Academic Integrity Hearing AI: What Panels Assess

An academic integrity hearing AI case turns on evidence, not on a detector score. Panels decide whether the institution has shown misconduct under its own policy. Your job is to account for how the work was made, bring the trail that proves it, and ask the panel to apply its normal standard of proof.

What is an academic integrity hearing actually deciding?

Most panels are not asked “did a detector flag this?” They are asked whether the evidence meets the institution’s standard for misconduct. That usually means three things: the act (did you use a prohibited tool, or submit work that was not yours?), intent or knowledge (did you know, or should you reasonably have known, that it was not allowed?) and the evidence (can the institution demonstrate both, at whatever threshold its policy sets?).

That last point is where AI cases get shaky. A detector score is a probability estimate produced by a model comparing your text with patterns it associates with generated writing. It is not a record of your keyboard. It cannot show who typed a sentence. When a panel treats a score as the finding rather than as one piece of evidence, it has quietly skipped the part where the institution proves its case.

This is why you should read the policy that applies to you, not a general one. Rules differ on whether AI assistance is banned outright, permitted with disclosure, or permitted for some tasks and not others. If the policy allows editing tools, grammar checkers or translation aids, say so, with the clause in front of you. Our guide to what your university actually allows walks through how to read those clauses before a hearing.

What to bring: your writing process as evidence

The strongest thing you can put in front of a panel is not an argument. It is a trail.

1. Drafts, in order. Cloud documents keep version history. Export it, or screenshot the version list with timestamps. An essay that grew over days looks different from one pasted in at 2 a.m.
2. Notes and research. Reading notes, highlighted PDFs, library loans, bookmarks, interview recordings, data files. Anything that shows the thinking before the sentences.
3. Outlines and messy intermediate files. Half-finished paragraphs are evidence of construction. Nobody generates a draft and then generates the scaffolding.
4. Correspondence. Emails to a supervisor, messages to a study group, feedback from a writing centre. These date your work and show the direction of travel.
5. Your own earlier work. Previous essays in the same style help show that the voice in question is yours.

If you used an AI tool inside the rules, for brainstorming, for checking grammar, for translating a source you then read, document that too. Concealing a permitted tool looks worse than declaring it. What sinks people is a gap in the record, not the tool itself.

If you want to see which specific sentences carry the signal in your own draft before anyone else scores it, run the text through the sentence-level AI detector. It highlights each line by AI probability, so you can point at the actual passages under discussion instead of arguing about a whole-document number.

Why does a detector flag writing you did yourself?

Because detectors measure pattern, not authorship. They look for signals like low variation in sentence length, predictable word choices, generic transitions and evenly distributed information. Human academic writing sometimes looks like that too, especially when it has been through a style guide, a translation, or several rounds of tidying.

We ran our own detector over 46 hand-written DeAIze guides. The mean AI score was 27%, the median 25%, and the range ran from 0 to 55%. Seventeen of the 46 were flagged at the threshold we used. Every one of those documents was written by a person. That is the honest picture of a detector in the wild: useful, noisy, and not a verdict.

There is a second, stranger finding from the same run. In the human-written sample, common AI tells appeared about 3 times per 100 words. In 12 unedited model outputs generated from fixed prompts, the same tells appeared about 1 time per 100 words. The signals people rely on by eye are not reliably present in machine text and are often present in human text, because humans write in habits. If a panel member says the prose “sounds like AI”, that is an impression, not a measurement, and it deserves the same scrutiny as any other impression.

How do you present your process at the panel?

Keep it chronological and concrete. “Here is how this essay was made” is easier to follow than “I would never do that.”

• Start with the assignment and the brief.
• Describe your research: what you read, where you found it, what you rejected.
• Walk through the drafting: when you started, how many drafts, what changed.
• Point to the artefacts as you go. Hand them over or have them on screen.
• Explain any tools you used, by name, and what you used them for.
• Stop. Let them ask.

Answer the question you were asked. If you do not know, say you do not know and offer to find out. Panels are usually trying to establish a sequence of events; a witness who is consistent and specific is more persuasive than one who is fluent and vague.

If English is not your first language, say so early and plainly. Non-native writers are disproportionately flagged, not because they cheat but because their prose often uses the same safe, regular constructions detectors associate with generated text. There is a whole guide on why AI detectors flag non-native English writers and it is worth reading before you speak, because it gives you language for something you may have felt but not been able to name.

What not to say

The fastest way to lose a hearing is to concede a finding you disagree with because you want it over.

Do not say: “I must have used it without realising.” If you did not use a prohibited tool, that sentence is a false admission, and it will be recorded as one. Panels act on admissions. You cannot walk it back later.

Do not say: “The detector must be wrong, so I have nothing to explain.” Detectors are often wrong. But refusing to engage with the evidence makes you look evasive even when you are right.

Do not say: “Everyone uses ChatGPT.” It is not a defence, and it invites the panel to treat your case as a policy question rather than a factual one.

Do not say: “I can’t remember.” For everything. Memory is allowed to be imperfect, but if your account of writing a 3,000-word essay contains no remembered detail at all, the panel will notice.

Do not volunteer theories about who reported you. It is irrelevant, it sounds retaliatory, and it shifts the conversation away from the evidence you came to present.

What you can say, and should, is simple: “I wrote this. Here is the record of how. I am asking the panel to apply its normal standard of evidence to the detector output, and I would like to know what that standard is.”

How do you ask for a fair standard of evidence?

Ask in writing, before the hearing if the process allows it. Three questions do most of the work:

1. What standard of proof applies? Is it the balance of probabilities, or something higher? Is that written in the policy?
2. Who bears the burden of proof? In most institutional processes the institution must establish misconduct, not the student disprove it. Get that confirmed.
3. What evidence will be relied on, and can I see it? You are usually entitled to the detector report, the flagged passages, the version the marker assessed, and any correspondence about the case.

Then ask the operational question: what would the panel accept as evidence that I wrote this? If the answer is “nothing”, you have learned something important about the process, and you can say so calmly, on the record. If the answer is a list, you now know exactly what to bring.

A detector score is one input. If the panel’s case rests on it alone, that is worth stating without heat: the tool reports a probability, not a fact about authorship. Our own numbers show why. A mean of 27% across 46 documents known to be human-written is not a system you would convict on.

How do different kinds of evidence compare?

Not all evidence carries the same weight. Here is how the usual items stack up in an AI misconduct case.

Evidence	What it shows	Weight in a hearing
Detector score (whole document)	A probability estimate about patterns in the text	Weak on its own; useful as context
Sentence-level detector output	Which specific lines carry the signal	Moderate; lets you explain individual passages
Document version history	How the text changed over time	Strong; hard to fabricate at scale
Research notes and sources	That the thinking predates the sentences	Strong when dated
Correspondence with supervisors	The direction of travel and timeline	Strong; independent of you
Your earlier essays	That the voice is consistent with your work	Moderate; supports the pattern argument
Verbal account alone	Your recollection of the process	Weak without artefacts to anchor it

Notice what sits at the top. The evidence that persuades is the evidence you generated while working, not the evidence you assemble the night before. If you have nothing in the top rows, say so honestly and lean on what you do have.

Before the hearing: a short checklist

• Read the policy that covers your assessment. Note the exact clause on AI tools.
• Request the documents: detector report, flagged sections, the submitted file, marker comments.
• Export your version history and save it somewhere you control.
• Assemble the trail: notes, sources, outlines, correspondence.
• Contact your student union, adviser or representative. Many institutions provide one, and having a second person in the room changes the tone.
• Write a one-page timeline of how the work was made. Take it in with you.
• Practise the account out loud once, so the first time you tell it is not under pressure.
• Ask about recording or minutes, and whether you may bring a supporter.

One more thing worth doing: score your own draft before the hearing, so you are not surprised by a number in the room. A sentence-level tool is more useful here than a whole-document score, because it shows which lines are driving the result. If a specific paragraph is flagged and you can explain why it reads that way, a quoted definition, a translated passage, a formulaic methods section, you have an answer ready instead of a shrug. Our guide on what to prepare when a detector flags your thesis sets out the same routine at longer length.

What the panel is really looking for

Strip away the procedure and panels are usually trying to answer one question: can this student account for this work? Not perfectly, not with forensic precision, but plausibly and consistently, with something to show for it.

That is a standard you can meet whether or not you ever touched an AI tool. It is also a standard worth holding the institution to. If the case against you is a percentage on a report, say so, politely and repeatedly, and ask what else there is. Sometimes the answer will be more evidence. Sometimes it will be nothing, and the panel will know it.

If you did use AI outside the rules, the honest route is to say what you used and why, and to accept the outcome the policy provides. Hearings are not a game to be won; they are a process for establishing what happened. The students who come out of them worst are rarely the ones who made a mistake. They are the ones who could not describe their own work.

Ready to see which sentences read machine-made?

Before any panel sees a number, you should. DeAIze’s AI detector scores the whole document and highlights each sentence by AI probability, so you can see exactly which lines carry the signal and prepare an explanation for them rather than a general denial. It won’t tell you what a panel will decide, and it isn’t a verdict. It is a map of the argument you are about to have.

If a flagged passage is genuinely your own writing that simply reads flat, the draft editor rewrites only the flagged sentences, checks that the meaning survived, and re-scores the result. You can try it on 200 words free at sign-up, and see pricing if you need more. Either way, bring the version history. The tool explains the number; the trail explains the work.

What if the panel’s case rests on a rewriter’s output?

Some cases arrive with a second report attached: a marker ran your draft through a rewriting tool, or you did, and now there are two numbers in the room. Treat them as two separate instruments, not as corroboration. A rewriter’s output is a suggestion about wording. It is not a record of authorship, and it does not confirm a detector’s reading just because the two appear on the same page.

Ask three things about any rewriting report:

1. Which text was scored? Your submitted file, or a version a tool produced from it? Those are different documents.
2. What does the number mean? Some tools report a similarity percentage, some a model score, some a share of qualifying prose. They are not interchangeable, and none proves who wrote the sentence.
3. Was the meaning preserved? A rewrite can quietly drop a negation, a citation or a qualifier. If a tool altered your claims, that is an editing problem you should be able to describe, not a confession.

If you are weighing up tools before a hearing, read how to judge a rewriting tool and what each tool returns so you can tell a score definition from a verdict. Bring the original file either way. The original is the evidence; the rewrite is commentary on it.

For context on how noisy our own instrument is: on 2026-10-13 we scored 1,549 passages from DeAIze guides not marked generated (metadata selection only, human authorship not independently verified) and the AI score mean was 28%, median 20%. That is our own measurement, not a third-party detector result, and it is a passage count, not a verified false-positive rate. It is enough to justify asking what a number is measuring before anyone treats it as a finding.