Most academic integrity disasters do not start with a student trying to cheat. They start with a student who was panicking at midnight, used a tool, and then had no idea how to say so.
The conversation that follows is where things go wrong. An accusation puts a student on the defensive, and from that point forward you are not discussing writing anymore. You are conducting an investigation.
Here is a different approach.
Open with the work, not the suspicion. "Walk me through how you approached this paragraph" is a question about process. "This looks like AI wrote it" is a verdict.
The first question invites an explanation. The second invites a defense. Students who did nothing wrong will freeze on the second, and students who did something wrong will simply deny it.
Before you mention a detection score, ask the student to describe their process. Where did the idea come from? What did the first draft look like? Which part was hardest?
You will learn two things. First, whether their account matches the text. Second, and more useful, whether they understand the argument they submitted. A student who cannot summarize their own conclusion has a problem regardless of who typed the words.
Only after that conversation should a detection result enter the room, and it should enter as one data point rather than a finding of guilt.
These lines are deliberately plain, and they work because none of them are threats.
That last question is the important one. It moves the conversation from judgment to learning, which is where it belongs.
Some cases are genuinely ambiguous. A student who writes in a formal register, works in a second language, or writes short answers can score higher than they deserve. If a student's account of their process is coherent, their other work is consistent, and the only evidence is a score, you do not have a case.
What you have is a reason to change how you assign writing, not a reason to sanction anyone.
That might mean requiring in-class writing, collecting drafts, or asking students to annotate their sources. All of those make authorship visible without turning a score into a verdict.
The single most effective thing an educator can do is say the rule out loud in week one, in plain language, and write it down.
Students follow rules that are clear. They guess at rules that are vague, and guessing produces exactly the midnight decision that starts this whole cycle.
A student who has talked with you about their writing once is dramatically easier to talk with a second time. Short process conversations, a comment on a draft, a question after class: these are cheap, and they are the reason an ambiguous score can be resolved in a five-minute conversation instead of an honor council hearing.
Detection tells you where to look. The conversation is what actually resolves it.