DetectorsFor teachers
Use the Detector the School Provides, and Keep Its Score Out of the Grade
Yes, when the school provides the tool and the teacher reads the score as a screen. A detector sorts a stack of papers into the few worth reading closely. Alone, it settles nothing. Turnitin's own guidance tells instructors to start a conversation rather than draw a conclusion, and GPTZero's FAQ says its results should not be used to punish students. Set the score that triggers a look before the term starts, and keep that number out of the grade book.
Olive Staff & OliveUpdated
Should a Teacher Run an AI Detector at All?
Yes, if the school already licenses one and the teacher treats it as a sorting step. Detection is already the common case: of 460 US middle and high school teachers surveyed in late 2023, 68 percent said they use an AI content detection tool regularly, up 30 points in a year 4. The question worth arguing is not whether to open the tool. It is what the number is allowed to do afterwards.
Teachers who use a detector regularly were likelier to say students at their school had got in trouble for using or being accused of using generative AI, 72 percent against 48 percent of those who do not 4.
That gap does not prove cause. It is a reason for care: a number can harden into a case whether or not it was strong enough to carry one. The decision a teacher makes is procedural rather than technical. Run it on everything or on nothing, because running it only on the papers that feel wrong turns a hunch into evidence and hands the hunch a percentage to wear.
What Do Turnitin and GPTZero Publish About Their Own Limits?
Both publish limits. Both are worth reading before the first flag of the term lands. Turnitin states its document-level false positive rate, below 1 percent, for documents with 20 percent or more AI writing 1. Its sentence-level rate is around 4 percent, so roughly one highlighted sentence in twenty-five is human-written 1.
- Turnitin added an asterisk in place of a percentage under 20 percent in May 2023, calling the score less reliable in that band, and raised its minimum to 300 words in the same update 2.
- GPTZero reports that in its high confidence band 99.1 percent of human articles are classified as human and 98.4 percent of AI articles as AI, and its FAQ says those results should not be used to punish students 3.
- Turnitin tells instructors to use a highlighted sentence to start a conversation rather than to draw a conclusion 1.
Put those three together and the shape of the tool is plain. A published rate describes a population of documents, and a teacher grades one student at a time. As of that 2023 update, under 300 words there was no reading at all and under 20 percent an asterisk instead of a figure 2; above those lines the document-level rate holds, while the sentence highlights a teacher reads carry the higher error rate, one in twenty-five 1. Every one of those caveats comes from the vendor's own page, and the overstatement happens downstream, in how a number gets read.
Set the Review Threshold Before the Term Starts
Write down the number that triggers a second look. Write it down in August, before any paper has come back. A threshold chosen after one paper comes back at 43 percent is a threshold chosen about one student. Turnitin's asterisk under 20 percent is a defensible floor to build on, since the vendor called its own reading less reliable in that band when it added the asterisk in 2023 2.
The threshold decides attention, not guilt: it is the point at which a teacher spends fifteen minutes on a paper instead of four, and that time goes on the version history and a conversation about the argument. What comes next deserves its own planning: the first move after a flag is never an accusation.
Tell the class the number in week one, and it stops being a trap.
Which Students Did the Detectors Get Wrong in the Stanford Test?
Non-native English writers, in the one study that compared. Seven detectors tested by Stanford researchers in 2023 were near-perfect on essays by US eighth graders. The same detectors misclassified more than half of 91 TOEFL essays by non-native English speakers, an average false positive rate of 61.22 percent 5. Rewriting those same human essays with richer vocabulary cut the rate to 11.77 percent 5.
That second figure is the one to carry. Vocabulary range moved the verdict while the authorship stayed exactly the same, which suggests the flag was tracking the prose rather than who wrote it. The study did not test students whose prose is plain for other reasons, so extending the finding to them is an inference from its explanation and not a result of it. Nor did it measure the detector a school licenses today, which makes an English learner a live risk to plan around rather than a finding to cite.
A separate survey points the same way. 1 in 10 teens told Common Sense Media that a teacher had flagged their schoolwork as AI-written when it was not 6. Black teens reported being wrongly flagged at about twice the rate of their White and Latino peers 6. The survey does not say whether a detector or a teacher's own judgment produced those flags. That pattern falls along lines a classroom can predict, and it is an argument for care in how a flag is used, not for pretending the tool says nothing.
Keep the Score Out of the Grade Book
A detector number never belongs in the grade book: a percentage can open a conversation and it can never close one. Vanderbilt's own arithmetic in 2023 applied Turnitin's stated 1 percent to all 75,000 papers it submitted in 2022, about 750 7, and that arithmetic, with other concerns, led the university to disable the tool. The same sum for one teacher with 120 essays a term assumes every paper is human-written and that a vendor's rate transfers to that class.
Newer in-house measurements exist, and none of them turns a score into evidence. A threshold can at least be argued against a measured ceiling, and one detector tuned for college-level academic writing publishes its own, measured September 2026: at human.olive.is, Human records 0 wrongly flagged by Human's default model of 600 human-written papers it had never been trained on; the ceiling on that is 0.498%. That is Human's own in-house measurement on its own held-out set, not an independent audit. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. Applying that figure to a middle or high school marking pile, or to another vendor's score, is an assumption rather than a transfer.
The rule that survives a real term is easy to state and hard to hold: the score decides who gets a conversation, and the conversation plus the draft history decides everything after that. Whether detectors are reliable enough to rest a case on is a different question from whether one is useful, and the district version of the same decision is a third.
What a teacher explains to a parent is the draft history and the conversation.
Common questions
Is it fair to run a detector on some papers and not others?
No. Running a detector only on the papers that feel wrong converts a suspicion into a record, and the record then looks like independent corroboration of the suspicion that produced it. Either every submission in an assignment goes through the same screen, or none does. A uniform screen can be announced in advance, and that is what makes its result answerable.
What if the school provides no detector?
The practice still works without one. A detector only picks which papers get a second look, and a teacher without one picks by reading. What settles a question was never the score: an in-class writing sample from week one, a draft trail in a shared document, and a conversation about the argument. Student work should not go into a free web detector on a teacher’s own account, which is a tool the school has not licensed.
What AI percentage should trigger a conversation?
Where the department sets that number, the department's number governs; where it has not, the teacher writes one down before the term. Turnitin called its own reading less reliable under 20 percent when it added an asterisk there in 2023 2, which makes 20 percent a defensible floor rather than a rule. Whatever the number is, it buys reading time, never a grade change. Publishing it in the syllabus removes the later argument about whether it was chosen for one student.
Can a detector score support an academic integrity finding by itself?
No. A finding needs evidence a person can examine: the draft history, an in-class writing sample, a conversation about the argument. The vendors draw the same line themselves. Turnitin tells instructors to use a highlighted sentence to start a conversation rather than to draw a conclusion 1, and GPTZero's FAQ says its results should not be used to punish students 3. The score picks the paper, and everything after that is ordinary teaching evidence.
Do AI detectors flag English learners more often?
One 2023 study says yes for the seven detectors it tested. Those detectors misclassified more than half of 91 human TOEFL essays, an average false positive rate of 61.22 percent 5, and were near-perfect on essays by US eighth graders. Rewriting the same essays with richer vocabulary cut that rate to 11.77 percent 5, which suggests the flag was tracking vocabulary range rather than authorship. Whether the detector a school licenses today behaves the same way on English learners is not something that study measured, and treating it as a live risk is a precaution rather than a finding.
What should a teacher set up before the first flag of the term?
Two things, both cheap. An in-class writing sample in week one, so a baseline exists in the student's own hand, and work that leaves a trail, such as drafts written in a shared document with version history. Neither depends on which detector the school bought, and both survive the detector being wrong. The threshold and the next step go into the syllabus at the same time.
References
- 1.Understanding the false positive rate for sentences of our AI writing detection capability Turnitin, 2023. turnitin.comTurnitin's document-level false positive rate below 1 percent for documents at 20 percent or more AI writing, its sentence-level rate of around 4 percent, and its instruction to start a conversation rather than draw a conclusion.
- 2.AI writing detection update from Turnitin’s Chief Product Officer Turnitin, 2023. turnitin.comThe asterisk on AI scores under 20 percent and the 300-word minimum for a reading.
- 3.Frequently asked questions GPTZero, 2026. gptzero.meGPTZero's high confidence band figures of 99.1 percent and 98.4 percent, and its statement that the results should not be used to punish students.
- 4.Up in the Air: Educators Juggling the Potential of Generative AI with Detection, Discipline, and Distrust Center for Democracy & Technology, 2024. cdt.org68 percent of 460 surveyed US middle and high school teachers using a detector regularly, and the 72 against 48 percent gap in reported student discipline.
- 5.GPT detectors are biased against non-native English writers arXiv (Stanford University; published in Patterns), 2023. arxiv.orgThe 61.22 percent average false positive rate on 91 TOEFL essays, and the drop to 11.77 percent when the same essays were rewritten with richer vocabulary.
- 6.The Dawn of the AI Era: Teens, Parents, and the Adoption of Generative AI at Home and School Common Sense Media, 2024. commonsensemedia.org1 in 10 teens reporting a wrong AI flag on schoolwork, and Black teens reporting it at about twice the rate of White and Latino peers.
- 7.Guidance on AI Detection and Why We’re Disabling Turnitin’s AI Detector Vanderbilt University, Brightspace and Instructional Technology Support, 2023. vanderbilt.eduThe 75,000 papers submitted in 2022 and the roughly 750 that a 1 percent error rate would have mislabelled.
7 sources, numbered by first appearance.
General guidance for teachers, administrators and students. What holds at one institution, on one assignment, may not transfer to another.
Human checks whether a piece of writing can be verified as written by a person, and shows the sentences behind the result. It does not report a probability that a person used AI, it does not check for plagiarism, and no number it produces stands for a student's honesty. A flag should be reason to look closer and not verified proof of cheating.