The Code Names the Standard and Says Who Carries the Burden

Read it off the code. The standard is a written text, not a norm the field shares. The University of Texas at Austin puts it in its rules on evidence, section 11-608: every hearing uses preponderance of the evidence, and Student Conduct carries the burden of proving the charges by it ⁹. Stanford's Charter of 2023 sets the bar higher, and sets it twice: clear and convincing evidence at Mid-Level Review, beyond a reasonable doubt at High-Level Review ².

Subsection 11-102(d) defines that standard as what is more likely to have occurred than not ⁹. A code that names a different standard names a different job for the panel.

Of the three codes read for this page, none names the standard beside the definition of the offense. Texas puts it in the rules on evidence in hearings, Stanford in the section on levels of review, where two of three panelists must agree at Mid-Level Review and four of five at High-Level Review ², and CUNY's academic integrity policy names no standard at all ³. An integrity policy and the hearing procedure under it carry their own revision dates, and the two can disagree about who decides what.

Goss v. Lopez, 419 U.S. 565 (1975), decided for Ohio public high school students suspended for up to ten days, holds that such a student gets oral or written notice of the charges and, on denying them, "an explanation of the evidence the authorities have and an opportunity to present his side of the story" ¹. The Court said longer suspensions and expulsions may require more formal procedures, and how far that floor reaches into a university hearing is a question for the institution's counsel. That is a floor on procedure, not a standard of proof.

The same alleged conduct can be decided under two different degrees of certainty, depending on the letterhead at the top of the page.

The institution brings the charge and supplies the evidence for it. Texas writes that down: Student Conduct has the burden of providing evidence and proving the charges ⁹. CUNY's policy was amended in 2022 and published in a 2024 version. It builds its hearing around the same shape: written notice of the charges, the right to appear, and the right to present or call witnesses ³. Nothing in it asks a student to prove authorship of their own paper.

The inversion is easy to see in a transcript and hard to see in the room. It starts when the first sentence of the meeting is a number, and every sentence after is a student accounting for it. The question has changed from what the institution can show to what the student can disprove. That question has no floor. Whether a detector score alone can support a finding is the narrower version of this, and it is worth settling in policy before it arrives in a room.

Saying three things out loud at the top of a hearing keeps the order straight:

• The charge, in the words of the code, with the clause number read out.
• The evidence the institution relies on, listed before the student speaks.
• The standard the panel will apply, named rather than assumed.

A student who then chooses to produce drafts is adding to a record that already exists. Ask for drafts before the charge has been described, and the student is carrying something no code handed them. A decision resting on that sequence stays fragile long after the sanction is served.

As one input, and a weak one standing alone. Turnitin said in 2023 that it does not make a determination of misconduct, even in the older space of text similarity, and that it provides data for educators to decide under their own policies ⁵. Pittsburgh's teaching center puts it shorter: AI detectors are not reliable enough to act as proof ⁶.

The vendor's own sentence-level figure is the part a panel needs beside a marked-up report. Turnitin puts its sentence-level false positive rate at around 4%, which it defines as a 4% likelihood that a specific sentence highlighted as AI-written might be human-written. It adds that the incidence climbs in documents containing a mix of human and machine writing, and climbs most at the transitions between the two ⁸. That is an in-house number, published by the vendor about its own product.

What a reading can support is narrow. It describes text. It carries no date and no device, no account and no intent, and a panel needs at least one of those. A percentage cannot say when the writing happened, who sat at the keyboard, or whether the student understood the argument they submitted. Corroboration a panel can check comes from elsewhere: the drafting record, the submission times the institution already holds, the student's account of their own argument, and the assignment's history in the course.

When those agree with the reading, the reading was never the case. When they contradict it, the reading was never the case either.

Pittsburgh also records that many detector vendors decline to publish a false positive rate at all ⁶. A panel writing any reading into the record can borrow the caveat that goes with Human's own readings. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating.

Everything the case would rest on if the detector output were struck. CUNY asks a faculty member who suspects a violation to review the facts and circumstances with the student first, wherever that is feasible ³. That conversation is where a record either forms or fails to, and the panel reads whatever survived it.

A record that meets a standard is made of things a third party can check months later.

The conversation is the part easiest to skip and the hardest to reconstruct later. A short oral follow-up after a flag that asks a student to walk through their argument gives a panel something to weigh. A Google Docs version history export does similar work from the other side, though it evidences a pattern of typing rather than authorship. The written decision says which of the two it relied on.

Detection sits outside every line of that list. It is a records question, and an institution can lose a case like this on appeal in the week before the hearing, when nobody wrote anything down.

To the grounds the institution's own procedure lists, and nowhere else. The University of Texas at Austin allows a conduct appeal on three, listed at Institutional Rules subsection 11-801(d): significant procedural error, discovery of new information that was not reasonably foreseeable to the accused student, and a sanction significantly disproportionate to the violation ⁴.

A hearing that let a percentage stand in for the case is closer to a procedural argument than an accuracy one, and Texas's list has a procedural ground and no accuracy ground ⁴.

An accuracy argument asks a panel to rule on a question it cannot settle, and it reads the same in every file. A procedural argument is specific: the code names a standard, the record does not meet it, and here is the line in the decision where a number did the work.

The cost of getting this wrong lands on one person. Jackie Davalos and Leon Yin reported for Bloomberg Businessweek in 2024 that Marley Stevens, a student at the University of North Georgia, said she was put on academic probation for a year after a disciplinary hearing determined she had cheated ⁷. She says she used only Grammarly's standard spell-checking and grammar features. A year on probation is not a clerical matter.

An institution designing an appeals process for detection cases can close most of this inside the procedure. Name the standard. Say in writing that a detector reading is not sufficient on its own. Require every written decision to describe what the case rests on with the reading set aside.