The Five-Part Detector Clause to Add to a School AI Policy Template

TeachAI's sample school guidance says two things about detection and disclosure. It tells teachers not to use technologies that purport to identify the use of generative AI, on the ground that their accuracy is questionable. It also requires that any use of an AI system by a teacher or a student be disclosed and explained ¹. That is a position on detection and a disclosure rule. A district that already licensed a detector needs a procedure instead.

Disclosure rules are the easy half to adopt, and they cover the honest case. Neither line covers the Tuesday afternoon.

A score comes back at 47% on a sophomore's essay. The teacher has 40 minutes before the next class, a student waiting in the hall, and no rule that says what a 47 means.

The gap is narrow and specific. Each of the 5 parts runs a sentence or two, drafted to drop into whichever template the district already adopted.

The first sentence of the clause decides every argument that follows, so it goes in ahead of any number. A workable version: a detector score is an estimate of how much of a document reads as machine-written, it is not evidence of what a student did, and no sanction follows from a score by itself. The rest of the clause is machinery for holding that line under pressure.

Vendor figures keep that sentence from reading as a courtesy. In a 2023 post, Turnitin put its sentence-level false positive rate at around 4%. Call that one highlighted sentence in 25 that may be human-written. The same post stated a document-level rate under 1% for documents marked 20% or more AI writing ³. Those are the vendor's own published figures, and they are from 2023.

Vanderbilt did a version of the multiplication. 75,000 papers went through Turnitin in 2022, and at 1% about 750 of them could have been labelled wrongly ². Each of those is a paper sent to review on the strength of a wrong label. That figure treats every submission as human-written and takes the vendor's upper bound as the rate. Both are assumptions, and a district doing its own arithmetic states them.

A separate question is whether any detector's rate holds across every writer. A 2023 Stanford study ran seven GPT detectors available at the time over 91 TOEFL essays written by non-native English writers. The average false positive rate came to 61.22%, against near-perfect accuracy on essays by US eighth graders ⁴.

Having ChatGPT rewrite those same essays with richer word choices cut the average to 11.77% ⁴.

So a machine-edited version of a human essay drew fewer flags than the human original. That is a flag tracking style rather than authorship.

Whether a district's licensed detector behaves the same way on its own English learners is not something that study measured. It is the reason to ask a vendor for a rate on that population, and a clause naming one flat number hides the question. The classroom version of that case runs through a non-native writer flagged by a detector.

A named person, and by preference not whoever has the score open. Review is the easiest step to leave out. It is also the step that turns a reading into a decision somebody can defend later. The City University of New York, a university system, has the faculty member who suspects a violation review the facts and circumstances with the student whenever feasible, before a report goes anywhere ⁵. A district clause can add a second reader.

What sits on the desk beside the score matters more than the score. Draft history from the document the student actually wrote in, an in-class writing sample from the same term, the assignment's own AI clause, a short conversation. Each of those is checkable by a third party. A detector reading is not corroboration of that kind: it can be re-run, but it cannot be explained.

Vanderbilt's objection to Turnitin was not only the rate. The vendor also gave no detailed account of how its tool decides a piece of writing is machine-generated ². That leaves an accused student little that is specific to rebut, and a clause that lets an unexplained number carry the case makes an appeal hard to argue.

The conversation is itself evidence, and how it gets run decides what it is worth. That is the step a flag most easily goes sideways on, and it is worked through in the oral follow-up after a flag.

Everything the reviewer saw, in writing, inside a stated number of school days. The notice names the assignment, the score, the detector, the setting it ran at, the specific sentences highlighted and the reviewer. One sentence in it says the score is an estimate rather than a finding. A notice that withholds the highlighted sentences is asking a student to argue with a number.

The deadline is the easiest line to leave blank. One workable pair is 5 school days from flag to notice and 10 from notice to decision, and the clause states those figures in days.

Promptly is not a deadline, and an open-ended case is a sanction nobody voted on.

One more line belongs in the notice: what the student may bring. Name draft history, a prior writing sample and a request to talk, and a letter that reads as a verdict starts to read as a question.

Parents arrive with the same question in a different register, and the front office reads from this same paragraph, which is the subject of explaining an AI detection policy to parents.

Someone who did not raise the flag decides it, on stated grounds, against a deadline. The two worked models below come from higher education rather than K-12, so a district adapts them rather than copying them. The City University of New York sets a floor worth borrowing for the hearing itself: written notice of the charges, the right to appear, and the right to present witness statements or call witnesses ⁵.

A detector case adds one more, which is the right to see the highlighted text and the setting that produced it. Stated grounds keep an appeal from turning into a second hearing on the same facts. The University of Texas at Austin limits a conduct appeal, under its Institutional Rules 11-801(d), to 3 grounds: significant procedural error, discovery of new information that was unknown or not reasonably foreseeable and was material to the decision, and a sanction significantly disproportionate to the violation ⁷.

A detector case argues most naturally from the first two. Draft history surfaces late. The procedural error a clause can prevent is a score treated as the case with no human corroboration gathered.

Then the record, which is where a cleared student can still lose. The file outlasts the hearing, and US federal law says what a parent or eligible student can do about it. The FERPA regulation at 34 CFR 99.21, as published in 2025, gives a parent, or a student who is 18 or enrolled in college, a hearing on request to challenge the content of an education record as inaccurate or misleading, once the school has declined a request to amend it ⁶. Where the hearing leaves the record standing, a written statement of disagreement may be placed in the file, and that statement travels with the record whenever the contested part of the record is disclosed ⁶. That describes the regulation and is not legal advice.

A clause silent on the record side leaves a cleared student carrying the flag anyway.

What a student is expected to do on the other side of this paragraph is set out in how an AI detection accusation gets appealed.

The threshold the district actually configured, carried into the policy text as a number. Name the tool. Name the mode, the percentage of machine content that triggers a review, and the date the setting was chosen. A threshold changed quietly in a vendor console is a policy change nobody ratified.

That number comes out of the contract rather than out of a template, and 5 questions decide what it is worth, asked of every vendor in writing before signature: what the false positive rate is, which population it was measured on, how large that set was, who ran the measurement, and what the tool does with short text and with text a person has edited. A rate with no cohort attached is a marketing figure, and both short and edited text turn up in a classroom.

Some of it is already published, and it is quotable as a vendor's claim rather than as a settled rate. In a 2023 post, Turnitin stated its document-level rate with a condition attached, under 1% for documents at 20% or more AI writing, and a separate sentence-level rate of around 4% ³. A district asks the vendor for the current figures in writing rather than quoting a post from 2023.

Vanderbilt disabled Turnitin's detector in 2023, and one of its stated reasons was that the vendor gave no detailed account of how the tool decides a piece of writing is machine-generated ². A vendor figure carries that same objection until the vendor explains how the tool decides, which is what those questions in writing are for.

The detector called Human publishes its own answer to the first of those questions at human.olive.is/research/model-card, with the date of the measurement beside it. As of September 2026: 0 false flags in 1,928 human documents it had never seen; the statistical ceiling on that is 0.155%. Human made that measurement in house, on its own held-out set, and nobody outside the company has audited it. Human is tuned for college-level academic writing, and that set is not a school district's marking pile, so applying the figure to secondary-school writing is an assumption, and a clause that borrows it says so.

Where the thresholds sit is published in the same form, and a clause borrows the wording rather than paraphrasing it: accusation-safe flags a paper above 15% machine content, standard above 6% (the default), sensitive above 2%. No setting turns an AI verdict into a Human one; an AI verdict needs at least 80% of the document to read as machine-written. This is an estimate from our detector. Treat a flag as a reason to look closer, not as a finding. A district that copies the mode and drops the 80% rule has written a clause implying a stricter setting settles something it cannot.

Whichever detector a district signs, the clause repeats that vendor's threshold and that vendor's cohort inside the policy text, so the next administrator does not have to go hunting for either one.