The Teacher’s Dilemma: Reporting a Suspected AI Submission Responsibly

When you suspect an AI-written submission, the responsible sequence is: verify quietly first, talk with the student second, and file a report only when independent evidence converges — because both failure modes, ignoring real misconduct and accusing an innocent student, do lasting harm.

Any honest guide to how teachers should handle suspected ai use has to start by admitting the dilemma is real. Report too readily and you may put an innocent student — statistically, often a non-native English speaker — through a process that marks them regardless of the outcome. Look away and you quietly devalue the work of every student who wrote their own essay at 1 a.m. There’s no path with zero risk. But there is a sequence that keeps you fair, defensible, and sane.

First, sit with what a suspicion actually is

A suspicion is a hypothesis. It might come from a detector score, a sudden jump in polish, an essay that reads like nobody in particular. Each of these is worth noticing; none is a finding.

Detector scores deserve special caution because they arrive dressed as findings. They’re probabilities from a classifier with a documented false positive rate, and the errors don’t fall evenly — the 2023 Stanford study by Liang and colleagues showed detectors flag non-native English writers far more often. Turnitin’s own guidance says the score shouldn’t be the sole basis for action, and Vanderbilt found the problems serious enough to disable the feature. We’ve written at length about the false accusation problem — students wrongly flagged for AI; the pattern in those stories is nearly always the same: a score treated as a verdict, and a process that started with an accusation instead of a question.

The “reads like AI” intuition needs the same humility. It’s real pattern recognition — you’ve read hundreds of essays — but it’s also exactly the intuition that misfires on formal, careful, slightly stiff human writing, which describes a lot of earnest students and most second-language writers.

So the first responsible act is nothing visible at all.

Step one: verify quietly

Before the student knows anything is on your mind, check what can be checked without them.

Version history. If the submission came through Google Docs or a platform that records revisions, look at how the document grew. Hours of accretion across days is strong evidence of authentic work. One paste, nine minutes, done — that’s worth a conversation.

Citations. Open every source. Fabricated references — real-sounding journals, plausible authors, nonexistent articles — are the closest thing to objective evidence in these cases, and they don’t depend on any classifier’s opinion.

The student’s baseline. Compare against in-class writing, earlier assignments, discussion posts. Voice drift across a semester is normal; a wholesale personality transplant between the midterm and the final essay is data.

Your own policy. Reread what your syllabus actually says about AI. If it bans generation but permits editing help, or says nothing at all, be honest about whether the conduct you suspect would even violate it. An ambiguous policy is your problem to fix going forward, not the student’s to be punished under.

If verification dissolves the suspicion — history shows organic drafting, sources are real, the voice matches — stop here. You never accused anyone, which is precisely the point of this order of operations.

Step two: the conversation

If the concern survives verification, meet with the student — and structure the meeting as an inquiry about the work, not a confrontation about the crime.

Ask process questions with genuine curiosity: How did you get started on this? What was the hardest part? What did you cut? Which source did you find most useful, and why? You already know what authentic answers sound like — specific, textured, a little rambling, anchored in the actual experience of struggling with the assignment. A student who can go beneath their own essay’s surface, defend a paragraph’s placement, or explain a choice you’d have made differently is giving you real evidence of authorship. A student who can only restate the essay’s points at the same altitude the essay states them is giving you evidence, too.

Two rules keep this fair. Don’t open with the detector score or an accusation — an inquiry can escalate later if warranted, but an accusation can never de-escalate back into an inquiry; the relationship damage is immediate and mostly permanent. And weigh nerves honestly: an anxious student stumbling on one question is noise, a pattern across the whole conversation is signal. If you teach younger students, there are extra considerations when the students are minors — parental involvement, developmental fairness, and records that follow a child — that make this care matter even more.

Sometimes this conversation ends the matter in the best way: a student explains their process convincingly, or admits to more AI help than allowed and you handle it as the teaching moment it is, with a redo and a clearer understanding. Not every resolved case needs to become a filed case.

Step three: report, close, or redesign

After verification and conversation, you’re in one of three places.

Evidence converges — file. Fabricated sources plus implausible version history plus a student who can’t engage with their own argument: file the report, and do it without guilt. At this point, *not* reporting is unfair to every student who did the work. Bring the file — what triggered the concern, what you checked, what was said — so the committee sees your reasoning rather than reconstructing it. Note what “converges” means: multiple independent lines. A high score plus a hunch is one line and an echo.

Evidence contradicts — close it, fully. Tell the student clearly they’re in the clear if they knew of the concern. A cloud of unresolved suspicion is a sanction you imposed without a hearing.

Evidence is ambiguous — close it, and redesign. This is the hardest one to accept and the most common. Filing on ambiguity just forwards your uncertainty to a committee with less information than you have. Instead, let the case change your assessments: scaffolded drafts, version-tracked submissions, brief oral defenses, in-class writing baselines. These make the next case easy — or, better, make it never arise.

One more calibration worth doing before any of this: spend ten minutes and test a detector against your own writing — your lecture notes, an old paper, an AI sample. Seeing where the tool is sharp and where it’s confidently wrong changes how you weigh it, and that judgment is exactly what your students are relying on you to have. There are more resources for educators in our library, and our pricing covers classroom use.

The dilemma doesn’t fully resolve — that’s what makes it a dilemma. But a teacher who verifies before suspecting aloud, asks before accusing, and files only on convergent evidence has done right by both the honest students and the flagged one. That’s the whole job.

Frequently Asked Questions

Should I report a student the moment a detector flags their paper?

No. A detector flag is a screening signal with a known false positive rate, not a finding — and the vendors themselves say a score shouldn’t be the sole basis for action. Before any report, gather what only you can gather: version history, comparison with the student’s other writing, and a conversation about the work. Reports filed on a score alone regularly collapse at the hearing stage, after the damage to the student is already done.

What should I check before talking to the student?

Four things: the document’s version history (a paper that grew over multiple sessions tells a different story than one pasted in whole), the citations (invented sources are objective, checkable evidence), the student’s writing baseline from in-class or earlier work, and your course’s actual written AI policy — because if the policy is silent or ambiguous, that’s a policy problem, not a misconduct case.

How do I raise the concern without accusing an innocent student?

Frame the meeting around the work, not the suspicion. Ask the student to walk you through their process — how they started, what they struggled with, what they’d revise. Students who did the work answer with specific, textured detail. Don’t open with the detector score or the word “cheating”; an inquiry can become an accusation later if warranted, but an accusation can never become an inquiry again.

What if I’m still unsure after the conversation?

Unresolved ambiguity is an answer: it means you don’t have a case, and filing anyway outsources the ambiguity to a committee with less information than you have. Close it, keep brief notes, and adjust future assessments to generate better process evidence — drafts, milestones, in-class writing. One unprovable instance costs less than one wrong accusation.

When is filing a formal report clearly the right call?

When multiple independent lines of evidence converge: fabricated citations, version history inconsistent with authentic drafting, a conversation where the student cannot engage with their own submission’s content, or an admission. At that point filing is not just defensible but required by fairness to the students who did the work themselves — and the file you bring should let the committee see exactly what you saw.