AI Content Detector

Five measurable signals, each scored and shown separately, combined into a single probability and a three-way verdict. Because every signal is arithmetic on your text rather than an opaque judgement, you can see exactly which habit in your writing pushed the score up — which is the only version of this tool that is useful.

Five Signals, All of Them Arithmetic

Claiming to detect AI writing from a single number is the standard weakness of this category. A detector that cannot tell you why it decided something gives you nothing to act on, and one trained on a particular model's output ages badly the moment that model is updated. This one is built from five measurements that can each be checked by hand.

• Buzzword density: the text is searched for a compiled list of over-used terms, and the count per thousand words is scaled so that eight hits per thousand scores one hundred. A document with four of them per thousand therefore scores fifty.
• Em-dash usage: em-dashes and en-dashes are counted and normalised per thousand words against a ceiling of six, which is where a hundred is reached. This is the signal most likely to catch heavily edited machine output, since the long dash is a stylistic tic rather than a grammatical requirement.
• Sentence-length variety: with fewer than four sentences the signal abstains and returns fifty, because there is not enough data to judge. Above that, a mix of more than forty percent short sentences of six words or fewer and more than fifteen percent long sentences of over fifteen words scores ten — the human pattern. All-medium sentences, with fewer than ten percent long ones, scores seventy.
• Repetitive sentence openings: the most common first word across all sentences is found and its share of the total computed. From a share of twenty percent upward the score rises linearly to one hundred at sixty percent.
• Transition-word density: a seventeen-entry list of connectives is counted per thousand words and scaled against a ceiling of twelve, so a passage leaning heavily on the same connective phrases is penalised without any judgement about whether the sentence is good.

Reading the Result Without Misreading It

1. Paste at least a few hundred words. Short samples are the main source of nonsense results: the variety signal abstains below four sentences, and a density calculated over fifty words swings wildly on a single hit.
2. Analyse, then read the per-signal breakdown before the verdict. The verdict is a composite and the breakdown is what tells you whether a high score is one strong signal or five weak ones — these mean quite different things.
3. Look at the highlighted hits, not just the counts. A buzzword list that flagged "robust" three times is telling you something specific about the draft; the number alone is not.
4. If the verdict is mixed, treat the two closest signals as the finding. Mixed is the honest middle state, and it is common for genuinely human text that has been through a grammar checker.
5. Re-run after editing to confirm the signal actually moved. A score that does not respond to removing the flagged terms means you edited the wrong thing.

What the Three Verdicts Do and Do Not Mean

Alternative	Notes
Likely AI-written (70 and above)	Usually means several signals fired together — high buzzword density alongside uniform sentence lengths and repeated openings. Worth reading the breakdown, because the same composite score can come from a machine draft or from a human writer who has internalised corporate register, and the tool cannot distinguish those.
Likely human-written (below 30)	Indicates natural variation: a genuine mix of short and long sentences, few hedge words, varied openings. It does not certify authorship. Heavily edited machine output with a human pass over it scores here, which is precisely what it should do given that the edit is real.
Uncertain / mixed (30 to 69)	The most common outcome on real business writing, and the least useful as a label but the most useful as a diagnosis. Mixed usually means one or two signals fired hard. Find them in the breakdown and decide whether they represent a writing habit you want.
What none of them mean	None of these is evidence of misconduct. Every signal measures a stylistic habit that both machine and human writers exhibit. Treat the score as a description of the prose, not of its origin.

Two Passages and the Breakdown Behind Each Score

The composite number is the least informative part of the output. Two short passages show why.

• A four-sentence passage of even, medium-length sentences with no short ones returns a variety score of seventy on its own — the tool is telling you that uniformity is the pattern, not that the content is machine-made. A human writer who writes consistently is penalised here, and should know it.
• The same passage with two three-word sentences inserted changes the variety score to ten, dropping the composite substantially without a word of the substance changing. That is a legitimate improvement in readability and also a demonstration of how shallow the signal is: the score moved because the rhythm moved.
• A passage scored high on buzzword density but low on everything else points at vocabulary rather than structure, and the highlighted hits list the exact terms. Removing them is a mechanical edit with a real effect on the score.
• Note that neutral prose with no strong signal can land in the mixed band by default. If three signals sit near their midpoints and two abstain, the composite will trend toward fifty, which is why mixed is common and why the label alone is not a finding.

Sensible Uses and the Trap

• Editing your own draft before submitting it somewhere that screens with a detector, using the breakdown as a list of specific habits to change rather than the score as a target.
• Teaching writing. Each signal has a name a student can act on — uniform sentence length, repeated openings, transition-word overuse — which is more useful than a percentage.
• Checking whether a piece has drifted into template register after several rounds of editing, which is the most common way human writing acquires machine patterns.
• Not for accusing anyone. A detector built from surface features cannot establish authorship, and using one to make an allegation of misconduct will produce false positives on careful writers and false negatives on edited machine text.
• Not for academic integrity determinations. Institutions that do this use systems with provenance data and a formal review process, and the reasoning for that should be obvious from the breakdown above.
• Not as a benchmark between models. The signal set was written against a particular style of output and a different model will score differently for reasons that have nothing to do with how human the text is.

Questions About the Score

Can this tell me whether a text was written by AI?

No, and it does not claim to. It measures five surface habits — buzzword density, dash usage, sentence-length variety, repeated openings and connective density — and reports how much of the text exhibits them. A careful human writer with a uniform rhythm will score high; machine output that has been properly edited by a person will score low. The number describes the prose, not its origin, and any tool that tells you otherwise is overstating what text statistics can do.

Why does a short passage give a strange result?

Because every signal is a density or a ratio, and ratios are unstable on small samples. The sentence-variety signal explicitly abstains below four sentences and returns a neutral fifty instead of guessing; other signals will move sharply on one flagged word. Paste a few hundred words before reading anything into the output.

What is the fastest way to lower a high score?

Read the breakdown and fix whichever signal is highest, rather than editing generally. If buzzword density is the top signal, removing the highlighted terms is a mechanical edit. If sentence-length variety is the top signal, the fix is rhythm: add at least one very short sentence and let one run long. If repeated openings are the problem, the first words of consecutive sentences are the thing to change.

My text is not sent anywhere, is it?

Not in the default mode. All five signals are computed in the page from word and punctuation counts, so the text never leaves the browser and the analysis works offline. Note that for a detector this matters more than usual, since the whole point is to check sensitive unpublished writing before it goes out. Connecting an AI provider would transmit the text; the built-in signals do not need one.

Last verified: October 3, 2026

Launch AI Content Detector

All Tools · AI Tools · About · How We Test · Contact · Privacy

Maintained by Huang Bin · Last updated October 3, 2026