AI detectors do not read for meaning. They estimate how predictable each word is, and report low surprise as machine authorship. Second-language instruction teaches you to be predictable on purpose, because predictable is what can be marked reliably. So the writing that earns good grades in class is the writing a classifier scores as machine-made.
This page explains the mechanism and names the specific constructions carrying the risk in your own draft, so you can edit the ones you choose to and keep the English you worked for.
A detector has no way of knowing who sat at the keyboard. What it has is a language model, which it uses to ask how surprising each of your words is given the words before it. Text that a model would have produced anyway scores as machine-written. Text that keeps surprising the model scores as human. That is the entire mechanism, and it never once looks at authorship.
Now think about how you learned to write academic English. You were given a list of linking words and told to use them. You were given a paragraph shape and told to keep it. When two constructions were available you took the one you were certain was correct, rather than the idiomatic one you had only half heard. Every one of those decisions is good practice, and every one of them lowers surprise.
A first-language writer working quickly produces something messier. An idiom that lands slightly wrong, a sentence that runs long and then one that stops dead, a register that slips for half a line. None of that is better writing. It is just less predictable, and a perplexity-based classifier reads unpredictability as a human being. The writer with less control gets scored as more human. That is backwards, and it is the classifier's problem, not a reason to write badly.
This is also why the accusation is so hard to answer. You are being asked to explain why your writing is regular, and the honest answer is that you worked hard to make it regular. What helps is being able to point at the exact constructions that produced the reading, which is what the rest of this page is for.
Each card below is a real entry in the 44-pattern registry running in this tab. The label and the reason are read out of the detector itself. The line above them is the version you were taught. Both descriptions are accurate at the same time, and that is the trap.
Firstly, Secondly, Lastly is the standard way to signal that an argument has parts.
Ordinal adverbs are rare in natural prose and common in generated outlines.
Moreover and Furthermore get taught as the markers of a formal register.
These connect nothing. If the next sentence follows, it follows without an usher.
Therefore and Thus are how you were shown to introduce a result.
Formal-register connectives cluster in generated text far more than in human drafts.
Announcing the conclusion makes the essay structure impossible for an examiner to miss.
Readers can see it is the last paragraph. The label is pure filler.
A paired construction demonstrates grammatical control, so it earns marks.
A correlative pair that adds emphasis without adding information. Drop the scaffolding.
Keep sentences short and clear so a long one cannot get away from you. Safe advice that flattens rhythm.
Human drafts vary sentence length sharply. A flat length distribution is what perplexity-based detectors score hardest.
Nothing here says the construction is wrong. "Firstly" is correct English and always was. The claim is narrower: these patterns now carry risk, they carry it whether or not you used a model, and knowing which ones you lean on is better than guessing. See all 44 tells →
The editor makes the thing you are being judged on visible, sentence by sentence, in language you can act on. You keep the decision on every line.
Being flagged is worse when nobody will tell you what triggered it. The registry ships 44 named patterns, each highlighted where it occurs and explained in one sentence you can disagree with. Once a habit has a name, it is a thing you can decide about rather than a mystery you carry into the next essay.
The editor measures your own draft first: sentence lengths, vocabulary range, punctuation habits. Suggestions are scored against that profile, so a rewrite has to still sound like you to be offered. The goal is your English with fewer risky constructions, not somebody else's English pasted over yours.
There is no upload step, it keeps working with the wifi off, and your text is never submitted to a detector.
If you are already facing an accusation, the editor exports an evidence report of your local drafting history, signed with ECDSA over SHA-256 in your browser. Anyone can check it on the verification page against the public key the report carries, without contacting us.
Nothing is uploaded. Named patterns and three whole-document statistics, no model, no account, no upload step. You do not have to take that on trust: open your browser's network tab, paste a paragraph, and watch nothing leave. Or switch the wifi off and check anyway.
Your text is never submitted to a detector. Not to GPTZero, not to Turnitin, not to any of them. If you are a student on a visa with an integrity meeting in the diary, that is the part that matters: unsubmitted coursework never enters a third party's scoring queue in order to be checked here.
Evidence reports work the same way. The signature is generated in your browser with ECDSA over SHA-256, and an institution verifies it against the public key inside the report, on a page that never needs to contact us.
Adding mistakes on purpose to relax a classifier costs you marks with the human reader who is grading you. The edits that work are ordinary editing: vary the rhythm, let one sentence run long and the next stop short, cut the connectives that connect nothing. Your grammar stays intact.
Commercial detectors are retrained without notice and disagree with each other on the same paragraph, so the number they return is theirs to move. What you get here is the specific list of constructions producing the machine-made reading in your draft, and a decision on each one.
Our score measures how machine-made your text reads. Who wrote it is a separate question, and the product keeps the two apart. Use the score to edit. Use the evidence report and your drafting history when authorship is what is actually in dispute.
Because the thing they measure and the thing you were taught point the same way. A detector estimates how predictable each word is given the words before it, and reports low surprise as machine authorship. Second-language instruction is built to make writing predictable: a fixed set of linking words, a fixed paragraph shape, the safe construction over the idiomatic one. That is competence, and the classifier reads it as a machine. Stanford HAI's 2023 study measured this directly and found 61% of TOEFL essays by non-native speakers falsely flagged as AI across seven detectors.
Not as a rule. Use the ones that are doing work and cut the ones that are only there to sound formal. If the next sentence already follows from the one before it, "Furthermore" adds nothing but risk. If it genuinely contrasts, keep the contrast marker. The registry shows you every instance so you can make that call sentence by sentence instead of banning a word.
Run the paragraph and read the tells first, so you can point at the specific constructions that caused the reading rather than arguing in general terms. Then export a signed evidence report of your local drafting history from the editor. Your institution can check it on our verification page against the public key the report carries. All of that is free and needs no account.
No. Scoring and highlighting run in your browser with no account and no upload step. Your text is never submitted to a detector.
The named tells are English. The negation pivot, the stock opener, the idle transition and the announced conclusion are properties of English generated text. The rhythm and burstiness measurements are language-agnostic, so a draft in another language still gets a cadence reading, but the named patterns will not fire.
What it does is show you every sentence that reads as machine-made and the reason it does, so you can decide what to change. Cutting those constructions lowers the score this tool gives you. What a third-party detector then reports is its own call, on a model it retrains whenever it likes.
Back to Humanize AI, or read about the humanizer itself.