AI detector for teachers

Every search for an AI detector for teachers is really two questions: can a tool tell me whether this essay was written by a student, and what am I allowed to do with its answer. The first has a clear answer from the published evidence, and it is no. The second is the one that matters, and the institutions that looked hardest at it in 2023 reached the same conclusion. This page lays out both, then gives a procedure for the flagged essay and one for the accused student.

What a detector's score measures

A detector measures how much a text resembles the default output of a language model: predictable word choices, even sentence lengths, signposted paragraphs, a formal register. That resemblance is real and measurable. It is also the profile of plain, careful, conventional writing, which is the kind of writing schools teach. Stanford researchers tested seven widely used detectors on 91 essays written by non-native English speakers for the TOEFL exam and found that the detectors flagged 61 percent of them as generated; on 88 essays by US eighth-graders the same tools were near-perfect. When the researchers enriched the TOEFL essays' vocabulary with ChatGPT, the false positive rate fell to 12 percent. The score measures vocabulary range and regularity. It does not measure authorship, and no later vendor improvement changes what is being measured.

The arithmetic that made Vanderbilt switch it off

Turnitin switched on its AI writing detection for its customers in April 2023, with less than a day's notice and no way to disable it, and claimed a false positive rate of 1 percent. Vanderbilt University's teaching centre did the only calculation that matters: the university had submitted 75,000 papers to Turnitin in 2022, so a 1 percent rate meant around 750 student papers a year wrongly labelled as partly written by AI, and the company had published no detail on how its tool reached its verdicts. In August 2023 Vanderbilt disabled the feature and wrote that it did not believe AI detection software was an effective tool that should be used. It pointed at the rest of the field for support: OpenAI had withdrawn its own classifier a month earlier for its low rate of accuracy, and other detector companies were pivoting or closing.

The rate is the vendor's claim on texts it chose; the Stanford test measured 61 percent on non-native writers. The arithmetic is the same at any rate: the wrong flags are real students.

Who in a classroom gets flagged most

The students whose writing is most regular: those writing in English as a second language, those who learned the five-paragraph essay well, those who use a grammar checker that smooths their sentences, those who write formally because they were told to. The Stanford finding is the measured case; the mechanism applies to all four. A teacher who runs every essay through a detector is, in effect, running a test that the most careful and the least native students fail most often, and that a student who asked a model to "vary the language" passes: the same study cut detection on generated essays from 100 percent to 13 percent with that one instruction.

When an essay is flagged: a procedure

Treat the flag as a question about the essay, never as a finding about the student. Vanderbilt's guidance to its own instructors after disabling the tool is the template: talk to the student first. Concretely:

If you are the student who was accused

You are in the position the arithmetic above predicts for several hundred students a year at a large university. What helps: collect everything that exists outside the text, the version history of your document, your notes and outline, your browser and search history for the days you wrote, messages to classmates or the teacher about the assignment, earlier essays in your own voice. Ask, in writing, which tool produced the accusation, what score, at what threshold, and what false positive rate the vendor publishes for writing like yours. Point to the Stanford study if English is not your first language, and to the Vanderbilt decision in any case. Ask for a conversation about the essay rather than about the score, and prepare for it the way you would for an oral exam: being able to explain and extend your own argument is the evidence that no detector can produce or refute. If your institution has an appeals procedure, use it; the burden of proof is on the accusation, and a percentage is not proof.

What to do instead of running a detector

Vanderbilt's advice to its instructors, and the practice that has spread since, moves the effort from after the essay to before it: say in the syllabus what use of AI is allowed and what is not; design assignments that ask for the student's own data, experience or class discussion, which a model cannot supply; collect outlines and drafts as part of the work; use short in-class writing to know each student's voice; and when something looks wrong, talk. None of this is new, and all of it is more accurate than a score. The detector on this site exists to show what the scores are made of; the honest use of it in a classroom is to paste in a text together and discuss why the meters move, which teaches more about writing than any verdict would.