A Detector Score Alone Does Not Meet a Code's Standard of Proof

Not on the score alone. Discipline follows a finding, and a finding has to meet a standard of proof. A detector reports a probability about text. It establishes nothing about a person. The vendor agrees: Turnitin states that it does not make a determination of misconduct, and that the decision rests with the educator and institutional policy ².

A score describes a document; a finding describes what a student did.

Nothing inside the report closes that gap. The University of Pittsburgh's Teaching Center says so without hedging: AI detectors are not reliable enough to act as proof ³. The same center switched off AI detection inside Turnitin and endorses no AI detection tool at all ³.

Here's a working test for a case file, and it takes about 10 seconds. Take the detector report out and read what's left. Nothing there? Then there's no case yet, only a suspicion nobody has worked.

Suspicion is a fine reason to look closer and a poor reason to charge.

Scale matters too. A false positive rate that looks small on a vendor page turns into a queue of wrongly flagged papers once it's multiplied by the human-written papers in a term. That step assumes a rate measured on the vendor's test set holds on one school's pile. The arithmetic, and its assumption, are worked through in what a false positive rate means across a whole term.

Whatever the institution's own code says, and the answer is usually one sentence long. The University of Minnesota's Board of Regents policy states that a finding of responsibility for violation of the Student Conduct Code must be based on a preponderance of the evidence ¹. Read that sentence before the hearing, not during it.

Codes differ. Some set a higher bar, so quote the institution's own words, not a half-remembered phrase.

On weight, Turnitin's own published numbers settle it. The company puts its document-level false positive rate below 1% for documents its model reads as 20% or more AI writing ⁴. At the sentence level the rate is around 4%. That's roughly 1 highlighted sentence in 25 turning out to be human-written, and Turnitin tells instructors to use the information to initiate a conversation rather than to draw a conclusion ⁴.

Two more details from the same vendor matter at a hearing. In May 2023 Turnitin raised the minimum length for an AI writing reading from 150 words to 300 words, so a short submission gets no reading at all ⁸. And since 8 July 2024, the report attributes no score and no highlights at all below 20%, showing an asterisk in place of a percentage ⁹.

A panel asked to find a violation on the report alone is being asked to turn a measurement into a conclusion. The company that made the measurement won't draw that conclusion itself. The standard in the code stops it, but only if somebody reads the standard out loud in the room.

Would a better detector change the panel's position? Less than a panel might hope. Human, a detector that gives a document a Human, Mixed or AI verdict, reports, as of September 2026, 0 wrongly flagged by Human's default model of 600 human-written papers it had never been trained on; the ceiling on that is 0.498%. That figure is Human's own in-house measurement on its own held-out sets rather than an independent audit. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. The measured operating points are on Human's research page. A lower published error rate changes how often a flag is wrong, never what a flag can carry.

Corroboration is evidence about how this particular paper came to exist, drawn from somewhere other than the detector: the drafting record, the student's own account of the argument, and what the same student has written before. None of it proves authorship on its own either. Put together, though, it's a record a panel can weigh.

William Quarterman, a UC Davis senior, cleared himself with the time-stamped edit history of the Google document he wrote in ⁷. It showed the exam response taking shape over two or three hours ⁷. That's what corroboration looks like: a record made while the writing was happening, kept somewhere the detector never touched.

Four kinds of record do the work.

• Version history exported with timestamps, showing a paper building up over days instead of arriving in one save.
• Drafts and research notes, including the page with a thesis crossed out.
• A meeting in which the student explains the argument and says where the sources came from.
• Earlier graded work from the same course, read for range and habit.

Four more things look like corroboration and aren't.

• A second detector agreeing with the first. Two readings of the same text are still readings of text, not evidence about the student.
• A percentage above a threshold the department invented for itself.
• An instructor's sense that the prose reads too evenly.
• A missing draft, when nothing in the assignment ever asked for one.

CUNY's code puts that meeting first. It tells a faculty member who suspects a violation to review with the student the facts and circumstances of the suspected violation whenever feasible ⁶. A thin case gets its first test there, and the oral follow-up after a flag works through how to run it.

If the student drafted in a browser-based editor, the record usually exists already, and nobody has asked for it. Exporting the version history as an exhibit turns a live document into something a panel can still read weeks later.

One sentence, placed in the integrity policy rather than in a teaching center FAQ, because a hearing reads the policy. A workable form: a report from an AI writing detector may prompt an inquiry and may be weighed alongside other evidence, and it may not by itself support a finding of responsibility.

Name the standard of proof in the same paragraph. Name who decides, too.

Five things the clause should settle, because each one has already caused an argument somewhere:

• Who may run a submission through a detector, and whether the student is told it happened.
• What the report is called in the file. An indicator, not evidence of authorship.
• What has to accompany the report before a charge can be filed.
• Which standard of proof applies, quoted from the code rather than paraphrased from memory.
• Who makes the finding, and on what record.

Put it in the code first. Then cross-reference it from the syllabus language, the LMS help page and whatever training an instructor gets in August. Why the code? A clause that lives only in a teaching center post is guidance, while a clause in the code is procedure, and procedure is what an appeal gets measured against. Drafting it is worked through in the integrity clause that names AI.

Keep two things out of the clause: a named percentage and a vendor name.

Write a threshold into policy and it becomes the thing every hearing argues about. It also goes stale the moment the model behind the score is retrained. Turnitin added its under-20% asterisk in May 2023 ⁸ and stopped printing any number beneath it in July 2024 ⁹, which is the kind of correction a vendor can make and a code can't. A vendor name goes stale faster, and it reads as an endorsement the procurement file may not support.

It hands the student a ready-made appeal wherever the code allows one on procedural grounds: the record doesn't meet the standard the code names. The file holds one automated reading and nothing else, and that's the whole argument. Where the burden sits is covered in the burden of proof at a misconduct hearing.

The student also arrives with procedural rights. CUNY's procedures give a student who denies the charge written notice of the charges, the right to appear and the right to present or call witnesses ⁶.

What moves a provost is volume. Vanderbilt ran the arithmetic on its own submissions. In 2022, 75,000 papers went to Turnitin, and at the 1% document-level false positive rate the vendor had published, around 750 student papers could have been incorrectly labeled as partly AI-written ⁵. Vanderbilt disabled the detector in August 2023, saying it did not believe AI detection software was an effective tool ⁵.

One wrongly charged student is a hearing. Put 750 papers that could have been wrongly flagged together and it's a policy failure with an institution's name on it. Writing the corroboration rule into the code is not generosity toward students; it is a cheap limit on the cases an institution cannot defend.

A vague appeal route slows the charge and the challenge alike, and the record decays while everyone waits. Designing the appeal path before the first case belongs in the same revision as the clause.