: generated text can hide, but human unevenness is difficult to fake.

Ten analyses a day without an account, each on up to twenty-five thousand characters. That covers a teacher checking a set of essays, a student verifying a couple of papers, or an editor going through a day's submissions. No card, no email, no sign-up: the tool is usable the moment the page loads.

A free account multiplies both limits by four. The counter runs over a rolling twenty-four hours rather than resetting at midnight, so it cannot be gamed by using two days' allowance back to back, and the limit is applied by account and by network to keep automated use from crowding out everyone else.

There is no premium tier for the detector itself. The limits exist to keep the service fast and free for people, not to sell a bigger version.

The text is sent to the server, analysed, and the result is returned. It is not stored, not added to any database, and not used to train anything. The only record kept is a usage counter — how many analyses were run, from which account and network — needed to enforce the daily limits.

Uploaded documents are handled the same way: the text is extracted in memory, analysed, and discarded. The file itself is never written to disk. If you analyse confidential material, this matters, and it is one of the reasons to prefer a tool that states its policy plainly over one that does not.

The analysis itself is performed by a language model hosted by a third-party provider under a business agreement that excludes training on submitted data. Nothing you paste here becomes part of any model.

An AI detector and a plagiarism checker answer two different questions, and confusing them leads to wrong conclusions. This page answers 'does this text look like it was generated by a machine?'. A plagiarism checker answers 'does this text exist somewhere else already?'. A text can be entirely generated and match nothing online; it can be copied word for word from a website and score as fully human.

If your question is about copying — a passage that seems lifted from a source, a submission you suspect of being pasted from the web — the plagiarism checker on this site actually searches the web for the distinctive passages of a text and shows you the pages where they appear. That is a different tool, with a different method, and it is the right one for that question.

For a complete picture of a text you are unsure about, running both is reasonable: one tells you about the writing, the other about its provenance.

A score between 40 and 70%, or any score with low confidence, means the tool could not call the text one way or the other. That is an honest answer, and it happens often with short, technical, translated or very formal texts. Forcing a conclusion out of an ambiguous result is where most misuse of detectors begins.

If you can, get more text. A longer sample from the same author, or the full document instead of an excerpt, gives the analysis more to work with. Read the markers in the list: they may point to a template or a translation rather than a generation. And if the decision matters, rely on the methods that do not depend on a detector at all — comparison with earlier work, a conversation with the author, drafts and version history.

A detector is a lens, not a judge. Used that way, it is genuinely useful; used any other way, it does harm.

• an ambiguous result is an honest result, not a failure
• analyse a longer sample when you can
• read the markers for hints of a template or a translation
• for important decisions, use evidence that does not depend on a detector

Consider two short texts about remote work. The first: 'Remote work has become increasingly prevalent in today's world. It offers numerous benefits, including flexibility and improved work-life balance. However, it also presents challenges such as isolation and communication difficulties. Organisations must therefore implement effective strategies to maximise its advantages.' Four sentences of fifteen to twenty words, each opening with a familiar move, no specific anywhere, a conclusion that could end any paragraph on any subject. The detector scores it high, and the markers say why: uniform length, stock transitions, no anchoring.

The second: 'I've worked from my kitchen table since March 2020. The first year was fine. Then the kid started school online in the same room, and I discovered that 'flexibility' mostly means being interrupted at 10:40 every morning. We fixed it with a cheap folding screen from IKEA — I'm not joking — and a rule about headphones.' A three-word sentence, a date, a brand, an aside, an opinion. The detector scores it low, and again the markers say why.

Neither result is a verdict. The first paragraph could have been written by a tired student at midnight, and the second could be generated by a model asked to sound personal. But the example shows what the tool actually measures, and why a high score on a formal, correct text should send you to the markers rather than to a conclusion.

A growing number of tools promise to rewrite generated text so that it passes detectors. Some of them work, partly, by doing what this page describes as the markers of human writing: varying sentence length, removing stock transitions, inserting an aside. A text treated that way will score lower here, because it genuinely has more of the surface features of human prose.

What such rewrites rarely add is substance. A humanized paragraph still has no date, no name, no specific claim that could be wrong, because the tool that rewrote it does not know any. The detector's list of markers will show that: rhythm markers pointing towards human, anchoring markers still pointing towards generation. Reading the list, rather than the score, is how you see through a cosmetic rewrite.

This is also why the detector reports both directions. A score is a single number; the markers are an argument, and an argument can be examined.

In a classroom, the useful process is: run the detector on the set, look at the scores as a whole rather than one at a time, read the markers on the outliers, compare those with the student's earlier work, and talk to the student. In an editorial team, it is: run submissions, read the markers on high scores, and ask the author about the ones where the writing has no anchoring. In recruitment, the same, with the cover letter as the text and the interview as the conversation.

In each case the detector does one job — it decides where a person's attention should go first — and the person does the rest. That division is not a limitation to work around; it is what makes the tool useful without making it dangerous. A process that ends at the score is not a process, it is a coin toss with a percentage sign.

• run the set, read scores as a distribution
• read the markers on the outliers
• compare with earlier work by the same author
• have the conversation — that is where the answer is

AI detector, AI checker, AI content detector: the same thing under different names — a tool that estimates whether a text was generated. False positive: human text flagged as generated; the most harmful error. False negative: generated text passing as human. Confidence: how much the estimate is worth on a given text, shown next to the score here. Markers: the observable features of style that the estimate rests on, listed under the result.

Anchoring: the presence of specifics — dates, names, figures, lived details — that generated text tends to lack. Register: the level of formality of a text; generated text holds one register, human text drifts. Humanizer: a tool that rewrites generated text to add the surface markers of human writing. Plagiarism checker: a different tool, which asks whether a text exists elsewhere rather than how it was written.

• Free AI chat
• Free plagiarism checker
• Free grammar checker
• Paraphrasing tool
• AI humanizer
• AI image generator