How to Appeal an AI Detection Accusation: The Procedure, Step by Step

Glossary

What we actually measured

There is a lot of advice online about being accused of AI use. There is almost nothing about the machinery you land in afterwards. So we went to the primary sources: 21 published student-conduct and academic-integrity procedures at US institutions, coded for four things — the deadline to file, the grounds an appeal may be based on, whether a support person is allowed, and the standard of proof.

The reason this matters more than another article about detector accuracy: the deadline is the one part of this that is not negotiable and not arguable. Miss it and the substance never gets read.

The filing window is short and remarkably consistent

Nineteen institutions, none longer than ten days, twelve of them at exactly five. Note the split between "business," "calendar," and "working" days — Swarthmore's five business days and West Virginia's five calendar days are not the same week. Read your own letter for which one applies to you, and count from the date on the decision notice, not the day you opened the email.

The grounds are a closed list, and they're mostly not about the facts

Sixteen of the policies enumerate the grounds an appeal may be based on. Here is how often each appears:

That fourth row is the finding that should change how you write your letter. At fewer than half of these institutions — 7 of 16 — can you appeal on the basis that the evidence didn't support the conclusion. The University of Iowa allows an appeal where "the decision was not supported by substantial evidence when viewed as a whole." UNH names "Sufficiency of Evidence — the finding did not meet the required standard of evidence, and the decision rendered is clearly erroneous." Chapman allows a respondent found responsible to argue "that the facts in the case were insufficient to establish that a violation of the Code occurred." Bowling Green allows "error in interpretation of evidence." New Mexico allows a decision "not in accordance with the evidence presented." Northern Kentucky allows "a finding was not supported by evidence." Georgia Southwestern allows a claim "that the finding was inconsistent with the weight of the information."

Everywhere else, arguing the merits again is explicitly out of scope. Swarthmore: "the appeal is not a new review of the underlying matter." UNC Charlotte: "It is not the purpose of the appeal process to provide for a new hearing at a higher administrative level." Pratt Munson: "Appeals are not full rehearings but are instead limited to reviewing the original case based on specific criteria." Three separate policies also state that dissatisfaction with the decision is not itself a ground.

What follows from this is the single most important sentence in this article: the place to win an AI-detection case on the facts is the first meeting, not the appeal. If you are still at the allegation stage — nobody has made a finding yet — the companion piece to this one, what to do when you're falsely accused of using AI, is the right guide, and you should stop reading this and go there.

The first 24 hours after a finding

1. Find the deadline and write it down. It's in the decision letter and in your institution's published procedure. Convert it to a calendar date immediately, including whether it counts business or calendar days.
2. Stop touching the file. Your document's revision history is the strongest evidence you have and every edit you make after the fact degrades it. Do not "clean up" the flagged passage.
3. Ask, in writing, for the record. Request the detector output itself (which tool, which version, what score, which passages), the written policy that authorises its use, and the case file the decision was based on. Do it by email so the request is timestamped. Whether they had to give you the underlying evidence — and whether they did — is often the procedural error your appeal will rest on.
4. Read the procedure, not just the letter. The letter tells you the outcome. The procedure tells you the grounds, the deadline, who decides, and whether you may bring an advisor.
5. Ask whether you may bring a support person. Six of the 21 policies we read state this right on the page a student would land on; the others are silent, which is not the same as "no." Brandeis grants the right "to be accompanied by an Advisor at all phases of a Student Conduct Process"; Chapman permits "one support person at the hearing." Where a policy is silent, ask in writing rather than assuming the answer is no.

Building the evidence dossier

Appeals are won on artefacts, not on assertions. Assemble these before you write a word of the letter.

• Version history export. Google Docs (File → Version history) and Microsoft 365 both keep timestamped revision records. Export or screenshot the full timeline, not a single frame. A real document shows dozens of sessions, false starts, and deletions.
• Dated drafts, outlines, and notes. Including handwritten ones — photograph them and keep the file metadata.
• Research trail. Library database access logs, browser history for the sources you cite, PDFs with your annotations, interlibrary loan records.
• Communications. Emails or messages to classmates, TAs, or the instructor where you discussed the assignment while writing it.
• An offer to demonstrate knowledge. Volunteer to talk through your argument, explain a specific choice on page three, or write a comparable passage under observation. This is the item instructors most often find persuasive and the one almost no advice page mentions.
• The detector's own limits, cited precisely. Turnitin's published guidance states its false-positive rate is measured at document level and is higher on short passages, and that scores should not be the sole basis for a misconduct decision. Our own test of GPTZero on 861 verified human sentences found 13.8% classified as AI or mixed — with the honest caveat that this was a sentence-level test, deliberately the harder case, against a vendor claim of ≤1% at document level. Both numbers are true. Use the caveat; an appeal officer who catches you overstating the research loses trust in everything else in your letter.

For the templates, checklists, and the citation pack in a form you can copy and fill in, we put them all on one page: the AI accusation defense kit. No account, no email address.

Writing the appeal letter

The structure that survives screening:

1. Line one: name the ground. Quote your institution's own wording. "I am appealing on the ground of significant procedural error, as defined in section X of the Student Conduct Procedures." Appeal officers screen for this before reading anything else. UNH's procedure says explicitly that the notice "must contain: the specific grounds for appeal; the specific outcome requested; and the appellant's reasons in support."
2. State the specific outcome you want. Reversal of the finding, a rehearing, or reduction of the sanction. Vagueness reads as a complaint rather than an appeal.
3. Evidence the ground, not the feeling. For procedural error: what the policy required, what happened instead, and how that changed the outcome. For new evidence: what it is, why it wasn't reasonably available earlier, and what it would have changed. Most policies require that third clause and reject appeals that skip it.
4. Attach the dossier as numbered exhibits. Reference them by number in the text.
5. Keep it short and unemotional. Chapman caps appeals at five double-spaced pages. Take that as a general ceiling even where none is stated.

Things to leave out: general argument that AI detectors are unreliable with no link to a named ground; anything about your grades, workload, or personal circumstances unless you are appealing the sanction's proportionality; anything you can't evidence.

The meeting or hearing

Most appeals are decided on the paper record. Where there is a meeting, it is usually short and narrow.

• Bring the advisor if you're permitted one, even if they may not speak. A second set of ears on what was said matters later.
• Answer the question asked. Volunteering "I only used it for grammar" turns an unproven allegation into an admitted one.
• Take notes with times and names. If you later discover the process wasn't followed, your contemporaneous notes are the evidence.
• Ask, before you leave, what happens next and by when. Berkeley's procedure commits to a written decision within 15 days of receiving the appeal; Clayton State within 7 business days. Knowing the return date tells you when to escalate.

The escalation ladder

Typical order, though names differ by institution:

1. Instructor (informal resolution)
2. Department chair or program director
3. Academic integrity office / dean of students — the first formal finding
4. First-level appeal: dean, appellate officer, or an appeals panel
5. Second-level appeal: provost, vice chancellor, or president. Berkeley routes to the Vice Chancellor for Student Affairs; UNC Charlotte to the Chancellor; BU from dean to provost
6. Student ombudsman — independent, confidential, and outside the chain; underused
7. External: your students' union or graduate student association, an education attorney, and for a matter of public record, the institution's grievance or accreditation channel

Two things worth knowing about step 4: a successful appeal usually produces a remand, not an exoneration — the case goes back to be heard properly. And several policies give you exactly one appeal. Shawnee State: "limited to one appeal and that decision is final."

What not to do

• Don't admit to end the discomfort. Clayton State's code states that students who accept responsibility "waive their right to a hearing and to an appeal." Lafayette's is narrower but the same in effect: for a student who accepted responsibility, the only remaining ground of appeal is inappropriate sanction. An admission is the one move that cannot be undone.
• Don't edit the flagged document. You are destroying your best evidence to lower a number that isn't the issue.
• Don't run your own writing through other detectors and submit the "clean" scores. Detector output is exactly the kind of evidence you are arguing is unreliable. It also won't hold: in our running measurement of detector accuracy, GPTZero and Vanguard agreed on which human sentences were "AI" at a Cohen's κ of just 0.213, and 23.3% of verified human sentences were flagged by at least one of three detectors. Citing a score you like while contesting one you don't undermines the whole letter.
• Don't rewrite the text with any rewriting tool, ours included, and submit that. It changes nothing about the allegation and it looks like tampering. We sell a rewriting tool and this is still the advice: it is the wrong instrument for this situation.
• Don't let it stay verbal. Follow every conversation with an email summarising what was said. If the process later goes wrong, the paper trail is what your procedural-error ground is built from.
• Don't miss the deadline while gathering perfect evidence. File on time with what you have; most procedures allow supplementary documentation after the notice of appeal.

If English is your second language

You are more likely to be flagged and less likely to be believed. Liang et al. (2023) found GPT detectors misclassified over half of TOEFL essays written by non-native English speakers as AI-generated, while classifying US-born eighth-graders' essays accurately. Our own sentence-level test found ESL learner writing flagged at 16.0%, second only to professionally edited journalism. Cite the study directly in the appeal; it is peer-reviewed in Patterns, and it is a stronger citation than any vendor blog.

Frequently asked questions

What do I do if a professor accuses me of using AI?

Before any finding is made, don't admit anything, ask in writing for the specific basis, freeze your document's version history, and read your institution's procedure so you know whether this is an informal conversation or step one of a formal process. The full pre-finding checklist is in our falsely accused guide.

Can a professor prove that I used AI?

Not from a detector score alone at most institutions. Turnitin's own guidance says its score is an indicator rather than proof, and Vanderbilt and UAB each disabled Turnitin's AI indicator and published their reasoning. Whether a score alone can sustain a finding at your institution is a question your written policy answers — and if the policy says it can't and the finding rests on it anyway, that is a procedural-error ground.

What if the AI detector was simply wrong?

That is a merits argument, and at 9 of the 16 policies we coded it is not by itself an appealable ground. Make it at the first meeting. On appeal, it has to be reframed: as new evidence (your version history, if it wasn't considered), as procedural error (the policy required corroboration and there was none), or as sufficiency of evidence if your institution is one of the seven that allows it.

How long do I have to appeal?

Between 3 and 10 days at every one of the 19 institutions we checked that publishes a deadline, with 5 days the most common. Check whether yours counts business or calendar days.

Sources

1. UC Berkeley, Student Conduct — Appeals. conduct.berkeley.edu/students/appeals
2. University of New Hampshire, Appeals Process & Procedures. catalog.unh.edu
3. Swarthmore College, Student Conduct Policies and Procedures. swarthmore.edu
4. Brandeis University, Rights and Responsibilities §18. brandeis.edu
5. UNC Charlotte, University Policy 407 — Code of Student Academic Integrity. legal.charlotte.edu/policies/up-407
6. University of Iowa, Accountability Procedures. dos.uiowa.edu
7. University of New Mexico, Student Disciplinary Procedure. pathfinder.unm.edu
8. Bowling Green State University, Code of Academic Conduct. bgsu.edu
9. Northern Kentucky University, Undergraduate Student Academic Integrity. nku.edu
10. Shawnee State University, Academic Misconduct. shawnee.edu
11. Clayton State University, Disciplinary Procedures. clayton.edu
12. Lafayette College, The Conduct Process. conduct.lafayette.edu
13. Chapman University, Student Conduct Procedures. chapman.edu
14. West Virginia University, Appeal Process. studentresponsibility.wvu.edu
15. SUNY ESF, Student Handbook — Sanctions and Appeals. esf.edu
16. Georgia Southwestern State University, Code of Conduct Procedures. gsw.edu
17. SUNY Empire State University, Student Conduct System. catalog.sunyempire.edu
18. UC Santa Cruz, Academic Misconduct. undergraduate.ucsc.edu
19. Boston University, Academic Conduct Code. bu.edu
20. Pratt Munson, Student Conduct Appeals. prattmunson.edu
21. CU Boulder, Honor Code and Student Code of Conduct. colorado.edu/sccr/honor-code
22. Turnitin, Understanding false positives within our AI writing detection capabilities. turnitin.com
23. Vanderbilt University, Guidance on AI detection and why we're disabling Turnitin's AI detector. vanderbilt.edu
24. University of Alabama at Birmingham, Ending the gen-AI detection war: turning off Turnitin's AI detection. sites.uab.edu
25. Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns 4(7), 100779. doi.org/10.1016/j.patter.2023.100779
26. Weber-Wulff, D. et al. (2023). Testing of detection tools for AI-generated text. International Journal for Educational Integrity. doi.org/10.1007/s40979-023-00146-z
27. GPTZero, Technology (stated ≤1% document-level false-positive rate). gptzero.me/technology
28. ToHuman, GPTZero false-positive rate across 861 human sentences
29. ToHuman, Do AI detectors work? The AI detector accuracy tracker (Q3 2026)

Methodology: on 2026-09-02 we pulled the published student-conduct or academic-integrity procedure for 21 US institutions, sampled from a Google site:edu query for pages containing both "grounds for appeal" and "procedural error," plus a small set of large public universities added by hand. Each page was read for four fields: the first-level appeal deadline, the enumerated grounds, whether an advisor or support person is permitted, and the stated standard of proof. Nineteen of the 21 publish a deadline and 16 enumerate grounds; the counts above use those denominators, which is why they differ between tables. Pages that render only via JavaScript were excluded rather than guessed at. This is a sample of published US procedures, not a representative census of American higher education, and it does not cover non-US institutions. Nothing here is legal advice.

Related reading

• Falsely accused of using AI: the full checklist — what to do before a finding is made, which is where the factual case belongs.
• AI accusation defense kit — the letter builder, evidence checklist and citation pack, in your browser, no account.
• We ran 861 human sentences through GPTZero — the false-positive number to cite, with the caveat that keeps it credible.
• Do AI detectors work? The quarterly accuracy tracker — inter-detector agreement, refreshed each quarter.
• Does Turnitin detect AI-generated content in 2026? — what the score is and isn't, plus what universities have published about it.

We build a rewriting tool, which means the honest thing to say here is that our product is not the answer to this problem — a documented process is. What we can do is put the templates, the evidence checklist, and the citations in one place you can work from: the AI accusation defense kit.