Should AI Detection Scores Be Disclosed to the Accused Student?

Yes — a student accused of AI use should see the full detection report, the tool and version that produced it, and its known error rates, because an accusation you cannot examine is one you cannot answer, and every serious conception of academic due process starts with access to the evidence. The harder questions are what “disclosure” has to include for it to mean anything, and why some institutions still resist — so let’s take those in order.

Key takeaways

• Basic fairness and most schools’ own integrity codes require showing accused students the evidence; a detection score is evidence like any other.
• A bare percentage is not meaningful disclosure — students need the flagged passages, tool and version, and documented false-positive context to respond at all.
• The tools’ error profiles make scrutiny essential: OpenAI retired its own classifier, Stanford research found heavy bias against non-native writers, and Vanderbilt disabled Turnitin’s detector outright.
• The anti-disclosure arguments (gaming, vendor secrecy, hearing friction) are weak precisely where the stakes are highest.
• Students accused on an unseen score should request the report in writing and build a process-based defense — version history beats algorithm arguments.

The baseline: you can’t rebut a number you’ve never seen

Strip away the technology and the situation is old. An institution accuses a person of an offense, citing evidence. Every integrity code worth the name — and, at U.S. public institutions, the constitutional floor sketched in cases like *Goss v. Lopez* — gives the accused notice of the evidence and an opportunity to respond. Nobody would accept “our expert says you cheated, but you may not see the expert’s report” for plagiarism, or lab misconduct, or anything else with transcript-level consequences.

An AI detection score is just evidence with worse-than-average reliability, which makes concealment *more* indefensible, not less. FERPA adds a practical lever in the U.S.: students may inspect their education records, and a detection report attached to their integrity case is plausibly exactly that. Yet the common failure mode isn’t a policy that says “hide the score” — it’s an instructor who says “Turnitin flagged this at 89%” in an email and treats the number as self-evident. The student is then negotiating with a verdict, not examining evidence.

What disclosure has to include to be real

Here’s where well-meaning institutions cut corners. Showing a student “AI: 89%” discloses almost nothing. For the score to function as examinable evidence, disclosure needs five parts:

1. The full report — which passages were flagged, at what confidence. This is what lets a student respond *specifically*: here’s the version history for that section, here’s why my methods paragraph reads formulaic.
2. Tool and version. Detectors change constantly; a score is a product of a particular model on a particular day. An unnamed tool can’t be researched, and its error rate can’t be cited.
3. The vendor’s own caveats. Turnitin has publicly discussed its false-positive rate and cautioned that its indicator shouldn’t be the sole basis for action. The accused student is entitled to have the panel read the vendor’s fine print, not just its headline number.
4. Reproducibility information. Was the paper scanned once? Did rescans give different numbers? Detection scores are not stable across runs and versions, and instability is relevant to weight.
5. The human interpretation. Who decided the score merited an accusation, and by what threshold? “The tool said 89” is not a decision procedure; someone applied a judgment, and that judgment is contestable.

This is the difference between disclosure as ritual and disclosure as due process. We’ve written more broadly about what due process looks like in an AI integrity hearing, and about whether detection scores should be evidence at all — disclosure is the minimum position even people who answer “yes, evidence” should accept.

Why the reliability record makes scrutiny non-negotiable

The case for transparency strengthens every time the tools are tested. OpenAI retired its own AI Text Classifier in July 2023, citing low accuracy — the maker of ChatGPT concluding it couldn’t reliably detect ChatGPT. A 2023 Stanford study (Liang et al., published in *Patterns*) found popular detectors flagged more than half of a set of non-native English speakers’ TOEFL essays as AI-generated while classifying native-speaker essays almost perfectly — a bias with obvious, ugly implications for exactly the students most likely to face accusations. Vanderbilt University disabled Turnitin’s AI detector in August 2023, explaining publicly that it couldn’t verify the tool’s claims and wouldn’t subject students to unverifiable accusations.

None of this proves any individual score wrong. It proves that scores *as a class* have known, documented failure modes concentrated on identifiable groups — careful formulaic writers, ESL students, neurodivergent students taught to write to template. Evidence with that profile can still inform a process, but only a process where the accused can see it, contextualize it, and put error rates in front of the decision-maker. Concealment converts a noisy signal into an unanswerable verdict.

The arguments against disclosure, weighed

“Students will learn to game the detector.” The accused student already submitted the work; showing them the report teaches them nothing they can retroactively exploit. As for future gaming — detector behavior is already probeable by anyone with a free account; anyone can check their own writing with a free AI detector before submitting. Secrecy toward the accused buys no security; it only buys unaccountability.

“Vendor terms restrict sharing.” Then the institution has purchased evidence it can’t use fairly, which is a procurement problem, not a student’s problem. A tool whose output can’t be shown to the person it accuses has no place in a disciplinary file.

“Every hearing will turn into a fight about the algorithm.” Some friction is the price of using probabilistic evidence on high-stakes questions. But in practice, disclosure *reduces* fights: students who see that only two paragraphs were flagged respond narrowly with drafts and history, panels apply the vendor’s own caveats, and weak cases die before hearings instead of during appeals — or lawsuits.

For students on the receiving end right now: request the report in writing, then build the affirmative case — version history, dated drafts, your ability to discuss the work cold. There’s a step-by-step appeal plan for accused students covering exactly this, and more on detection and academic integrity across the blog.

Frequently Asked Questions

Do students have a right to see their AI detection score?

Usually yes, in practice if not always by that name. Most institutional integrity codes promise accused students access to the evidence against them, and public universities in the U.S. owe students basic due process — notice of the charge and the evidence, plus a chance to respond — under cases like Goss v. Lopez. Records-access laws add another route: FERPA gives students the right to inspect their education records, which an AI report attached to their case plausibly is. The gap is that some instructors informally accuse based on a score they never show. If that happens to you, politely requesting the full report in writing is both reasonable and usually effective.

What exactly should a school disclose along with the score?

A bare percentage is close to meaningless, so meaningful disclosure includes: which tool and version was used, the full report showing which passages were flagged, the score’s documented false-positive rate and the vendor’s own caveats, whether the document was scanned once or multiple times with different results, and who interpreted the score and how it factored into the accusation. Turnitin itself has said its AI indicator shouldn’t be the sole basis for an accusation, and that context belongs in front of the student and the panel. Without those pieces, the student is being asked to rebut a number nobody can explain.

Why do some institutions resist showing students the score?

The stated reasons are usually gaming — if students see exactly what was flagged, they’ll learn to evade the tool — plus vendor confidentiality and a fear that scores will be picked apart in every hearing. The gaming argument is the weakest: detection scores are already easy to probe from the outside, the accused student has already submitted the work, and due process doesn’t get suspended because evidence is inconvenient to defend. The picked-apart fear is actually an argument for disclosure — evidence that can’t survive scrutiny shouldn’t decide a student’s transcript. Institutions confident in their process show their work.

Can a student challenge an AI detection score once they see it?

Yes, and effectively — that’s precisely why disclosure matters. Detector scores are probability estimates with documented failure modes: OpenAI retired its own classifier in 2023 for low accuracy, a 2023 Stanford study found detectors flagged more than half of non-native speakers’ TOEFL essays as AI, and Vanderbilt disabled Turnitin’s AI detector over reliability concerns. A student who can see which passages were flagged can respond specifically: show version history and drafts for those sections, explain stylistic factors like ESL phrasing or formulaic academic tone, and ask the panel what error rate the tool has on writing like theirs. None of that is possible against an undisclosed number.

What should I do if I’m accused based on a score I haven’t seen?

Ask, in writing, for four things: the full detection report, the tool name and version, the institution’s policy authorizing its use, and the procedure for contesting the finding. Stay polite and factual — you’re invoking process, not picking a fight. In parallel, assemble your affirmative evidence: version history, dated drafts, notes, browser history from writing sessions, and your ability to discuss the work fluently. If the process still won’t show you the score, escalate through the appeal channel and cite your school’s own evidence-access language, which almost every integrity code contains. Most cases are won by process and paper trail, not by arguing about the algorithm.

The bottom line

A detection score that stays hidden isn’t evidence — it’s an oracle, and oracles have no place in a process that can end a degree. Disclose the report, the tool, and the error rates, and the score becomes what it always should have been: one noisy signal, weighed by people who can see it, answered by a student who can finally respond to what’s actually being claimed.