What a Student Can Put in Front of a Panel: Version History, Drafts, a Walk-Through

Almost nothing counts as proof. A framework published in Frontiers in Artificial Intelligence in 2026 puts the limit plainly: process documentation should be read as evidence of engagement rather than proof of independent authorship, because process records can themselves be staged ¹. The University of Pittsburgh's teaching center makes the same point from the detector's side: its guidance states that detectors are not reliable enough to act as proof ². So a certificate isn't the goal.

A panel weighing an allegation is comparing two accounts of how a document came to exist. One is a percentage from a model the student cannot inspect. The other can include timestamps and earlier versions. It can include the sources the argument leans on and a conversation with the person who wrote it. Only one of those accounts can be questioned in the room.

The same framework lists what to keep while a paper is still being written: drafts, revision logs and reflective checkpoints ¹. Its reasoning is that such documentation is less easily outsourced ¹. That's advice for the week before a deadline. Reading it the week after an accusation is the uncomfortable part.

Four kinds of record, and they work together. Version history comes first, because the application sets its timestamps and the student doesn't, though the Frontiers framework warns that a revision history can itself be staged ¹. Next come dated drafts and reading notes, with the sources behind the argument. Earlier graded work in the same voice is third. Last is a walk-through of the argument.

In Google Docs, version history is visible only to people who can edit the file, so a document shared as view-only hasn't shared its history ³. Name a version to stop it being folded into a neighboring one; a file holds up to 40 named versions ³.

Two conditions decide whether that record exists at all. Google says deleting version history is permanent and cannot be undone ³. And Word keeps version history only for files stored in OneDrive or SharePoint, so an essay saved to a laptop's own drive has no Microsoft 365 version history to produce ⁴. A record nobody kept can't be handed over later.

In one published case a student did produce it. William Quarterman, a UC Davis senior, was referred for academic dishonesty after his exam went through GPTZero, and he produced the edit history of the Google document he'd written it in ⁵. It showed time-stamped evidence of the paper being built over two or three hours ⁶.

His family also showed that the same detector frequently flagged famous human writing, including Martin Luther King's I Have A Dream speech and passages from Genesis ⁶.

How to export a history, and what it can't settle on its own, is worked through in reading a document's version history as evidence.

Very little, and it can cost something. A second reading is a second estimate, and agreement between two estimates is not verification. Pittsburgh's teaching center notes that headline claims of 98 to 99 percent accuracy are common, that third-party evaluation finds high false positive rates, and that many vendors decline to publish a false positive rate at all ².

There's a practical problem too. A low reading doesn't delete the first one. Now the panel holds two numbers and has been given no method for choosing between them, and the student who produced the second number is the person with the most to gain from it. It arrives with the wrong fingerprints on it.

The deeper issue is that the first reading is hard to interrogate. Vanderbilt gave that as a reason for switching Turnitin's AI detector off in August 2023, writing that, to date, the vendor gave no detailed information about how it determines whether a piece of writing is AI-generated ⁷.

One use survives. Where two tools disagree on the same paper, the disagreement is an argument about detectors, not about one student. Put it in writing with the vendors' own published limits beside it, and keep it apart from the account of what happened the night the paper was written.

Start with four, all published by Turnitin itself. Turnitin puts its document-level false positive rate below 1 percent, and states that the figure holds only for documents its model reads as 20 percent or more AI writing ⁸. Its sentence-level rate is around 4 percent. That means roughly one highlighted sentence in twenty-five may be human-written ⁸.

The other two concern thresholds. Turnitin raised its minimum from 150 words to 300 in May 2023, so a reflection under 300 words gets no reading at all, which isn't the same as a clean one ⁹. It also added an asterisk to any score under 20 percent, marking the reading as less reliable in that range ⁹, and since 8 July 2024 that asterisk appears instead of the number, with no score and no highlights attributed below the threshold ¹¹.

The arithmetic a panel rarely hears is the one Vanderbilt did out loud. Using 1 percent as the ceiling on the below-1-percent figure, the university noted that 75,000 papers went to Turnitin in 2022, so around 750 could have been incorrectly labeled ⁷. The sum assumes every one of the 75,000 was human-written and that the vendor's rate holds on Vanderbilt's papers, so it is an upper-end illustration rather than a count. A rate that sounds small on one document is a large number of documents across a campus.

A fifth number matters where English isn't a first language. A 2023 Stanford study published in Patterns tested seven detectors on 91 TOEFL essays by non-native English writers. The detectors wrongly flagged more than half, an average false positive rate of 61.3 percent ¹⁰. The figure averages seven detectors as they stood in 2023, measured on TOEFL essays, so applying it to a different detector or a different student's paper is an assumption, and a panel should hear it as one. The full version of the Vanderbilt sum sits in the false-positive arithmetic a panel should do, and the TOEFL finding in what a non-native writer can bring to a flag.

Read any figure beside the population it was measured on. Turnitin says its below-1-percent document-level figure was validated on a test of 800,000 documents written before GPT ⁹. Human publishes one such figure, its own in-house measurement as of September 2026 and not an independent audit: 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%. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. Its public check reads up to 1,200 words, one paper a day, with no account, so it cannot read a whole long paper at once, and a second reading is still only a second estimate. Human's detector scores the text in memory and de-identifies any personal information for additional training. No rate, however low, stands in for the version history.

Quietly, and about the paper rather than the percentage. Turnitin's own guidance to instructors, published in 2023, is to use a highlighted sentence to start a conversation rather than to draw a conclusion ⁸. Put that sentence in front of any panel that has been handed a score and nothing else. The meeting is where the four records get used.

A walk-through works because it's specific. Which source changed the argument? What did the third paragraph look like before it was cut? Which claim was hardest to support? A writer can usually answer those in detail. Nobody can ask a score what it meant.

Quarterman's case ended about a month after the accusation, when the university dropped it after a written appeal ⁵. The university official talked with him and his instructor, did further research into indicators of AI-generated text, and concluded he most likely wrote what he submitted ⁵. The reporting does not say which piece of evidence moved the official, whose stated grounds were the two conversations and his own research ⁵.

Where a conversation has already hardened into a formal allegation, the procedural route is different, and it's set out in appealing an AI detection accusation. What to put in the first email, before any of that, is in what to say to a professor after an AI accusation.