AI detector, done honestly · Free · Nothing uploaded
AI Detector: the Signals, Not a Verdict
This page measures the signals AI detectors rely on and gives no AI percentage, because no detector is reliable enough to judge who wrote a text. Paste at least 100 words to see sentence length variety, vocabulary spread, repeated phrases, repeated openers and words that rose after ChatGPT, each explained. Nothing is uploaded.
The example on the tool: a formal paragraph varies its sentences 0.14, an informal note 0.88; on vocabulary the order flips.
No AI percentage, on purpose. No detector has been shown to tell reliably who wrote a text, and these figures describe style, not authorship. Never use them, or any detector’s score, to accuse anyone. What the studies found.
Nothing typed yet. The text is kept in this browser only.
| Signal | Your text | Second text |
|---|---|---|
| LengthWords, sentences and paragraphs | – | – |
| Sentence length varietyStandard deviation of sentence lengths over their mean | – | – |
| Vocabulary spreadShare of different words in each 50-word stretch, averaged | – | – |
| Repeated phrasesThree-word phrases used more than once | – | – |
| Most repeated openerShare of sentences that start with the most common first word | – | – |
| Words that rose after ChatGPTUses of nine word families two studies measured rising | – | – |
Sentence lengths, in order
Each sentence of your text appears here as a bar, its length in words.
Repeated phrases
- Shown once the text has 100 words.
Sentence openers
- Shown once the text has 100 words.
Words that rose after ChatGPT
- Shown once the text has 100 words.
Paste a text, and another to compare it with if you like. Every figure comes with what it describes and why it is not proof; nothing leaves this browser.
An AI detector is software that estimates whether a text was written by a language model, usually from how predictable its words are and how much its sentences vary. This page measures those features and stops there: it counts, it does not judge, and it must never be used to accuse anyone.
- No sign-up
- Nothing uploaded
- No AI percentage
- Every figure explained
- Works offline
Key facts
| Detail | Answer |
|---|---|
| Price | Free, no sign-up, nothing to install |
| Gives | Five measured signals with what each describes; no AI percentage or verdict |
| Minimum text | 100 words; shorter texts give unstable figures |
| Compare | A second text side by side, such as earlier writing by the same person |
| Not measured | Perplexity, which needs a language model |
| Language | English |
| Where the text is kept | This browser only |
| Evidence cited | OpenAI 2023, Liang et al. 2023, Weber-Wulff et al. 2023, Kobak et al. 2025 |
How to read the writing signals in a text
Paste, read each figure with its note, compare if it helps.
Paste at least 100 words
Into Your text. Below that, only the length is shown.
Read each signal
Every row gives the figure and a plain line on what it describes.
Look at the bars
Each sentence is drawn as a bar, so an even rhythm or a varied one shows at a glance.
Check the lists
Repeated phrases, the commonest openers and any of the nine listed words.
Compare, if it helps
Compare with another text adds a second column: another draft, or earlier writing by the same person.
Decide nothing about the writer
Use the figures to improve the writing. They cannot say who wrote it.
Who uses a signals check, and for what
People who want to know what a detector would look at, without being handed a verdict.
Writers checking their own drafts
Sentences of one length, a phrase used four times or ten sentences that start the same way, found before a reader notices.
Students worried about a detector
See the features detectors weigh in your own essay, privately, before you hand it in.
Students who were flagged
Set the flagged essay beside writing you did earlier, and see whether the figures match the way you write.
Teachers and markers
Understand what a detector’s score rests on before deciding whether it means anything, and why it should not decide a case.
Writers in a second language
See why plain, careful English is the style detectors misread, with the study that measured it.
Editors and content teams
Find repeated phrasing across drafts that went through many hands, human or not.
Confidential text
Reports, applications and anything else that should not be pasted into an online detector.
Teaching about AI
A lesson on how detectors work, and fail, with figures students can produce themselves.
The five signals: what each describes, and why none is proof
Detector companies say their tools weigh perplexity, burstiness and style. Two of those can be measured without a language model, and this page measures them and three more. Each is a fact about the text. None is a fact about who wrote it:
| Signal | How it is measured | What the value describes | Why it is not proof |
|---|---|---|---|
| Sentence length variety | The standard deviation of sentence lengths, in words, divided by their mean (the coefficient of variation). It is the usual stand-in for what detector companies call burstiness | Low: sentences of much the same length. High: short and long mixed | Instructions, summaries, legal text and writing taught to a pattern are even by design. Anyone can vary sentence length on purpose, and a model can be asked to |
| Vocabulary spread | The moving-average type-token ratio: the share of different words in every 50-word stretch, averaged (Covington and McFall, 2010). Unlike a plain ratio, it does not fall just because a text is longer | Low: the same words coming back within a few sentences | A technical text repeats its key terms because it must. Writers working in a second language often keep to the words they are sure of, and that is the pattern Liang and colleagues found detectors penalise |
| Repeated phrases | Every three-word phrase used more than once in the text, leaving out phrases made only of short function words such as in the and of a | A high count: a text that reuses its own wording | Repeating a key phrase is often good practice: a definition, a product name, the question an essay answers |
| Most repeated opener | The share of sentences that begin with the most common first word, from five sentences up | A high share: sentences built on one frame, such as It is… or Students… | Lists, instructions and lab reports start the same way on purpose, and so do many children’s writing exercises |
| Words that rose after ChatGPT | Uses of nine word families that two studies measured rising sharply in 2023 and 2024: delve, underscore, showcase, potential, findings and crucial (Kobak and colleagues, 15 million PubMed abstracts) and meticulous, intricate and commendable (Liang and colleagues, AI conference peer reviews) | A high count: a text using words that became much more common once language models arrived | Every one is an ordinary English word. The studies measured a rise across millions of texts; one use in one essay says nothing about that essay |
The example shows why no single figure decides anything. Try an example loads two texts of about 110 words each. The formal paragraph has a sentence length variety of 0.14, its sentences running from 9 to 14 words, and the informal note 0.88, from 1 to 34 words. On vocabulary spread the order flips: 0.87 for the formal paragraph and 0.78 for the note. A detector that turns figures like these into one percentage has to decide how much each counts, and it does not show you how.
Below 100 words the page shows the length only. Short texts give unstable figures: one extra sentence moves every number, and GPTZero’s own page says its accuracy is lower on paragraphs and single sentences than on whole documents.
How accurate are AI detectors? What the studies found
The strongest evidence on AI detectors comes from the people who built one and from independent tests. All of it points the same way:
| Source | What was tested | What it found |
|---|---|---|
| OpenAI, AI Text Classifier (2023) | OpenAI’s own detector, launched 31 January 2023 | Identified 26%% of AI-written text as “likely AI-written” and labelled human writing as AI 9%% of the time. Withdrawn on 20 July 2023 for its “low rate of accuracy” |
| Liang, Yuksekgonul, Mao, Wu and Zou, Patterns (2023) | Seven widely used detectors on 91 TOEFL essays by non-native English writers and 88 essays by US eighth-graders, all written by people | The TOEFL essays were called AI-written 61.22%% of the time on average; 89 of 91 were flagged by at least one detector and 18 by all seven. The eighth-grade essays were judged almost perfectly. When ChatGPT was asked to simplify their wording, the eighth-graders’ false positive rate rose from 5.19%% to 56.65%% |
| Weber-Wulff and colleagues, International Journal for Educational Integrity (2023) | 14 tools: 12 public detectors and two commercial systems, Turnitin and PlagiarismCheck | Found them “neither accurate nor reliable”, inclined to call text human-written, and easily thrown by light editing or paraphrasing |
| GPTZero (the maker’s own page, 2026) | Its own detector | Says no AI detector is 100%% accurate, accuracy is lower for paragraphs and sentences, and results “should not be used to punish or as the final verdict” |
| Kobak and colleagues (2024; Science Advances, 2025) | Word frequencies in more than 15 million PubMed abstracts, 2010 to 2024 | At least 13.5%% of 2024 abstracts were processed with language models, a figure for the whole body of abstracts. The method counts excess words across millions of texts and cannot say which abstracts they are |
Why non-native writers are flagged. Liang and colleagues traced the bias to the measure itself: detectors score text with low perplexity, meaning predictable word choices, as machine-like, and writing in a second language tends to use common words in common orders. Asking ChatGPT to make the TOEFL essays’ wording more elaborate cut their false positive rate from 61.22%% to 11.77%%, which shows how easily the measure moves with vocabulary rather than with authorship.
Detectors have been updated since 2023 and their makers report lower error rates, but those figures are the makers’ own. No independent, peer-reviewed study has shown a detector reliable enough to decide about one person’s text.
Why a small error rate still wrongs many people
A false positive rate sounds small until it meets a large number of honest people. Take a class, a department or a magazine’s inbox with 1,000 pieces of work, 100 of them written with AI, checked by a detector that catches 90 of those 100. The question that matters to the person flagged is not the detector’s accuracy but this: of the pieces it flags, how many were written by the person who handed them in?
| False positive rate | Where the rate comes from | Honest work flagged | AI work flagged | Flags in all | Flags that are wrong |
|---|---|---|---|---|---|
| 1% | A low rate of the kind detector makers claim for whole documents | 9 | 90 | 99 | 9% |
| 4% | The sentence-level rate Turnitin reported for its own checker in 2023 | 36 | 90 | 126 | 29% |
| 9% | OpenAI’s own classifier on human text, before it was withdrawn | 81 | 90 | 171 | 47% |
| 61.22% | The average for essays by non-native writers (Liang and colleagues) | 551 | 90 | 641 | 86% |
Even at 1%, 9 of the 900 honest writers are flagged, and one flag in eleven points at someone who did nothing wrong. The rarer AI use is in a group, the larger that share becomes: with 20 AI pieces in the 1,000 and the same 1% rate, 10 of the 28 flags are wrong. This is the base rate fallacy, and it is why a detector’s accuracy figure cannot tell you how likely one flagged person is to be guilty.
If your writing was flagged as AI
A detector’s flag is a number from a tool that its own makers say is not a verdict. If your writing has been flagged, these are the things that carry weight, roughly in the order to do them. This is practical guidance, not legal advice, and your school’s or employer’s own policy decides the process:
Ask what was used and what it said. The tool’s name, the score, and which passages it marked. A score with no tool named cannot be checked or challenged.
Show how the work was made. Google Docs keeps a version history (File, then Version history), and Word does for files saved to OneDrive or SharePoint. Notes, outlines, earlier drafts, browser history of your sources and the time stamps on files all show writing that grew over time.
Compare with writing you did before. Paste the flagged text above, press Compare with another text, and paste something you wrote earlier, ideally before 2022 or in an exam. If sentence variety, vocabulary spread and openers are much the same, that supports the plain explanation: this is how you write. It is evidence of a consistent style, not proof of authorship, and the same is true of the detector’s score.
Offer to talk the work through. Explaining your argument, your sources and why you chose them is something only the writer can do well, and many institutions treat that conversation as the real test.
Cite the evidence on detectors. OpenAI withdrew its classifier for low accuracy, Liang and colleagues found non-native writers flagged 61%% of the time, and GPTZero says its results should not be used to punish. All three are in the table above.
For teachers and markers. A detector score is a reason to look at a piece of work more closely, if that, and never a finding on its own. Several universities, Vanderbilt among them, have switched off detection features over reliability concerns. A conversation about the work, drafts and the student’s other writing tell you more than any score.
What this page does not measure, and why
Some things said to give AI writing away are measured by no one reliably, and some need tools this page does not load. This page leaves them out, and says why:
| Signal | What it is | Why it is left out |
|---|---|---|
| Perplexity | How predictable each word is to a language model, the measure most detectors start from | It needs a language model to score the text, which means sending the text to a server or downloading hundreds of megabytes. Its result also depends on which model scores it |
| A trained classifier | A model trained on labelled human and AI text that outputs a probability | That probability is exactly the verdict the studies above found unreliable, and it cannot explain itself |
| Watermarks | A hidden statistical pattern a model maker can build into its output | Only the maker can test for its own watermark; no public checker covers the models in use, and editing weakens it |
| Long dashes, “delve” lists, rule of three | Habits that circulate online as signs of AI writing | Of these, only words with a measured rise are counted, and with their source. Dash use and lists of three are long-standing human habits with no published study behind them as signals |
| Facts and sources | Invented references, wrong quotations, made-up figures | Checking them needs the sources themselves. An invented reference is a real problem in any text, whoever wrote it, and the citation generator shows what a complete reference contains |
What the five measured signals are good for is the writing itself. Sentences of one length read flat, a phrase used four times reads careless, and ten sentences that start with The read like a list. Those are worth fixing for the reader’s sake, which is the only reason to change your writing because of anything on this page.
This page, GPTZero, Paperpal or Grammarly
AI detectors compared on what their own pages say, checked 31 October 2026.
| This page | GPTZero (free) | Paperpal (free) | Grammarly (free) | |
|---|---|---|---|---|
| What you get | Five measured signals, each explained | An AI percentage, sentences highlighted | Human, mixed or AI bands, sentences highlighted | A basic label |
| A verdict on who wrote it | None, by design | Yes | Yes | Yes |
| Where the text goes | Nowhere; measured in this browser | Their servers | Their servers | Their servers |
| Free limit | None set | 10,000 characters a scan | 5 scans a day, 1,200 words each | About 1,000 words a check |
| Shows how each figure is worked out | Yes | Names the factors only | Names the methods only | No |
| Compares with your own earlier writing | Yes, side by side | No | No | No |
| Measures perplexity | No | Yes | Yes | Not stated |
More writing tools
Every writing tool on the site runs in your browser, needs no sign-up, and sends nothing anywhere.
Format & compare6
Sort, recase and compare text, generate words and format citations.
Alphabetize a list
A to Z word by word or letter by letter, by last name past Jr. and van, Mc with Mac, Room 2 before Room 10.
Open alphabetize a list →Compare text
Two texts line by line, words or characters marked, invisible and look-alike characters named, similarity %.
Open compare text →Capitalize sentences
First letters of sentences only, nothing lowered; Dr. and e.g. passed over, etc. listed to check, I and IJ.
Open capitalize sentences →Random word generator
Words from the named 7,776-word EFF list by length and letters, no repeats, the words that fit counted.
Open random word generator →Citation generator
APA 7, MLA 9 and Chicago 18 for books, articles, web pages and videos, the rule behind every part shown.
Open citation generator →Grammar checker
21 named rules in the browser: a or an, could of, their is, repeats, spacing, US and UK spelling; nothing sent.
Open grammar checker →
All 12 writing tools, by category: Counts & lengths (5) · Format & compare (7)
Frequently asked questions
How do AI detectors work?
Most score a text on perplexity, how predictable each word is to a language model, and burstiness, how much sentence length and structure vary, then pass those and other features to a classifier trained on human and AI text. The classifier outputs a probability, shown as a percentage. This page measures sentence variety and four other signals, and stops before the probability.
How accurate are AI detectors?
Not accurate enough to judge one person’s text. OpenAI’s classifier caught 26% of AI text and flagged 9% of human text before it was withdrawn in 2023. Liang and colleagues found seven detectors called essays by non-native English writers AI-written 61% of the time on average, and a 14-tool study called them neither accurate nor reliable.
Can AI detectors be wrong?
Yes, both ways. They flag human writing, most often plain, formal or second-language English, and they miss AI text that has been edited or paraphrased. Even a low false positive rate wrongs many people when most work is honest: at 1%, a detector checking 900 honest essays flags about 9 of them. GPTZero says its own results should not be used to punish.
Why does my writing get flagged as AI?
Usually because it is predictable in the way detectors measure: common words, even sentence lengths, a formal register and few personal touches. That describes much good academic and second-language writing. Paste your text above to see its sentence variety and vocabulary spread, and compare it with something you wrote earlier to see whether the figures match your usual style.
What if my original work is wrongly flagged as AI-generated?
Ask which tool was used and what it marked, then show how the work was made: version history in Google Docs or Word, notes, drafts and sources. Offer to talk the work through. You can compare the flagged text with your earlier writing here, and cite the studies on detector errors. Follow your institution’s process; this is not legal advice.
Are AI detectors biased against non-native English writers?
The best-known test found they were. Liang and colleagues (Patterns, 2023) gave seven detectors 91 TOEFL essays written by people: on average 61.22% were called AI-written, and 89 of the 91 were flagged by at least one detector. The same detectors judged US eighth-graders’ essays almost perfectly. Plainer vocabulary was the cause.
What is burstiness in AI detection?
Burstiness is how much sentence length and structure vary through a text. People tend to mix short and long sentences, and detectors treat an even rhythm as a sign of a machine. This page measures it as the coefficient of variation of sentence lengths and draws every sentence as a bar, but an even rhythm is also normal in instructions, summaries and formal writing.
What is perplexity, and why is it not measured here?
Perplexity is how surprised a language model is by each word: low perplexity means predictable text. It is the measure most detectors start from, but it needs a language model to compute, which means sending the text to a server or downloading hundreds of megabytes, and the result depends on the model used. This page measures only what it can count in your browser.
Why does this page not give an AI percentage?
Because no detector has been shown to produce one reliably, and a percentage looks like a finding. A score of 80% is read as an 80% chance someone cheated, which it is not. The signals are shown one by one with what each describes, so you can see what a detector’s score would rest on without being handed a verdict.
Is my text stored or sent anywhere?
No. Everything is measured in your browser, and the text is kept in this browser’s storage on your device so a reload does not lose it. Nothing is sent to a server, which matters for unpublished work and confidential documents. Clear empties the boxes, and Bring back the text restores them if you clear by mistake.
Related guides
Longer answers to the questions behind the numbers.
- Counting
Why word counts differ between tools
The same sentence is 26 words to one counter and 39 to another. Four counting rules side by side, and how to meet a hard limit whichever one a form uses.
- Reference
Types of counters explained
Synchronous vs asynchronous, mod-N, ring and Johnson, industrial and PLC counters, plus mechanical and software counters.