What to Write to a Professor After an AI Flag, and What to Attach

Five short lines, sent within 24 hours of the message that prompted them. The first names the assignment. After that: a plain statement that the work is the student's own, a question about which tool produced the reading and what the report shows, an offer of the drafting record, and a request for a short meeting. Nothing else belongs in a first reply.

• Name the assignment. "I got the note about my Week 9 essay and I want to respond properly."
• State it once. "I wrote that paper myself, and I can show how it came together."
• Ask for the reading. "Could I see the report, and which tool produced the score?"
• Offer the record. "I still have the version history, my research notes and two rough drafts, and I can send them over today."
• Ask for time. "Could I come to office hours this week to walk through the argument in the paper?"

The whole message runs shorter than the paragraph above it. It argues nothing about detection accuracy and it concedes nothing either, which is the point. An opening that debates detectors invites a reply about detectors, and the paper drops out of the conversation for a week.

Send it from the school account, keep a copy, and send it soon. A reply that arrives the same afternoon reads differently from one that arrives after the department has already scheduled a hearing, and nothing in those five lines is hard to write.

Two questions carry most of the weight: which tool produced the score, and what the report actually claims. Turnitin's 2023 guidance states that it makes no determination of misconduct, and that it supplies data for educators deciding under their own academic and institutional policies ². Its 2023 note to instructors is blunter still: use the information to start a conversation rather than to draw a conclusion ¹.

Both sentences come from the vendor, which is what makes them quotable. Asking to see the report is not a challenge to the professor; it puts the vendor's own guidance on the table where both people can look at it.

The report will not explain itself either. Vanderbilt's instructional technology office disabled Turnitin's AI detector in 2023, and one of the reasons it published is opacity: Turnitin gives no detailed information about how it determines whether a piece of writing is AI-generated ⁴. So a student answering a Turnitin percentage is rebutting a conclusion whose reasoning the vendor had not published as of 2023 ⁴. For another tool, ask what that vendor has published about how its score is reached.

What the number counts, and what it cannot carry alone, is set out in an essay that came back flagged.

The drafting record is evidence that existed before the flag did, and it is the easiest to hand over. William Quarterman, a UC Davis senior, was falsely accused after a professor ran his exam response through an AI detection tool. He answered with the time-stamped edit history of the Google document, which showed the paper taking shape over two or three hours; his family also showed the professor's tool flagging famous works as AI-generated, and he was eventually cleared ³.

What travels well in an attachment:

• The version history, shared as a link or exported to a file.
• Research notes, in whatever state they are actually in.
• The rough drafts, including the one with the opening paragraph that got cut.
• Source PDFs with the highlighting and margin notes still on them.
• Earlier graded work from the same course, which carries a voice the professor has already read.

Plenty of honest papers leave no trail at all. Written offline. Drafted longhand and typed up clean. Composed in a phone note and pasted into the submission box in one go. An empty version history is the absence of a record, not the presence of one.

Where the history is thin, the substitutes are the traces made while the work was happening. Library loan records. The download timestamps on the source PDFs. Browser history from the week of the deadline. A message to a friend complaining about the second source.

Assembling that file, piece by piece, is the record a student usually already has, and reading the curve of a document's growth is what a Docs version history shows.

The meeting is about the paper. CUNY's 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 ⁵. Arrive ready to talk about the argument, the sources and the cuts.

Four questions answer themselves when the paper is the student's own. Where the argument came from. Which source was hardest to use. What got cut between the second draft and the version handed in. What another week would have changed. None of those needs rehearsing.

Where a grammar checker was part of the drafting, name it in the meeting and say what it did. Grammarly's own support page says its red and blue underline corrections should typically not move Grammarly's own AI detection percentage, while using its generative assistant to meaningfully rewrite sentences and full paragraphs will raise it; the same page says its scores may differ from Turnitin's, GPTZero's and others' ⁸.

The apology is the thing to watch in the room itself. Thanking a professor for the meeting is ordinary courtesy. Apologizing, in a tense room, for something that did not happen can read as the beginning of a concession. Thank the professor for the time and leave it there.

Leave the accuracy research out of the opening. Seven detectors tested on 91 TOEFL essays written by non-native English speakers wrongly flagged more than half of them, an average false positive rate of 61.3% ⁶. That 2023 finding is real, and it answers a question about the instrument rather than a question about the paper.

Where the writer is a non-native English speaker, the same research is the reason to ask for a second reader, and it is worked through in a detector flagging a non-native English writer.

Where the institution has a written process, the case moves off the professor's desk and onto it. CUNY and the University of Texas at Austin both publish theirs. CUNY's 2024 policy gives a student who denies the allegation written notice of the charges, the right to appear before the committee, and the right to present witness statements or call witnesses ⁵. Those procedures are worth reading before the first meeting.

Appeals are narrow. The University of Texas at Austin limits a student conduct appeal to three grounds, listed on its appeals page under Institutional Rules 11-801(d). A significant procedural error. Information that was not known or reasonably foreseeable to the accused student. A sanction significantly disproportionate to the violation ⁷.

Read that middle ground carefully. A version history that existed all along is not new information under a test like that. The practical consequence is to put the whole record in at the first stage, while it still counts as evidence.

A student assembling that record can also weigh what a detector's own published figures are worth. Human, which reads college-level academic writing at human.olive.is, reported as of September 2026: 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%. Those figures are in-house: Human measured them on its own held-out sets, and no independent audit stands behind them. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. A ceiling calculated across a population of documents still says nothing about the single paper a committee is reading.

How an appeal is written, and which of those grounds a detector case usually sits under, is the appeal itself.

The record that settles these cases is almost always the one made before anybody was arguing.