We are building a standing audit of how often detectors flag human writing. Open the Detector Observatory.
Try an example:
Scoring starts as soon as you type.
Never sent to a detector

Switch to Check Risk for a detectability score and the sentences behind it. No account, no upload to check, no word limit.

Free AI detector · the score, opened up

This AI detector shows the four measurements behind your score.

The score above measures how detectable your text reads, not whether a machine wrote it. This page takes the number apart: the four inputs, what each one is worth, and the sentence that produced every finding.

Reading the number

What the score means.

Four bands, with the cutoffs written down. Below 25 words the meters stay blank, because a rate computed over two sentences is noise.

0 – 14

Low

Nothing in the registry fired often enough to matter and your sentence lengths vary the way human drafting varies. A detector can still flag you.

15 – 34

Moderate

A handful of stock constructions, or rhythm that has started to flatten. Usually one or two paragraphs carry the whole score. The report names them.

35 – 54

Elevated

Tells are stacking: negation pivots, three-item lists, idle transitions, near-identical sentence lengths. This is the range where human writing gets accused most often.

55 – 100

High

The text reads the way generated text reads, whoever wrote it. Every contributing sentence is highlighted, so you see the reading before anyone else runs it.

The four inputs, and what each one is worth

Here is the whole composite, with the weights the engine uses. Tell density carries most of the score. Rhythm, vocabulary and hedging make up the rest.

When a passage has fewer than three sentences the rhythm reading is meaningless, so its weight folds back into tell density instead of scoring as zero risk.

58%
Tell density
Weighted count of registry matches per 100 words. High-severity patterns count for more than low ones.
22%
Rhythm uniformity
How little your sentence lengths vary. This is the same property commercial tools call burstiness.
12%
Stock lexicon
Rate of the vocabulary generated text over-produces: delve, leverage, foster, robust, myriad.
8%
Hedging
Rate of stacked qualifiers and distancing frames, the "it could be argued that" register.
Findings, not verdicts

Every finding comes with a reason.

The registry ships 44 detectors: 41 surface patterns plus 3 whole-document cadence readings. Each one highlights the exact characters that triggered it and carries a written explanation of why it fired, so you can judge the finding clause by clause and keep the lines you meant to write.

"not X, but Y" pivothigh · Structure

The negation pivot is one of the strongest single predictors of machine text. Say the positive thing directly.

Stock opener ("evolving landscape")high · Stock openers

Scene-setting that describes no scene. Detectors weight it heavily because almost nothing else opens this way.

AI lexicon: "delve"high · Stock lexicon

Post-2023 generated text uses this verb at many times the human rate. It is close to a fingerprint.

"it could be argued that"medium · Hedging

Six words of distance between you and your own claim.

Structure · 11Stock openers · 7Idle transitions · 5Stock lexicon · 14Hedging · 4Cadence · 3→ Read all 44 tells
Under the hood

How commercial AI checkers work.

Almost every AI checker on the market is built from the same two measurements. Here is what each one reads, and what it cannot see.

Perplexity

A language model reads your text and records how surprised it is by each next word. Predictable word choices give low perplexity. Generated text is predictable by construction, because the model that produced it was choosing likely words. So is careful, correct, textbook prose written by a person.

Burstiness

The variation in that surprise across a document, usually tracked alongside sentence length. Human drafting tends to lurch: a long clause, then four words. Generated paragraphs even out. So do dictated paragraphs, translated paragraphs and anything that has been through a grammar tool.

The output is a probability, not a measurement

Those numbers go into a classifier trained on examples someone labelled. What comes back is that classifier's confidence, printed as a percentage. It is not a reading taken off the text the way a thermometer reads a room. Nothing in a sentence records its own origin, so an AI checker is inferring, and it inherits every bias in whatever it was trained on.

The research

AI detectors get it wrong, measurably.

Stanford HAI researchers ran 91 TOEFL essays, all written by people, through seven commercial detectors. 61% of the essays by non-native English speakers were falsely flagged as AI-generated. Essays by native speakers came through nearly clean. Liang et al., 2023.

Figures in this build are illustrative. They model the shape of the published literature and are not measurements.
Detector · June 2026 runFalse positives · native EN· ESL writers· assistive tech
Detector A1.3%11.7%18.7%
Detector B2.2%19.7%24.1%
Detector C3.8%21.6%30.1%
Detector D2.5%15.2%21.4%
Preview of 4 of 9 detectors, across 4 writer cohorts.
Who gets falsely flaggedOpen the ObservatoryIf you write in English as a second language
Privacy

The checker runs in your browser.

Every detector that returns a percentage received your document to do it. This one never gets it. Scoring is JavaScript that runs in this tab, over a string that stays in this tab.

01
No request carries your text
Checking runs against a string in this tab. Open your browser network tab while you type and watch nothing leave.
02
Never sent to a detector
Your draft does not go to GPTZero, Turnitin, Copyleaks or anyone else, on this page or anywhere else in the product.
03
Works in airplane mode
Load the page, disconnect, keep checking. The score, the highlights and the reasons all come from code already in the tab.
Next step

Scored high? Here is what actually helps.

A high reading has two causes and they need opposite responses. Either the draft really is machine-flat and needs rewriting, or you wrote it yourself and a detector is about to be wrong about you. Work out which before you touch a word.

You drafted with AI and it reads like it

Fix the writing, not the score. The humanizer walks the highlighted clauses one at a time and rewrites them inside your own measured style, so the result sounds like you rather than like a different machine. You approve every rewrite before it lands.

Open the AI humanizer

You wrote it and you are still being accused

Keep your voice and bring proof instead. Draft in the editor and export a signed evidence report: your local edit history and drafting cadence, sealed with a key your browser generated. Anyone can check the signature without an account.

The measurements

What goes into the number.

Four readings make the score. Here is what each one counts, what it is worth, and what it reads in your draft. The result is a prompt to reread a paragraph, not a verdict on who wrote it.

Tell density · 58%

A weighted count of registry matches per 100 words. High-severity patterns count for more than low ones, so four soft hedges do not outweigh one "not just X but Y".

Rhythm uniformity · 22%

How little your sentence lengths vary across the draft. Commercial tools call this burstiness. Dictated and heavily edited prose evens out the same way generated prose does.

Stock lexicon · 12%

The rate of vocabulary generated text over-produces: delve, leverage, foster, robust, myriad. Counted per 100 words, so a long draft is not penalised for its length.

Hedging · 8%

Stacked qualifiers and distancing frames, the "it could be argued that" register. It is the smallest input and the fastest one to clear by hand.

AI detector: common questions

Is this AI detector free?

Yes. The checker on this page is free and has no account, no sign-up and no word cap. It runs in this tab, so the score and the reasons are never metered.

Does a high score mean a detector will flag my text?

A high score means your text carries the features commercial detectors key on: predictable phrasing, flat sentence rhythm, stock vocabulary. Those tools run their own classifiers and publish no thresholds, so read the score as a warning about your draft rather than a prediction of their output.

Does a low score mean I am safe?

A low score means the obvious tells are gone. Detectors still flag human writing, especially from second-language, autistic and assistive-technology writers, so the report also shows you the sentences a reader is most likely to question.

Do you upload or store what I paste?

The checker on this page is JavaScript running in your browser over your own buffer. No request carries your text while you check it, and your text is never sent to a detector. Open your network tab and watch.

Can an AI detector prove who wrote something?

No. Text carries no record of its own authorship, so every detector infers from surface statistics. If you need to show your work, the editor exports a signed evidence report with your local drafting history, and anyone can check the signature on the verifier page without an account.

Does it work offline?

Yes. Load the page once, then turn off your connection. The checker runs in your browser, so the score, the highlights and the reasons all come from code already in the tab.

Keep reading

Check it yourself.
The checker runs in your tab.

Free. No account. No word limit.

Open the Full ReportHumanize a Flagged Draft