How to Build an Academic Integrity Case That Doesn’t Rely Solely on a Detector

A defensible academic integrity case treats a detector score as a reason to look closer — never as the finding itself — and builds instead on process evidence: drafts, version history, and a conversation with the student about their own work.

That single sentence is most of what ai misconduct investigation best practices come down to, but the details matter, because the details are where investigations go wrong. A percentage on a screen feels like proof. It isn’t, and every serious authority in this space — including the detector vendors themselves — says so. Here’s how to build a case that survives scrutiny, protects innocent students, and still holds actual misconduct accountable.

Why the score can’t carry the case

Start with what a detector actually outputs: a probability that text resembles typical machine writing. Resemblance is not authorship. Careful, formal human prose can score high; lightly edited AI prose can score low. The error rates are real and asymmetrically distributed — the 2023 Stanford study by Liang and colleagues found detectors flag non-native English writers at dramatically higher rates. Turnitin’s own guidance tells instructors its score “should not be used as the sole basis” for an accusation, and Vanderbilt University found the reliability concerns serious enough to switch the feature off entirely in 2023.

We’ve written a full breakdown of whether detection scores hold up as evidence; the short version is that a score is a screening signal with a known error rate, and screening signals are where investigations begin, not end.

AI misconduct investigation best practices, step by step

1. Treat the flag as a hypothesis

When a detector flags a paper — or your own reading raises an eyebrow — write down what specifically seems off. “Score of 91%” is one data point. “The essay’s voice differs sharply from this student’s in-class writing, cites two sources that don’t exist, and discusses a topic we never covered” is a pattern. Hypotheses get tested; verdicts get defended. Start with the former and you’ll behave better throughout.

2. Gather process evidence before the conversation

The strongest evidence in these cases is almost never the finished text. It’s the trail the text left behind:

• Version history. Google Docs and Word Online record the document growing. A paper written over six sessions across four days looks nothing like a paper pasted in as three large blocks in nine minutes.
• Drafts and outlines. Prior submissions, planning documents, annotated bibliographies.
• Fabricated citations. Invented sources are checkable, objective, and don’t depend on any classifier. If the paper cites a journal article that doesn’t exist, you have a real finding no one can dispute.
• Internal consistency. Does the essay reference class discussions, personal experience, or feedback from earlier assignments? Generic text tends to float free of the course it was written for.

Note what’s missing from this list: a second detector score. Running the text through five tools and averaging them adds no independent information — they’re all measuring the same statistical surface, and their disagreement (which is common) only demonstrates the point.

3. Have the conversation as an inquiry, not an ambush

Invite the student to talk about the work, and mean it. Ask how they approached the assignment, what they struggled with, what they cut and why, what they’d change with another week. A student who wrote the paper answers these questions with the specific, slightly messy texture of real memory. A student who didn’t tends toward abstraction, or reproduces the essay’s talking points without being able to go beneath them.

Two cautions here. First, interview nerves are real, and a shaky answer from an anxious eighteen-year-old is weak evidence on its own — look for the pattern across the whole conversation, not one stumble. Second, never open by announcing the detector score. It converts the meeting into an accusation, and if the flag is false, you’ve told an honest student their integrity is already in doubt. The damage from that conversation doesn’t fully undo.

4. Document contemporaneously

Every case that ends in a sanction may end up in front of a committee, and committees read files. Record what triggered the concern, what tool (if any) produced a score and what its documented false positive rate is, what process evidence you collected, and what was said in the conversation — written during or immediately after, not reconstructed a month later. If your process was fair, the file proves it. If you can’t write the reasoning down in a way that would persuade a neutral reader, that itself is information.

5. Match the conclusion to the evidence

Three honest outcomes exist. Supported: multiple independent lines — process evidence, fabricated sources, the conversation — point the same direction; proceed through your institution’s formal process, where the student gets the due process an academic integrity hearing about AI is supposed to provide. Unsupported: the evidence contradicts the flag; close it, tell the student clearly they’re in the clear, and consider what triggered the false alarm. Ambiguous: the evidence doesn’t resolve — and this must be closable as “no finding.” An integrity process that can’t say “not proven” isn’t a process; it’s a conviction machine with paperwork.

Designing the next assignment so this gets easier

The best investigators quietly become assessment designers, because the cheapest evidence is the kind the assignment generates automatically. Scaffolded milestones (proposal, annotated bibliography, draft, final) create a paper trail as a side effect. Requiring submissions through a platform with version history does the same. Short in-class writing gives you an authentic baseline of each student’s voice. An oral component — even five minutes — makes the “can you discuss your own argument” test routine rather than accusatory.

None of this is about surveillance. It’s about making the honest path legible, so that when a question does arise, the answer is usually already sitting in the record.

The standard that protects everyone

The uncomfortable truth about detector-led cases is that they invert the burden of proof: a number appears, and the student is asked to prove a negative. Best practice puts the burden back where it belongs — on the evidence — which protects innocent students from statistical bad luck and, just as importantly, makes the sanctions that *are* imposed stick, because they rest on facts a committee can verify rather than a probability it has to take on faith.

If you’re an educator trying to calibrate your own intuition, it’s worth spending ten minutes to run a passage through a detector yourself — your own writing, a student-style sample, an AI draft — and watch how the numbers move. Understanding what the tool sees, and what it can’t, is the first of the best practices. There are more guides on detection and policy in our library, and our pricing page shows exactly what we offer educators.

Frequently Asked Questions

Can a detector score alone prove a student used AI?

No. Detectors are probabilistic classifiers: they measure how statistically similar a text is to typical machine writing, and they produce false positives at documented rates — more often on non-native English writing, short passages, and heavily edited text. Turnitin’s own guidance says its AI score should not be the sole basis for an accusation. A score can justify looking closer; it cannot carry a case by itself.

What evidence is stronger than a detector score?

Process evidence. Version history from Google Docs or Word showing the document grow over hours, early drafts and outlines, research notes, browser-based citation trails, and the student’s ability to discuss their own argument in detail. A five-minute conversation about why they made a specific choice in paragraph three usually reveals more than any percentage.

How should an investigation conversation with a student start?

As an inquiry, not an accusation. Open with the work, not the suspicion: ask the student to walk through how they built the piece, what they found hardest, what they cut. Students who wrote their work answer these easily, often eagerly. Presenting a detector score as a verdict in the first sentence poisons the conversation and, if the flag is false, does real damage to a student who did nothing wrong.

What should be documented during an AI misconduct investigation?

Everything, contemporaneously: the initial concern and what triggered it, which tool produced any score and its documented error rate, copies of all drafts and version histories collected, notes from the student conversation taken during or immediately after, and the reasoning connecting evidence to conclusion. If the case later reaches a hearing or an appeal, the file — not memory — is what gets examined.

What if the evidence is genuinely ambiguous?

Then the case fails, and that has to be acceptable. Academic integrity processes borrow their legitimacy from fairness, and fairness means some guilty parties go unsanctioned because the evidence wasn’t there. The alternative — sanctioning on ambiguity — guarantees punishing innocent students at scale. Redesigning future assessments to generate process evidence is the productive response to an unprovable case.