Appeals and False Positives: Why Students Are Winning

What Appeal Data Can Actually Show

The originally cited coaching page provides guidance on oral defences, not a verifiable study of appeal outcomes. It does not publish the numerator, denominator or selection method needed to support an 89% success claim. For now, the appropriate conclusion is that a general success rate has not been established.

What it reflects is a fundamental evidentiary asymmetry. Plagiarism was historically objective. You could point to the source document. The match was there or it wasn't. AI generation is entirely different. There is no source document. There is no origin point. There is a probabilistic percentage from a proprietary algorithm that cannot name the AI tool used, cannot identify when the text was generated, and cannot distinguish between a student who used AI and a student who writes formally.

Institutional guidance is clearer than a generic success statistic: Cornell's Center for Teaching Innovation recommends gathering objective, verifiable evidence and speaking with students before advancing an academic-integrity allegation.

An appeal outcome depends on the facts and the institution's burden of proof. A detector score may prompt questions, but it should not decide the outcome without corroborating evidence.

Why the Score Alone Fails in a Formal Hearing

Academic integrity review panels, when functioning correctly, apply a burden of proof. The institution makes a claim. The institution provides evidence. The student responds. The panel decides.

A detection score is evidence that a text's statistical properties resemble AI output. It is not evidence of authorship. Those are different claims. The panel's job is to determine authorship. The score cannot do that.

When a student arrives at a hearing with Google Docs version history showing 47 distinct editing sessions across 8 days, rough drafts from two weeks prior, a browser history showing research sessions on the relevant dates, and the ability to explain the hardest part of their argument under mild questioning, the panel has a direct conflict between probabilistic software output and human-documented process. The human-documented process wins because it addresses the actual question. The software output does not.

Cornell University's Generative AI Academic Integrity Guidelines, published in 2025, named exactly this standard: "Faculty should follow up their initial concerns about a generative AI-related violation with tangible evidence, such as the inclusion of advanced references that exceed the scope of the course, or situations where the student is unable to verbally explain, expand upon, or defend their submitted work when queried." That guidance acknowledges what the appeal data confirms: a detection score is not tangible evidence. It is a trigger for further investigation.

When suspicion arises, the most reliable approach is a structured conversation targeting the cognitive experience of writing, not a second detection pass. The four-question oral framework that reveals authentic engagement is far more defensible in a hearing than any dashboard percentage.

The Columbia Data: Reported AI-Use Charges

Columbia University's Center for Student Success and Intervention recorded 79 charges involving unauthorized AI-tool use in its 2024–25 annual report. That was the second-most-reported academic-integrity charge category, behind cheating (135). The report does not say that 79 appeals were filed or describe an 89% success rate.

These are charges reported by one university in one academic year. They cannot tell us how many appeals were filed, upheld or overturned elsewhere.

The numbers in this table measure different things: institutional charges, survey responses and procedural rules. None supplies a reliable national AI-misconduct appeal-success rate, and they should not be combined to estimate one.

Gengyan Tang, a researcher on academic misconduct services at LSE, described what has emerged in 2026: "What is often described as 'academic misconduct appeal assistance' is no longer a marginal service. It has evolved into a platform-oriented business model that systematically captures students precisely when they are likely to feel uncertain or anxious." The commercial defence industry is a symptom of an institutional failure to build a defensible process.

The Administrative Cost Schools Aren't Counting

The visible cost of an AI misconduct appeal is the hearing itself. The hidden cost is everything that leads to it and everything that follows.

Filing a formal charge requires the instructor to produce charging documents, provide relevant syllabus statements, document the timeline of their intervention, and in most institutions participate in at least one formal hearing. For a single case, that is several hours of work. For an institution running detection across tens of thousands of submissions, the false positive rate generates a caseload that no administrative office is resourced to handle at full quality.

The result is predictable and well-documented in educator forums: instructors stop reporting. They see a suspicious paper, they run the detection tool, they get a flag, they calculate the documentation burden, and they decide it isn't worth it. They grade the paper or assign a lower mark without a formal process. This behaviour defeats the entire purpose of running detection. The tool produces anxiety and administrative waste without producing accountability.

Investigating authorship questions can create a real administrative workload: educators must review the available records, give students a chance to explain their work, and document decisions consistently. The size of that burden should not be assigned a general percentage without a directly supporting survey.

What Institutions Must Fix Before the Next Appeal Season

The research from 2024 to 2026 produces a clear institutional checklist. These are the structural fixes that reduce both false accusation rates and successful appeals against legitimate ones.

Make process documentation part of the grade from day one. Requiring outlines, annotated bibliographies, and draft histories as graded components of every major assignment builds the evidentiary record before any suspicion arises. If a student fails to provide process documentation, the grade reflects that absence. No accusation required. No appeal possible on those grounds.

Policies should tell students, assignment by assignment, which AI uses are permitted, which require disclosure and which are prohibited. Clear expectations give both students and review panels a fair point of reference.

Train faculty on what constitutes actionable evidence. A detection score is a trigger for investigation, not a verdict. Faculty who understand this distinction approach the initial conversation correctly. Faculty who treat a dashboard percentage as proof risk making allegations that lack sufficient corroborating evidence.

Decide institutionally on detection, not individually. A teacher who disables detection in their own workflow but whose school still runs it institutionally is still exposed to escalation from outside their classroom. The decision to continue or discontinue AI detection belongs at the administrative level, with a published policy that teachers can reference when a parent or colleague escalates a flag.

Building the evidentiary record before any suspicion arises removes the adversarial dynamic entirely. Schools that have built process documentation into their assignment structure don't end up in appeal hearings, because there's nothing to contest, and the reason detection-first approaches consistently fail is that they invert that logic, chasing evidence after the fact rather than requiring it upfront.

Frequently Asked Questions