How to have the "I think you used AI" conversation without wrecking trust

The stakes, stated plainly

This conversation has two possible errors, and they are not symmetric. If a student cheated and talks their way through, you lose one data point — the next assignment gives you another. If a student didn't cheat and leaves your office having been treated as a liar, you may have permanently broken their trust in you, in the course, and in the institution — and if they are one of the students detectors systematically mis-flag (a 2023 Stanford-led study found non-native English speakers' genuine writing falsely flagged at an average rate of 61.3% across seven detectors), you have also confirmed their worst suspicion about how they are seen.

The student in the best-known false-accusation story — a UC Davis student cleared by his Google Docs history — said afterward that what he wished for was simple: that his professor had asked him about the work before treating a score as a conclusion. That is the whole design principle of this guide. The conversation is an inquiry, not an ambush.

Before you schedule anything

Check what you actually have. A detector score is a statistical estimate, not an observation — the vendor itself says it should not be the sole basis for action. Before raising it with a student, write down what, beyond the score, made you concerned: a voice shift from earlier work, claims no source supports, a hallucinated citation, content outside the course's frame. If the honest answer is "only the score," it is worth knowing that a verified list of universities reviewed exactly that evidence basis and turned the tool off. A score alone can justify a conversation; it cannot justify an accusation.

Check your process. Know your institution's integrity procedure before the meeting: whether an informal conversation triggers reporting obligations, who carries the burden of proof (the institution, in most formal processes — the UK ombuds says this explicitly), and what the student's rights are. You cannot run a fair conversation if neither of you knows the rules.

Decide what outcome would satisfy you. If no account of the writing process could change your mind, skip the conversation and file the formal report — an interview whose outcome is predetermined is not due process, it is theater.

Opening: three sentences that set the frame

"I want to talk about your essay. Some things about it raised questions for me, and I'd rather ask you directly than guess. Walk me through how you wrote it."

What this does: it is honest (questions exist), it is specific (this essay), and it hands the floor to the student in the one arena where an honest writer has an overwhelming advantage — describing their own process. What it avoids: the word "cheating," the detector score as an opener, and any statement of conclusion.

Do not open with "Turnitin says this is 98% AI." Leading with the score does two destructive things at once: it tells the student the verdict precedes the hearing, and it converts the conversation into a referendum on a number neither of you can interrogate.

The middle: ask about process, not vocabulary

Questions an author can answer and a non-author cannot:

• "What did you start with — an outline, a draft, notes?"
• "Why this order of argument? What came first in earlier versions?"
• "What did you cut?"
• "This source in paragraph four — how did you find it, and what does it say beyond the sentence you cited?"
• "Which part was hardest to write?"
• "If you had another week, what would you change?"

Note the shape: these are questions about decisions, and writing is thousands of decisions. A student who made them can narrate them — haltingly, imperfectly, but with the texture of real memory. A student who didn't will produce smooth generalities that answer a different, easier question.

Two kinds of questions to avoid:

• Vocabulary quizzes ("define 'hegemonic' as you used it on page 2"). These punish thesaurus use and second-language writers, not misconduct, and a prepared cheater can survive them anyway.
• Gotcha re-performance ("rewrite paragraph three right now while I watch"). Timed performance under accusation measures anxiety. If a supervised writing sample is genuinely needed, offer it as an option with notice — many honest students will volunteer it, and the offer itself is informative.

Invite evidence, explicitly: "If you have drafts, notes, or version history, I'd genuinely like to see them." Students rarely know their process trail counts as evidence; telling them converts the meeting from interrogation to joint review. (What each kind of evidence is worth — and where version history falls short — is its own topic.)

Reading the result honestly

If the account is fluent and the evidence coheres: say so, in the meeting. "Thank you — that answers my questions" costs nothing and repairs most of the harm of having asked. If the concern arose from a detector score, consider telling the student what flagged it and why you no longer rely on it; students talk, and the ones who hear that the process was fair stop writing in fear of their own fluency.

If the account is vague but the student is anxious: distinguish can't-recall from never-knew. Offer a second, lower-pressure path: "Send me your drafts or version history this week." Anxiety under authority is not evidence of guilt — the students most harmed by false flags (second-language writers, anxious students, students on visas) are precisely the ones who interview worst.

If the account collapses: don't extract a confession in the room, and don't negotiate a private penalty. Both create exactly the procedural defects that ombuds and courts have faulted institutions for. Close neutrally — "I'm still concerned, and I'm going to refer this through the integrity process so you get a proper chance to respond" — and let the process with evidence standards and appeal rights do its job.

What never to say

• "Just admit it and we'll keep it informal." A pressured admission can follow a student onto their transcript, and an innocent, exhausted student will sometimes take the deal. If the case is real, it survives due process.
• "The software is never wrong." It is, by its maker's own published account.
• "This doesn't sound like you" as the whole case. Improvement is the goal of the course; "too good" is not an allegation.
• Anything you would not want read aloud at an appeal hearing — because if the accusation is wrong, that is where it will be read.

Fixing it upstream

Every one of these conversations is cheaper to prevent than to run. The structural fixes — assignments that generate their own process evidence, baselines, and consent-based process records — are covered in alternatives to AI detectors. The shortest version: tell students in the syllabus to keep drafts and version history, and give the ones who want it a way to carry verifiable proof of their own writing process — so that when a question does arise, the conversation starts with evidence on the table instead of a score and a stare.