Seven Questions to Ask a Student Suspected of Using AI

In order, they ask why this topic, what came out of the draft, which source disappointed, what a cited claim argues, how a key term is used and what the next paragraph would say, with the question about which tools were open saved for last. Each one asks about a decision. Decisions stay with a writer longer than finished sentences do, though three weeks thins both. The set runs about 10 minutes.

• "Why this topic, and what was the other option?" A real choice leaves a rejected one behind.
• "What came out of the draft, and why did it go?" Cutting leaves a memory a finished text does not show.
• "Which source was the most disappointing?" Checks whether the sources were read or only cited.
• "What does the third source actually argue?" Did the reading go past the abstract?
• "How is that term being used here?" Checks that the paper's vocabulary belongs to the writer.
• "What would the next paragraph say?" If the argument is still live in the writer's head, the next step comes easily.
• "Which tools were open while this was written?" Asks for the declaration plainly, once, with no trap in it.

Seven is a ceiling rather than a target, and the sequence counts for more than the total.

A meeting can answer itself by question three. Stop there, and don't count it as going easy. The last four are for the meeting where the early answers came out thin, and thin could as easily be nerves.

The job itself is narrower than it sounds. Harvard College lists the oral presentation among assignment designs, as a way to evaluate comprehension and let students explain their process, including how AI was used ¹. That is a design for a whole class. Borrowed for a follow-up, it points the same way: comprehension and process, with no confession asked for.

The first meeting is a different conversation with a different job. Run it as an examination and it spends the one thing it can produce.

With details, and often unimpressive ones. A student who wrote the paper names what went wrong. A source turned out to be about something else. A paragraph got moved twice and then deleted. The conclusion was drafted before the middle existed. Fluency isn't the signal. Detail is, about work that no longer shows up anywhere in the finished text.

The smooth answer is the one that should slow a teacher down. Someone working from a paper they didn't produce tends to paraphrase it back in its own vocabulary. That account has no rejected options in it and no dead drafts, because the finished text is the only record of the work that student has.

Nerves look like guilt. They aren't.

A 2026 framework in Frontiers in Artificial Intelligence makes oral defense one of four pillars of assessment built to survive generative AI, next to process documentation ². It also warns that process records can be staged, so they show engagement and fall short of proof of independent authorship ². The framework doesn't say it, but an interview can be rehearsed too.

A draft history is something examinable, and a version history read properly answers part of the process question a meeting cannot.

Send the questions ahead. Allow notes, accept a written answer, and hold the same meeting with students nobody suspects. Otherwise a quiet student loses an oral exam they ought to win. The point is to hear about the work, and nothing about that requires performing confidence in front of the person who holds the grade.

A student who writes plainly, or in a second language, may also find an oral meeting harder. No cited study measures that, so the design assumes it. Bloomberg Businessweek reported in 2024 that the students most likely to be wrongly accused are those who write in a plainer, more mechanical style. Neurodivergent students and second-language speakers are among them, according to the students, academics and detector developers it interviewed ³.

Word choice alone moved the reading in a 2023 Stanford study. Seven detectors read 91 TOEFL essays written by people and flagged an average of 61.22 percent of them as machine-written ⁴. Once ChatGPT rewrote the same essays for richer vocabulary, with the ideas and the writers unchanged, the figure was 11.77 percent ⁴. Those were 2023 detectors and test essays, so the figure shows the direction of the risk and gives no rate for any current product or classroom.

An oral meeting set up for one student may count as a policy or practice under disability law, where the student has a disability and the institution is covered. In the United States, Section 504 of the Rehabilitation Act of 1973 bars discrimination against a qualified individual with a disability in any program receiving federal financial assistance ⁵. Title II of the Americans with Disabilities Act of 1990 requires public entities to make reasonable modifications to policies and practices unless the change would fundamentally alter the service ⁵. Private institutions fall under a different part of the ADA, Title III ⁵. Neither law is about detectors, and none of this is legal advice. The institution's own procedure says where a request to change the meeting goes.

Cheaper still is an in-class sample from every student in week one. A writing baseline gives more to compare against, and it stays available all term.

Nothing in it proves use. A convincing account fits a student who wrote the paper. It fits one who studied a generated paper closely just as well. The University of Pittsburgh's teaching center states that detectors are not reliable enough to act as proof ⁷, and an interview built on top of a reading inherits that limit. The meeting narrows a question. It doesn't close one.

Turnitin's own 2023 note on sentence-level false positives tells instructors to use the information to start a conversation rather than to draw a conclusion ⁶. The seven questions are that conversation. Treat them as a test with a pass mark and they become the conclusion the vendor warned against.

It fails both ways.

A fluent performance doesn't establish authorship, since a student who turned in generated work can study it for an hour beforehand. A halting one establishes nothing either. Three weeks after a paper is finished, almost anyone's memory of it is thin.

The meeting changes one thing. The next step. That might be no referral at all, a second piece written under observation, a request for the draft history, or a referral with the student's own account attached. Sometimes the account and the paper don't fit together. Then a student's denial has a sequence of its own, and it doesn't start in this room.

The note takes a few minutes and carries five things: the date, the assignment, the questions as asked, the answers in the student's own words, and what was agreed next. Write it while the wording is still exact. If nothing further happens, that note is the only surviving record of the student's account, and it is still the record if the question comes back in March.

Record the questions verbatim. Don't summarize them. A summary is where a teacher's later belief leaks backwards into what was supposedly asked.

If a case does reach a panel, the account written that same afternoon is what the panel can rely on as a record of what was said. Memory of a 10-minute conversation isn't.

A meeting booked off a Human reading starts from a report that marks, sentence by sentence, how much of the paper reads as machine-written or machine-edited, with a document verdict of Human, Mixed or AI, and Human declines to score anything under 50 words. The public check reads up to 1,200 words, one paper a day, with no account; the school plan, for anyone with a school or university address, covers up to 50 papers a month. Human's detector scores the text in memory and de-identifies any personal information for additional training. Measured in September 2026 on its own held-out sets, in-house rather than an independent audit, the detector had 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%. A rate from those documents applied to one student's paper is an assumption. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. Whether a detector report is kept in a student's file is a records question for the institution. Keep the same-day note either way; of the two, only the note is written in the student's own words.