Teachers detect AI writing in three ways, and only one of them is a tool. The first is a detector, usually an AI-writing indicator attached to the similarity service your work is already submitted through. The second is the marker reading the work. They have your earlier essays, they wrote the brief, and they sat through the seminar, so work that stops sounding like the student who handed in the last three assignments is the strongest signal anyone has. The third is the record around the submission: timestamps, save counts and version history the platform kept whether or not anyone looked. The detector percentage is the weakest of the three. It is an estimate about how the writing reads, there is no source document behind it, and two tools can disagree about the same paragraph on the same afternoon. The things that actually settle a case are the citation that does not exist and a conversation about your own argument. This page explains all three, what a percentage can and cannot support, and what you can check in your own draft before you hand it in. The version written for markers themselves is at for teachers, and it is a fair thing to point them to.
Most detection starts inside the system you already submit through. Similarity checking has been built into university submission portals for years, and several of those services now report an AI-writing figure next to the match report. Your instructor usually sees a number they did not calculate and cannot inspect, produced by a vendor that set the threshold somewhere above them. Whether your institution has it switched on, who sees it, and what number turns it into a flag are questions for your department, and they are better asked before a meeting than during one.
The second tool is ad hoc. An instructor who suspects something pastes a paragraph into a free web detector between classes, and whether that happens at all depends on the person marking. The third is not a detector at all. It is the record the platform kept anyway, and it is often more informative than either of the first two.
Most cases start with a person rather than a score. The marker holds your earlier essays, wrote the brief and sat in the seminar. Work that stops sounding like the student who handed in the last three assignments is the strongest signal anyone has, and getting to it costs one folder search, because nobody else holds that comparison.
The rest is the texture of the prose, and every item below is visible in your own draft for free. The registry of AI writing tells highlights the exact phrases as the page loads, and the AI vocabulary checker and the readability checker run in the same tab, unlimited, with nothing to sign up for.
Style raises a question. Content answers it. The fastest way a marker turns a suspicion into something concrete is to pull one citation. A reference that does not exist, or that exists and does not say what the essay claims it says, takes a minute to check and cannot be argued with.
The same goes for course-specific material. An essay that never touches the reading the seminar spent a fortnight on, or that applies a framework the module explicitly set aside, or that answers a question adjacent to the one that was set, is missing the things that were only ever said in the room.
The last check is a conversation. Which part gave you the most trouble, why did you cut the paragraph that used to be third, which source changed your mind. Somebody who wrote the essay answers in a sentence. Each of those checks produces something both sides can examine and discuss. A percentage does not, and in a fair process it carries less weight than students expect.
A detector score is an estimate that a passage looks machine-made, computed from writing style alone. A plagiarism match is a different kind of claim: it names the document it matched and quotes the line, so you can open the source and judge the match yourself. An AI reading has nothing to show you. There is no source document to show, and that difference is the whole story, set out in detail at AI detection vs plagiarism.
The threshold that turns the number red is a vendor's choice and is usually unpublished, so flagged means the score crossed a line somebody else drew. If a tool returns a similar share of flags in every cohort, that says more about the tool than about any student in it.
Then there is who gets caught by mistake. Detectors respond to prose that is clean, even and correct, which is what careful students, heavily taught writers and second-language writers produce. Liang and colleagues at Stanford HAI found in 2023 that 61 percent of TOEFL essays written by non-native speakers were falsely flagged as AI across seven detectors. Grammar assistance and dictation both flatten sentence rhythm, and a score reads the finished sentence without seeing which tool helped shape it. The full argument, with what to say if it happens to you, is at why detectors flag human writing.
Where the work was written in a platform that keeps version history, there is a record of how the document was built, and it is more informative than any score. A steady accumulation across four evenings reads differently from one paste at two in the morning.
It is not proof either way. Plenty of people draft in one application and paste the finished text into another, and plenty write in one sitting on paper or on a phone. A missing revision trail is not an admission, and treating it as one penalises a working habit rather than a behaviour. The version of this argument written for markers is at for teachers.
If you want a record you control, write in the drafting record tool. It logs timestamps, how much you added and cut, and whether anything large arrived in a single paste, then exports a report signed in your browser with an ECDSA P-256 key. Anyone you hand it to can check the signature in their own browser, without contacting us: it confirms the report has not been altered and that it belongs to that exact draft. No tool that reads finished text can prove authorship, so what this gives you in a meeting is specifics to point at instead of a general denial.
The part you control is knowing what your writing reads like. Paste at least 65 words into the checker and you get a score from 0 to 100 plus one of three readings: reads as AI-generated, borderline, or reads as human-written. A run costs one credit whatever the length, and the free allowance is 10 credits a day per network, with no account, no login, no card and no email address.
A score is a reading of style, not a verdict on who wrote the text. What it is good for is telling you whether your prose has drifted into the shape described above, at which point the free highlighters name the sentences and you fix them. The rewriter is there if you would rather do it in one pass: it rewrites every sentence from 25 words up to 12,000 characters a run, roughly 2,000 words, at 1 credit per 50 words.
One thing to be plain about. Rewriting a draft you were permitted to generate is ordinary editing. Where AI is not permitted, that rule covers the draft however it is edited. Read the policy that applies to you. If you were flagged on work you wrote yourself, the steps to take in the first hour are at you wrote it and were accused.
Sometimes, and rarely from a detector alone. The reliable signals are ones a person notices: work that does not sound like your earlier work, sources that do not check out, an essay that answers a nearby question rather than the one set. A tool can raise the question. It cannot settle it.
Several of the services universities submit through now report an AI-writing figure alongside the match report. Whether your institution has it switched on, who sees it, and what threshold turns it into a flag are questions for your department, and they are better asked before a meeting than during one.
Rewriting changes the surface of the text a score reads, so a number can move. It does not change whether the essay engages with the seminar, whether the citations exist, or whether it sounds like you, which is what a marker looks at. Where AI is not permitted, the rule covers the draft however it was edited.
Ask what tool produced the flag and what threshold it crossed, then bring what your process left behind: drafts, notes, sources, version history, and your ability to talk about the argument. False positives fall hardest on clean, careful and second-language writing, which is worth saying calmly and with the research to hand. There is a full script for that meeting at you wrote it and were accused.
It is a record of how the document was built, which is useful supporting context, but it is not proof of authorship. Gaps and large pastes are normal for people who draft elsewhere. Bring it as one piece of evidence alongside your notes, your sources and your account of the argument, rather than as something that closes the question on its own.