Reply to a Disputed AI Flag in Writing Before the Meeting

Two or three days ahead, a short note a parent can read twice with no teacher in the room. Four things go in it, and the fourth is a date. Everything else is left out, including anything about the student's character or about work handed in last term. A parent who reads it should be able to tell exactly what is being claimed and exactly what is not.

Writing it down first is not a formality. A case that exists only out loud gets assembled in the room, in front of a parent who arrived angry. Every reason added under that pressure sounds like a search for better grounds.

• the assignment, the submission date, and the grade as it currently stands
• the clause of the school's AI policy being applied, quoted rather than summarized
• what the detector reported, and the sentence saying what it does not settle
• what gets examined next, and the date the decision lands

Anything a teacher would not want read aloud at a hearing six weeks later stays out. Before quoting the policy, read it again: What the school's policy actually promises a parent sets out how far the handbook clause and the classroom practice can drift apart.

The wrong flag is common enough to be familiar. Ipsos surveyed 1,045 US teens aged 13 to 18 for Common Sense Media in spring 2024. 1 in 10 said a teacher had flagged their schoolwork as AI-written when it was not ³. Black teens reported it at about twice the rate of their White and Latino peers ³.

The parent on the phone may be right.

In plain terms, the tool reported how much of the text reads as machine-written, and that is a statement about writing rather than about a child. Turnitin's own guidance says the company does not make a determination of misconduct, and that the decision rests with the educator and institutional policy ¹. Quote that line in the meeting, and attribute it to the vendor instead of offering it as an opinion.

The number carries no estimate of intent, and it knows nothing about the student who submitted the file. Saying that plainly, in the teacher's own words, gives the parent a reason to move off the tool and onto the paper.

In June 2023 Turnitin put its sentence-level false positive rate at around 4% ², or roughly one highlighted sentence in 25 that the student may have written. On a page carrying twenty or so highlights, that rate makes at least one wrong highlight more likely than not, assuming Turnitin's test set is a fair stand-in for the paper in hand. Nothing on the page marks which highlight it is.

Bias is the question a prepared parent asks second, and it has a measured answer. Seven detectors tested on 91 TOEFL essays written by people wrongly labeled more than half of them, an average false-positive rate of 61.3% ⁴. A student who learned English in the last few years sits closer to that failure than a student who did not. The authors suggest the same holds for any writer with a narrow vocabulary, because the tools read predictable text as machine-made ⁴.

What a score alone can and cannot support is the policy version of the same question, and in some schools an administrator has already answered it in writing.

Read the highlighted sentence. That is all a highlight asks for.

Material the parent can inspect without trusting the teacher: the document's revision history, the drafts in the order they were saved, and writing the student produced in class. William Quarterman, a UC Davis senior, answered an academic dishonesty referral with the time-stamped edit history of his Google document, which showed the take-home exam built over two or three hours ⁵. His family put that beside a demonstration that the same tool flagged famous published texts, and he was cleared ⁵.

Drafts that grow paragraph by paragraph across several days say something a finished file cannot, and the order matters more than the volume. One upload at 11.47pm says less, and by itself it settles nothing about who did the writing.

Absence of a trail is not evidence of anything. A paper typed offline and mailed as an attachment often leaves no history a school can inspect. What a drafting history actually shows is narrower than most meetings assume.

The third kind of evidence is quick and travels well. Ask the student to walk through two paragraphs the teacher picked, out loud, with the paper closed, while the parent listens. A writer who has forgotten a source sounds different from a writer who never read one. A parent in the room can hear that difference without being told what to conclude from it.

FERPA gives a parent the right to inspect and review a child's education records at any US school that takes federal education funding, and the school has to comply within a reasonable period and in no case more than 45 days after the request ⁶¹¹. That right arrives with the dispute, not with the hearing. It sits with the parent until the student turns 18 or enrolls in a college. Then it passes to the student ¹⁰.

Where the school keeps the detector report with the student's file, it is likely an education record under 34 CFR 99.3, which covers records directly related to a student and maintained by the school or by a party acting for it ⁹. A report that sits only on the vendor's system is the harder question, and it turns on whether the vendor is acting for the school.

Amendment is the second right, and it is the one that changes how the note gets written. Where a school declines to amend a record a parent or eligible student says is inaccurate or misleading, the rules at 34 CFR 99.21 provide for a hearing to challenge the content of that record ⁷.

If the hearing leaves the record unchanged, the parent or eligible student may still place a written statement of disagreement in the file. That statement stays with the contested part of the record and goes out with it whenever that part is disclosed ⁷.

So the file gets written for a reader who disagrees with it.

This is US federal law as it stands in 2026, not advice about any particular case. State statutes and district policy sit on top of it. How an appeals process gets designed is the administrator's side of the same paperwork, and a teacher who has read it writes a better note.

On the date the note named, if the detector reading is still the only thing holding it up. Withdrawal means the grade is restored and nothing conditional is attached to it. Moira Olmsted, a student at Central Methodist University, had her grade changed and still carried away a warning that another flag would be treated as plagiarism ⁸. The case ended and the cost did not.

That warning is the sentence to refuse to write. Written into the file, it lowers the standard of proof for every paper that follows, for a reader who has none of the context.

The entry after a withdrawal is one line: the allegation was reviewed and not substantiated. No conditions, no summary of the conversation, no note about how anyone behaved in the meeting.

A teacher explains the reading twice, once in the note and once in the room, and that is easier when the tool's own description is short. At human.olive.is, Human reports how much of a piece of writing reads as machine-written, sentence by sentence, with a document verdict of Human, Mixed or AI, and it is tuned for college-level academic writing.

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 figure was measured in September 2026. It is Human's own in-house measurement on its own held-out set rather than an independent audit, and the research page at human.olive.is/research/model-card names the run behind it. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. A ceiling that low still leaves a single reading short of a case.

Length changes what any reading is worth. Human returns nothing at all below 50 words, and between 50 and 149 words the reading is weak because there is less text to work from. What another detector does with a short text is that vendor's page to answer.

A number describes writing, never the person who handed it in.