Five Checks in Order, and the Detector Is the Last One

Work in a fixed order, and put the software last. Four of the five checks produce evidence a student can answer: a comparison with earlier work, a reference list that either resolves or doesn't, a drafting history, a conversation. The fifth produces a reading the student cannot answer the way they can answer a question about their own sources.

1. Prior writing. Read the paper beside something the student wrote in class, under supervision, earlier in the term. 2. The references. Open two or three of them and confirm the article, the journal and the page range exist. 3. The drafting history. Ask for the version history of the file itself, not a screenshot of a panel. 4. A walk-through. Ask the student to explain one paragraph: the source choice, the counterargument, what got cut. 5. A detector, last. Read the score as a reason to look again at the first four checks.

No single check settles anything. Two that agree justify a meeting, and the meeting adds evidence; the department's procedure decides. A paper whose voice has changed and whose reference list does not resolve is one situation. A paper that reads oddly and checks out on every page is a different one.

Set the paper beside something the same student wrote under supervision earlier in the term. A short week tends to drop this check, and it is the one a detector reading a single file cannot make. In an ACL 2025 study, five annotators who frequently use language models for writing tasks beat most of the commercial and open-source detectors tested ¹. By majority vote they misclassified 1 of 300 articles, and they held up against paraphrasing and humanization ¹.

That study measured readers who know what model output looks like, not teachers who know a student. Its annotators had never met the authors and were voting on published-style articles rather than coursework, so it says nothing about comparing a paper with the same student's earlier work. A teacher who has read four of a student's paragraphs knows what that student sounds like. That is a separate kind of evidence, and the kind a student can answer.

Signals worth two minutes:

• Vocabulary that jumps a register the student has never used in class, with no matching gain in the quality of the argument.
• A thesis that contradicts what the same student argued out loud in seminar two weeks earlier.
• Citation formatting the course never taught, turning up correct and consistent for the first time.
• Habitual errors that vanish. A student who has misused the semicolon since September suddenly never does.

A five-minute in-class writing sample in week 2 is what makes this check possible in week 10, and without one the comparison runs against memory, which is not evidence anyone can show a dean. Collect the baseline early.

Open the citations the argument leans on hardest and search the title and the digital object identifier. The APA Style Blog records its own author asking ChatGPT for five supporting references ². The fifth was not a real article, and the digital object identifier given for it belonged to a different paper ².

One search each tests four things: a real author paired with a journal that never published them, a page range that does not exist in that volume, an identifier that resolves to something unrelated, a quotation that does not appear in the source it is attributed to.

A fabricated reference is strong evidence that nobody checked the source, which is an academic problem on its own terms, and much weaker evidence about who wrote the prose around it. Both findings belong in the meeting. Only one of them is about authorship.

It can show a paper being built, session by session. It cannot show that a paper was not built somewhere else. One case shows both halves. William Quarterman, a UC Davis senior, was referred for academic dishonesty after his professor ran his take-home exam through detection software that included GPTZero ⁹.

Among the things he offered in reply was the time-stamped edit history of the Google document, which showed the exam assembled over two or three hours ³. The university dropped the case about a month later, writing that it had no reliable evidence he had not written the text ⁹.

Two mechanics decide whether that evidence exists at all. Google's help page states that browsing earlier versions of a file requires permission to edit it, so a document shared read-only has not shared its history ⁴. The same page warns that deleting version history is permanent and cannot be undone ⁴. Word keeps version history only for files stored in OneDrive or SharePoint, so a paper written offline and emailed in has none to show ⁵.

Ask for access to the file itself, since a screenshot of a panel shows nothing that can be checked, and read the shape of the timeline rather than its length. A paper built across four sessions in a week looks nothing like one that appears in a single paste and is then tidied. The three patterns a Docs version history can show are worth knowing before the meeting, and what a teacher can see about copying and pasting covers the rest of that trail.

A missing history is not evidence of anything.

A conversation produces evidence a score cannot: the student's own account of how the paper was made. Turnitin says plainly what its own highlight is for: use the information to start a conversation rather than to draw a conclusion ⁶. The questions that work are about process.

Ask why that source and not the obvious alternative. Ask what the strongest objection to the thesis was, and what happened to it. What got cut from the draft, and why?

An answer that names the source that was rejected and the paragraph that was cut is evidence to weigh beside the other four checks, never a tell on its own. Hesitation proves nothing either. A student working in a second language, or one who has just been called to a meeting about honesty, can answer haltingly about a paper written alone.

Write the questions down beforehand and use the same ones for everyone whose paper reaches this step; that is what makes the conversation fair. Say what prompted the meeting, plainly, in the first sentence.

Because detector errors have been measured falling harder on some writers than others. Bloomberg Businessweek ran GPTZero and Copyleaks over 500 college application essays submitted to Texas A&M in the summer of 2022, before ChatGPT was released ⁷. Between 1 and 2 percent came back falsely flagged, and some of those flags claimed near total certainty ⁷.

A 2023 study points at who is exposed. 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 percent ⁸. The authors of that study caution against using detectors in evaluative or educational settings at all ⁸.

In a 2023 post, Turnitin put its own sentence-level false positive rate at around 4 percent, roughly one highlighted sentence in every twenty-five, and its document-level rate at less than 1 percent for documents with 20 percent or more AI writing ⁶.

Human, the detector this page puts fifth, reports one figure of this kind: 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 is Human's own in-house measurement, as of September 2026, on a held-out set of its own and not an independent audit. It says nothing separately about second-language, neurodivergent or plain writers, and a rate measured on one set of documents and applied to one teacher's marking pile is an assumption, worth saying as one. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating.

The order matters because of what each kind of evidence can survive. A student can answer a question about their own reference list, their own timeline, their own paragraph. A reading, whether a percentage or highlighted sentences, cannot be cross-examined. Whether a class should run a detector at all is a department decision, and it belongs somewhere other than a Tuesday afternoon with one paper open.