How to tell if a student used ChatGPT without an AI detector

Signals that do not hold up

Most suspicions start with the prose. The paper is smoother than expected, the vocabulary is older than the student, the paragraphs all run to the same length, and something about it sounds like a machine. None of that is evidence, and every one of those impressions also describes a well-prepared student, a student who worked with a tutor, a non-native English writer who studied from textbooks, or a student who used a grammar tool the syllabus allows.

Style and vocabulary

Polished prose has no author signature. A confident register, a five-paragraph shape and a fondness for transitions are taught in every writing course. If a paper reads differently from a student’s earlier work, that is a reason to look at the process and to ask, not a reason to conclude.

Detector percentages

A detector score is an estimate about prose, and the research on its errors is consistent. Liang, Yuksekgonul, Mao, Wu and Zou at Stanford found in Patterns (2023) that GPT detectors flagged 61.3 percent of TOEFL essays by non-native English writers as AI-generated. Common Sense Media reported in 2024 that 20 percent of Black teens said they had been falsely accused of AI use, against 7 percent of white teens. Turnitin states a document-level false-positive rate under 1 percent for documents with more than 20 percent AI writing, which sounds small until you multiply it across every paper a department reads in a year. In Newby v. Adelphi University (2026 NY Slip Op 26021), a student’s 100 percent Turnitin flag was read as human by Grammarly and ZeroGPT, and the court found the finding “without valid basis” and ordered the record expunged.

Source: Liang et al., Patterns, 2023; Common Sense Media, 2024; Turnitin, “Understanding false positives within our AI writing detection capabilities”; Newby v. Adelphi University, Justia; all checked 2026-09-08.

A percentage, then, is at most a reason to open the process record. It is never the finding, and a case built on it alone tends to fail in the meeting or, increasingly, in court.

Signals that hold up

Process evidence

The strongest signal is a record of how the paper was written. If the student composed in a recorded editor, the replay shows whether the text was typed or pasted, where the pauses fell, which passages were deleted and rebuilt, and the order in which the argument formed. Generated text arrives in a paste, or it is retyped from another window in a steady rhythm with no deletions and no pauses, and either way it looks like transcription rather than composition. That is not a judgment about the prose. It is a description of how the document entered the editor, and the student can see the same replay you can.

Document history, with its limits

A Google Docs or Word version history is weaker but still useful. It shows that a document existed at intervals and how it changed between them. It does not distinguish typed text from pasted text, the author can turn it off, and it can be fabricated: auto-typer tools such as Phrasly Human Auto Typer replay pasted text as keystrokes into Docs or Word so the history looks written. Read a Docs history as one signal, and read a history with no revision, one large insertion, or a single session at three in the morning as a reason to ask, not as proof.

Source: learn.microsoft.com Q&A on Word version history; phrasly.ai/human-auto-typer; checked 2026-09-08.

A conversation about the argument

Ask the student to explain the thesis, why the second source was chosen over an obvious alternative, and what they cut. A student who wrote the paper can do this, imperfectly. A student who did not usually cannot, and the difference is obvious within a few minutes. Keep the conversation about the work, not the accusation, and it also becomes the best feedback the student gets all term.

Sources that do not exist

Check every citation. A fabricated source, a real author attached to a title they never wrote, a page number in a book that ends earlier, a quotation that does not appear in the cited work: these are concrete, checkable, and hard to explain. They are also the one kind of evidence a hearing will accept without argument, because anyone can verify it.

A decision sequence

1. Write down what raised the question. If it is only the prose or a percentage, note that, and do not contact the student yet.
2. Open the process record if one exists. Look at pastes, revision density, and the shape of the sittings. If the paper was written in a recorded editor, this step usually settles the matter one way or the other.
3. If there is no record, pull the document history and read it with the limits above in mind.
4. Check every citation against the actual source.
5. Count the independent signals. One signal means ask the student a question about their work. Two independent signals mean a formal conversation, and only then a referral if the conversation does not resolve it.
6. Write the first message to the student as a question about the paper, not as an allegation.

The first message to the student

The first message sets the terms of everything after it. Keep it short, keep it about the work, and do not name a tool or a score.

I have been reading your essay and I would like to talk with you about how it came together. Could you come to office hours this week, or suggest a time? Please bring whatever notes, outlines or earlier drafts you have. I want to go through the argument with you and hear how you built it. This is a conversation about the paper, and it is where I start whenever I have a question about a piece of writing.

Sample first message to a student

If the student wrote the paper, that message gives them the chance to show it and costs them nothing. If they did not, it gives them a fair chance to say so before anything becomes formal. Either way, you have not put an accusation in writing that you may later have to withdraw.

What not to do

• Do not send the student a detector screenshot. It invites an argument about the number and moves the conversation away from the work.
• Do not ask the student to prove a negative. Ask them to explain the paper, which is something a writer can do.
• Do not run the paper through several detectors until one agrees with you.
• Do not treat a refusal to meet, nervousness, or a weak explanation under pressure as a second signal. Those describe most students in a meeting with a professor.
• Do not skip the sources. It is the most concrete check and the one most often left undone.

What a record can and cannot settle

A keystroke-level record shows how text entered the editor: typed or pasted, in what order, with what pauses and rewrites. It does not show whether the ideas were the student’s, whether someone dictated the outline, or whether a patient student transcribed generated text slowly. It narrows the question and gives both of you the same evidence to look at. The finding is still yours to make, and the conversation is still where it gets made.