Building an Academic Integrity Hearing Around AI: What Counts as Due Process

A student walks into your office with a printout in hand: a 96% “AI-generated” score on their final essay. They’re upset, you’re uneasy, and the syllabus says academic misconduct can mean a failing grade. Before anyone signs anything, there’s a question worth slowing down for: is what happens next a fair process, or just a verdict with extra steps?

This guide is about the process, not the verdict. If you teach, chair a committee, or write the policy everyone else follows, here’s how to build an academic integrity hearing for AI allegations that holds up, treats students fairly, and still lets you address real cheating when it happens.

Key takeaways

• Due process rests on four pillars: clear notice, a genuine chance to respond, an impartial decision-maker, and a documented, appealable decision.
• An AI detector score is a probability, not a finding of fact. It can open an inquiry; it shouldn’t close one.
• The burden of proof belongs to the institution. “Prove you didn’t use AI” is the wrong question.
• Process evidence (drafts, version history, an oral walkthrough) usually beats any single number.
• Decide your standard of proof, evidence rules, and appeal path *before* a case lands on your desk.

What “due process” actually means in an academic setting

Due process in school discipline isn’t a courtroom, and it doesn’t have to be. Courts have long held that what’s owed depends on what’s at stake. A two-point deduction needs almost nothing. Expulsion needs a lot. An AI-cheating allegation that could fail a student for the term sits high enough on that scale that you owe them a real procedure.

Strip away the legal vocabulary and you’re left with a few plain commitments:

• Notice. The student learns the specific allegation, the assignment involved, and the evidence behind it — not just “we think you cheated.”
• Opportunity to be heard. They get to tell their side before a decision, not after.
• A neutral decision-maker. The person who reports the suspicion shouldn’t also be the one who rules on it.
• A reasoned, written outcome. The decision says what was found, why, and how to appeal.

None of that requires lawyers or a hearing room. It requires that you don’t decide the answer before the student has spoken.

Why AI allegations strain the usual process

Plagiarism used to come with a smoking gun: a source document the copied text matched. You could lay them side by side. AI allegations rarely give you that. The “evidence” is often a probability score from a tool, and that’s a different kind of claim.

Here’s the part that trips people up. A detector score is a probability, not proof. Detectors work by measuring statistical features of text — how predictable each word is (perplexity) and how much sentence-to-sentence variation there is (burstiness) — and comparing them to patterns common in machine-generated writing. Predictable, evenly-paced prose scores “AI-like.” But plenty of human writing is predictable and even-paced: a careful student, a non-native speaker, someone trained to write in a formal register. The tool can’t tell you *who* wrote the text or *how*. It tells you the text resembles a pattern.

It’s worth remembering that even the company that built ChatGPT couldn’t make this work reliably. OpenAI retired its own AI Text Classifier in July 2023, saying it was no longer available because of its low rate of accuracy. If the model’s maker can’t detect its output dependably, a third-party score deserves the same humility.

Where the score belongs in your process

The fix isn’t to ban detectors. It’s to put them in the right spot. A detector score is a trigger, not a verdict. It can justify a closer look; it can’t carry a sanction by itself.

A workable order of operations looks like this:

1. Flag. A score, an instructor’s instinct, or a tip raises a question. Fine — that’s allowed.
2. Inquire. You gather corroborating evidence and talk to the student before charging anything.
3. Charge (or don’t). If there’s a real case, you state the allegation in writing with notice.
4. Hear. The student responds to an impartial decision-maker.
5. Decide and document. You apply your stated standard of proof and write down the reasoning.
6. Appeal. The student can challenge the outcome through a defined path.

Skipping straight from step one to step five — score to sanction — is exactly the move that gets reversed on appeal and erodes trust in the whole system.

The burden of proof is yours

This is the single most common place AI cases go wrong, so it’s worth stating bluntly. The institution carries the burden of proving misconduct. The student does not have to prove innocence.

Asking a student to “prove they didn’t use AI” sounds reasonable until you try to do it. How would *you* prove a negative about your own writing? The burden sits with the accuser because, in practice, it’s the only side that can be held to it fairly.

Most academic policies use the preponderance of the evidence standard — more likely than not, often described as just past 50/50. That’s a lower bar than a criminal trial, but it still demands actual evidence pointing toward a violation. Suspicion plus a high score does not get you there on its own. Whatever standard you pick, write it down and apply it the same way every time.

A mini-scenario: how a fair inquiry plays out

Here’s a hypothetical to make the steps concrete. Maya, a sophomore, turns in a literature essay. Turnitin’s AI indicator reads 88%. Her instructor, Dr. Okafor, is concerned but doesn’t email a charge. He does this instead:

• He pulls the assignment’s version history from the school’s Google Workspace. The doc shows 40 minutes of typing across three sessions, with edits, deletions, and a paragraph she clearly rewrote twice.
• He asks Maya to walk him through her argument in office hours. She explains why she chose her central example and what she cut. She knows the paper cold.
• He notes Maya is a multilingual writer whose prose tends to be clean and formal — the kind of low-perplexity style detectors over-flag.

Put together, the file now says: a high score, but a documented writing process, fluent author knowledge, and a known reason the detector might be wrong. Dr. Okafor closes the inquiry. No charge, no hearing, and — just as important — a written note explaining why, in case anyone asks later.

Now flip it. If Maya couldn’t discuss her own paper, the doc was pasted in one block at 2 a.m. with zero revision, and she’d recycled phrasing across three classes, *that’s* a case. Notice the difference isn’t the score. It’s the corroboration.

Evidence that actually helps

When you build a case file, weight the kinds of evidence that a person — not just an algorithm — can evaluate:

• Process artifacts. Version history, draft files, browser or LMS timestamps, research notes. These are hard to fake and easy to explain.
• The student’s account. An oral walkthrough of the work reveals authorship faster than almost anything. Grading drafts and process, not just the final product, builds this in from the start.
• Pattern across submissions. A sudden, unexplained shift in a student’s voice is more telling than one number.
• Context. Is the student a non-native English writer? Research matters here. A 2023 Stanford study with James Zou as senior author found multiple detectors flagged non-native English writing far more often than native writing, because simpler vocabulary reads as “AI-like.” That doesn’t decide the case — it just lowers how much weight the score deserves.

Different institutions handle this with different tools and budgets, so it’s worth knowing what the different plans cover before you standardize on anything. The point isn’t the tool; it’s calibrating how much skepticism a score deserves before you sit on a panel.

Build the policy before the case

The worst time to figure out your procedure is mid-allegation, with a stressed student in the room. Get these on paper ahead of time:

• Your standard of proof, stated plainly.
• A rule that no detector score alone supports a sanction.
• A defined separation of roles — reporter, investigator, decision-maker.
• A written notice template and a fixed response window.
• A real appeal path with a different reviewer.

For more on the surrounding craft — talking to students, designing assignments that don’t invite shortcuts, reading scores correctly — browse our other guides for educators.

Frequently asked questions

Can an AI detector score alone justify a sanction?

No, and most policies that survive scrutiny say so outright. A detector returns a probability from text patterns, not evidence that a particular student used a particular tool. Use the score to start an inquiry, then build the case from corroborating evidence. If your file is just a percentage, you don’t have enough.

Who has the burden of proof?

The institution. Whoever brings the allegation must meet your stated standard, usually “more likely than not.” Demanding a student prove their innocence flips the burden and is a frequent due-process failure. Students may offer evidence; they’re never required to.

What standard of proof should we use?

Most academic settings use preponderance of the evidence — more likely than not. It’s lower than “beyond a reasonable doubt” but still requires real evidence, not a hunch. Write it into policy and apply it identically every time.

Does a student have a right to see the evidence?

Yes. Notice of the specific allegation and access to the underlying material is core to due process. Tell them which tool flagged the work, what it returned, and what else you’re relying on. Nobody can answer a case they can’t see.

What if the student is a non-native English speaker?

Treat it as a real factor. Peer-reviewed research has shown detectors over-flag non-native writing because predictable, formal prose reads as machine-like. That doesn’t prove innocence or guilt — it means the score is less reliable for that writer, so weigh process evidence and a direct conversation more heavily.

The bottom line

A fair AI hearing isn’t about being soft on cheating. It’s about being right more often, and being seen to be fair when you are. Give students notice, let them respond, keep your decision-maker neutral, and never let a single probability stand in for a finding of fact.

Want to understand the scores your students are seeing before you weigh them? Try PaperBleach and check a draft for yourself — knowing how detection behaves is the first step to judging it fairly.