An AI detector for teachers
A named verdict and every sentence scored — evidence you can discuss.
50 free scans a day · 2,500 words per scan · submissions are scored in memory and never stored
After the flag
The workflow between
a score and a decision
Most departments hand teachers a detector and no process. The tools page you are reading is, deliberately, mostly process: this is the sequence that holds up — to an integrity board, to a parent, and to the student who didn't do it.
Triage privately
Run the submissions that worry you. Treat every result as a private note about where to spend your attention — not as a finding, and never as something to read aloud to a student as an opening move. The score selects which three papers out of eighty get your careful hour.
Read against what you know
Compare the flagged work with the student's earlier writing and in-class voice. Detectors can't hear a voice break between assignments or notice that an essay ignores everything discussed in seminar. You can, and that observation is stronger evidence than any percentage.
Ask for process, then talk
Version history, outlines, notes — a Google Doc that grew over days settles most cases immediately, as often in the student's favor as not. Then have the conversation: a student who wrote the essay can summarize its argument and rephrase a paragraph on request. The gap when they can't tells you more than the tool did.
Act only on convergence
A referral that will hold up — and be fair — rests on convergent evidence: the flag, plus absent process, plus a conversation that didn't add up. A score alone should never carry a sanction, which is also what every serious detector vendor, including this one, will tell you in writing.
Two syllabus lines make all of this easier: tell students what tools you use and what a flag triggers, and require version history on from the first assignment. The second converts most future disputes into a two-minute check — usually in the student's favor, which is the point.
Duty of care
Who false positives
actually land on
Every detector's errors — ours included — are not spread evenly across a classroom. Knowing where they concentrate is part of using the tool responsibly.
Non-native English writers
Peer-reviewed research has found detectors flagging writing by non-native English speakers at substantially elevated rates — simpler, more uniform phrasing resembles model output statistically. If your classroom includes second-language writers, your false positives will concentrate on them unless your process corrects for it.
Formulaic and highly regular stylists
Students trained into rigid five-paragraph structures, technical writers, and some neurodivergent writers produce naturally uniform prose that sits close to every detector's decision boundary. A flag on a student whose style has always looked this way is telling you about the style, not the semester.
The asymmetry of harm
A missed AI essay costs some fairness diffusely. A false accusation lands on one student, by name, with consequences — and the students most likely to be falsely flagged are often the least equipped to contest it. This asymmetry is why process, not scores, has to carry decisions.
What we will not claim
No detector can tell you who wrote an essay
What this page will not promise: that any detector can tell you who wrote an essay. Ours reports how strongly text resembles machine writing, with the evidence shown per sentence — and it has the same structural limits as the whole field: weaker on short and heavily edited text, and error rates that concentrate on non-native and highly uniform writers. Used to direct your attention and open conversations, it will earn its place in your workflow. Used as a verdict machine, any detector — ours included — will eventually wrong a student who deserved better.
Related reading for educators: how Turnitin's AI indicator works and what its score means, and honest accuracy answers on GPTZero and ZeroGPT, the tools your students are checking themselves with.
How to run a scan
Check the papers that worry you, not the whole stack into a spreadsheet
Paste the submission. 50 scans a day, up to 2,500 words each, no account, no procurement, minimum 50 words before the engine will score. You get a named verdict — Human-Written, AI-Assisted, or AI-Generated — and a score on every sentence, shaded in place. That is something you can put on the table. A bare percentage is not.
Use Normal mode when a false flag would hurt a student. Use Aggressive when missing a generated draft would hurt the course. Those are different errors. The report will not choose between them. Short passages, formulaic five-paragraph essays, and fluent English written as a second language sit close to the decision boundary. Treat those results as a reason to look, not as a reason to skip the looking.
Then follow the workflow on this page: triage privately, read against earlier writing and in-class voice, ask for version history and outlines, talk, act only on convergence. A Google Doc that grew over days settles most cases immediately, as often in the student's favor as not. A student who wrote the essay can summarize its argument and rephrase a paragraph on request. The gap when they cannot tells you more than the tool did.
We do not publish a headline accuracy figure. We do not publish pass rates against Turnitin, GPTZero, or anyone else. Different classifiers disagree on the same essay; that is normal. Text is scored in memory and never stored, which matters when you are pasting writing that is not yours, often by minors. The privacy policy is the longer version. Support is one inbox on the contact page.
Two syllabus lines make the rest of this easier: name the tools you use and what a flag triggers, and require version history on from the first assignment. Transparency deters the behavior you are screening for and makes any later process visibly fair. Detection works best as a known part of the course, not a trap. A score alone should never carry a sanction — which is also what every serious detector vendor, including this one, will tell you in writing.
Who this page is for
A detector that allocates attention, not a verdict machine
The SERP is full of tools and empty of process. This page is, deliberately, mostly process: the sequence that holds up to an integrity board, to a parent, and to the student who did not do it.
A grading stack you cannot read twice
Eighty papers, three that worry you. The score selects which three get your careful hour. It does not finish the hour. Run those privately. Do not read the percentage aloud as an opening. Point at sentences and ask how they were written.
A student whose earlier work does not match this voice
That observation is stronger evidence than any percentage. Detectors cannot hear a voice break between assignments or notice that an essay ignores everything discussed in seminar. You can. Combine it with version history before you treat the flag as a finding.
A classroom that includes second-language writers
Peer-reviewed research has found detectors flagging non-native English at elevated rates — simpler, more uniform phrasing resembles model output statistically. Your false positives will concentrate there unless your process corrects for it. A flag on a student whose style has always looked this way is telling you about the style.
Not: a spreadsheet of percentages as the gradebook
A missed AI essay costs some fairness diffusely. A false accusation lands on one student, by name. That asymmetry is why process, not scores, has to carry decisions. If your institution already runs Turnitin, this page is not asking you to rip it out. It is here for the papers you need to look at yourself, with evidence you can discuss.
FAQ
What teachers ask us
Cost, reliability, policy, student privacy, and how this fits alongside the tools your institution already runs.
Is this AI detector free for teachers?
Yes — the scanner on this page runs without an account or card: 50 scans a day, up to 2,500 words per scan, minimum 50 words to score. That covers a normal grading stack checked selectively — which is how detectors should be used anyway. Paid plans add volume for departments that need it.
There is no procurement, no LMS integration, and no seat licence. If your institution already runs Turnitin, this page is not asking you to rip it out. It is here for the papers you need to look at yourself, with sentence-level evidence you can discuss, and a process you can defend.
What does the report give me that a percentage doesn't?
Something you can put on the table in a student conversation. The verdict is one of three named categories — Human-Written, AI-Assisted, or AI-Generated — and every sentence is scored and shaded in place. 'This paragraph reads as generated, these two read as yours — walk me through how you wrote it' is a discussable observation. A bare '72% AI' is an accusation with no handle on it.
Under an AI-authored verdict every line is shaded, strongly where it gives itself away and softly where it does not, and never painted as human. Text run through a humanizer is built so that no individual sentence looks wrong. Painting those green would clear AI work line by line, which is the one direction this tool must never err in.
Can I rely on the verdict to report a student?
No — and you should distrust any tool that says yes. Every detector, including this one, has error rates in both directions, concentrated on particular kinds of writers. Use the report to decide where to look and what to ask. Let version history, drafts, and the student's own account carry the decision. Our verdicts are evidence to weigh, never proof.
What about AI-Assisted — is that cheating?
That's a policy question only your syllabus can answer, and it is the case most policies still miss. AI-Assisted means the text carries both signatures — a model draft edited by a person, or the reverse. Some courses permit it with disclosure; some prohibit it. What the verdict gives you is the accurate description of what happened, so your policy — whatever it says — can be applied to the actual situation rather than a rounded one.
How does this compare with Turnitin's AI indicator?
Different classifiers; expect different results on the same essay — that's normal across all detector pairs, not a defect. The practical differences: Turnitin lives inside institutional workflows and shows instructors a document-level percentage; this tool is free at point of use, needs no procurement, and shows per-sentence evidence. Our Turnitin explainer covers how its indicator works, what the score means, and its accuracy record in detail.
Do my students' submissions get stored or used for training?
No. Text is scored in memory and never stored. That matters for student work in particular — you are pasting writing that isn't yours, often by minors, and where it goes is a real question to ask of every tool in your stack. Here the answer is: nowhere.
Should I tell students I use an AI detector?
Yes, in the syllabus, along with what happens when something is flagged. Transparency deters the behavior you're screening for, makes any later process visibly fair, and — combined with requiring version history on from the first assignment — converts most future disputes into a two-minute check instead of a hearing. Detection works best as a known, fair part of the course, not a trap.
A syllabus line that names the tool and the process also protects you. A student who was not told that a detector is in use has a fair complaint even when the draft is generated. A student who was told, and whose Doc shows a night of work, has a two-minute exoneration. Both outcomes are the point of saying it out loud.
How much of a submission do I need to paste?
At least 50 words, and continuous prose rather than a bullet outline. The free tier is 50 scans a day at up to 2,500 words each. That is enough to check the papers that worry you, which is how a detector should be used — not as a blanket scan of eighty files so a percentage can be pasted into a spreadsheet. Confidence rises with length. A two-sentence abstract is a weak sample.
What if the student used AI for an outline or for grammar only?
That is a policy question, and it is the case most syllabi still underspecify. The detector will sometimes return AI-Assisted on writing that started as a model outline and was then drafted by the student, and sometimes return Human-Written on writing that was grammar-checked in a chatbot. Neither result tells you whether the course allowed that use. Write the rule down. Then use the report to see whether the text matches the rule you wrote, not a rule you implied.
Spend your careful hour on the right three papers
Free, no account, nothing stored — a verdict with sentence-level evidence, and a process you can defend.