A Turnitin AI Flag Reaches the Instructor First; the Code Decides the Rest

The percentage lands in the instructor's report view, and the vendor's part ends there. Turnitin's 2023 guidance states that the company makes no determination of misconduct and instead provides data for educators to make an informed decision under their own academic and institutional policies ¹. The detector reports a reading. A person decides what it's worth, under the school's own rules.

Turnitin told instructors that year to use the information to start a conversation rather than to draw a conclusion ². In May 2023 the company added an asterisk to any document score under 20 percent, saying the reading is less reliable in that range, and raised the minimum document length from 150 words to 300 ³. For reports generated since 8 July 2024 it shows no score and no highlights for a reading between 1 and 19 percent, printing an asterisk in place of the number ⁹.

Scale matters before panic sets in. Of the 38.5 million submissions Turnitin had processed for AI writing by 14 May 2023, 9.6 percent came back above 20 percent AI writing and 3.5 percent between 80 and 100 percent ³. By subtraction, about nine in ten sat at or below the asterisk line, and Turnitin notes the counts include assignments where AI use was allowed ³.

What the percentage counts, and what a 30 percent reading does not claim about any one paragraph, is set out in what a Turnitin AI score actually means.

The instructor decides first, and the institution's code takes over from there. Three routes appear in the policies cited below: a question in office hours that ends the matter, a grade penalty applied by the instructor, or a referral to the office that handles academic integrity. The code says which of those is permitted, and whether an instructor may resolve a case alone. The vendor's product is silent on both.

Codes differ. The differences decide what a student's first week looks like. The City University of New York, in its 2024 academic integrity policy, asks a faculty member who suspects a violation to review the facts and circumstances with the student first, wherever that is feasible ⁴. Ohio State, on the integrity page its Office of Academic Affairs carried in 2026, routes alleged academic misconduct to a formal hearing before a standing university committee ⁵.

Two universities, two first moves, one flag.

Lowering a grade is the third path, and the way back from it is different again. Purdue's 2025 integrity page points a student whose grade has been reduced unfairly for alleged academic dishonesty to the grade appeals system ⁶.

A student who doesn't know which of the three routes is running can't tell which deadline is already counting down.

Sanctions come from each institution's code, not from the vendor, so no general list exists. What is possible at one school lives in that document and, where the code lets an instructor set a grade penalty, in the course syllabus. The student agreed to that code at enrollment; read it beside the syllabus.

One documented case gives the shape of it. Bloomberg Businessweek reported in 2024 on Marley Stevens, a student at the University of North Georgia, placed on academic probation for a year after a disciplinary hearing that followed a Turnitin flag. By her own account she had used only Grammarly's standard spell-checking and grammar features ⁷. A year of probation, after a hearing that found against a student who says she used only a spell-checker, is not a hypothetical.

What that case shows about process is worth separating from what it shows about the tool. The hearing produced the sanction; what it weighed beyond the percentage, the report does not say. No cross-institution rate for what a first AI case costs was found for this article, so a page quoting one should say where its number came from. The gap between a reading and a finding is the subject of why a detector score alone cannot support a finding.

Two documents answer this question for one specific student, and a search engine answers it for nobody. The syllabus states what the course permits and who to contact. The code sets out the process that follows a referral. Not every code states every step; ask the integrity office, in writing, about a missing one. Print both before drafting a reply.

Four lines to find, in this order:

• Whether the AI clause in the syllabus names grammar and spelling tools separately from drafting tools.
• Which office owns integrity cases, and whether an instructor may resolve one without referring it.
• What notice is owed before a meeting, and the deadline for a written response.
• The grounds an appeal is limited to, which are often narrower than the arguments a student wants to make.

That last line is the one most easily misread. The University of Texas at Austin, on its request-an-appeal page as accessed in September 2026, limits a student conduct appeal to three grounds under its Institutional Rules, subsection 11-801(d) ⁸. They are a significant procedural error; new information that neither the accused student nor the Dean of Students knew or could reasonably have foreseen at the time of the hearing, and that was material to the decision; and a sanction significantly disproportionate to the violation. A version history that existed all along is not new information under a test like that. So the record goes in at the first meeting rather than at the appeal.

How that filing is written, once a case has moved past the instructor's desk, is worked through in the appeal itself.

Answer in writing the same day, and keep it narrow. Ask which tool produced the reading and what the report shows. State once that the work is the student's own, then offer the drafting record: version history, research notes, rough drafts, earlier graded work from the same course. Accuracy research belongs in a later answer about the instrument rather than in the first message.

The drafting record is the part that cannot be created after the fact, which is why it goes first. Turnitin puts its sentence-level false positive rate at around 4 percent, meaning roughly one highlighted sentence in 25 may be human-written ². That figure argues about a tool. A time-stamped history argues about a paper.

The wording of that first message, and what to leave out of it, is set out in the email after an AI flag.

Human, a detector tuned for college-level academic writing, publishes a false-flag figure a student can weigh the same way, and it counts whole documents rather than the sentences Turnitin's rate counts. The arithmetic is at human.olive.is: 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%, measured September 2026. That is Human's own in-house measurement, on documents held back from training, and no independent audit has repeated it, which is one more reason the drafting record leads and the argument about instruments waits. Which run produced the figure is named on Human's research page rather than here. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating.

A reply sent the same afternoon can reach an instructor before a referral is filed.