Score-as-Verdict: The AI-Detection Due-Process Ladder

Score-as-Verdict: The AI-Detection Due-Process Ladder

A framework for where your institution sits — and where the lawsuits are landing. Free to reuse (CC BY 4.0).

A student submits an essay. A detector returns a number — 74%, 100%, "likely AI." A grade is zeroed or a misconduct finding is entered. No second reader, no disclosure that a machine was involved, no real chance to respond. The number was the verdict.

We have a name for that pattern: Score-as-Verdict. It is the single most litigated mistake in AI-detection enforcement, and every few weeks it produces another reversal, appeal, or lawsuit. This piece maps the six ways schools fall into it, and lays out a five-rung Due-Process Ladder you can score your own institution against. It's built from the actual case record — see our running AI-cheating lawsuits tracker and the 170+ policies in the GradPilot AI Policy Observatory — and it's deliberately vendor-neutral. The point is not "which detector is best." The point is what you do with the number.

Why Score-as-Verdict is indefensible — statistically and legally

Start with the statistics, because they set the ceiling on what a detector score can prove:

• 750 of 75,000. When Vanderbilt disabled Turnitin's AI detector in August 2023, it did the arithmetic: at Turnitin's own claimed ~1% false-positive rate, roughly 750 of the 75,000 papers submitted in a year could be wrongly flagged. A tiny percentage times a large denominator is a lot of falsely accused students.
• 61.3%. A peer-reviewed Stanford study (Liang et al., Patterns, 2023) found detectors flagged 61.3% of non-native-English (TOEFL) essays as AI while classifying US-student essays near-perfectly — and that a little vocabulary enrichment cut the false-positive rate to 11.6%. Translation: the tools partly measure fluency, not authorship. (Caveat: 2023-era detectors, small sample — but the direction is well-replicated. More in our explainer on why detector false positives are effectively unavoidable.)
• 94%. In a blind test at the University of Reading (Scarfe et al., PLOS ONE, 2024), 94% of fully AI-written exam submissions went undetected — and outscored real students 83% of the time. Detectors miss the guilty even as they flag the innocent.
• OpenAI shut down its own detector in July 2023 for low accuracy (26% true-positive, 9% false-positive). The maker of ChatGPT could not reliably detect ChatGPT.

Even the vendors concede the point. Turnitin's own guidance says the AI indicator "should start a review rather than end one" and does not itself determine misconduct. A probability score is a screening signal. Treating it as proof — Score-as-Verdict — asks a number to carry a burden its own maker won't put on it.

Then there's the law. Where students have been disciplined on a detector score alone, courts and ombudsmen have been increasingly willing to intervene — not because "AI detectors are inadmissible," but because the process was defective. The dividing line in the litigated cases is almost always corroboration: schools that had independent human evidence tend to win; schools where the score was the whole case tend to lose.

The six ways institutions fall into it

Across the disputes we track, the same failure modes recur. Most blown-up cases show two or more at once.

One correction worth making loudly, because nearly every viral write-up gets it wrong: the widely cited Adelphi student win (Newby) was a New York State Article 78 proceeding — not a federal precedent. It binds no one outside that record; its reasoning is that the school failed to follow its own written procedures. In fact the litigated federal scoreboard so far is mixed-to-school-favorable (the University of Minnesota won its case; preliminary injunctions were denied in the Yale and Hingham matters). The clear student wins have come through state court, internal appeals, and ombudsmen — which is exactly why process, not the tool, is the lever. For the myths that have grown up around these rulings, see AI-cheating lawsuit myths.

The Due-Process Ladder

Each rung closes specific failure modes. A school's rung is the highest level it fully meets. Most institutions using detectors today sit at Rung 0 or 1.

Why these anchors matter: this isn't GradPilot inventing a standard. The EU AI Act already classifies education uses — explicitly including "detecting prohibited behaviour during tests" — as high-risk, and requires human oversight before an AI assessment becomes final (its high-risk education obligations were deferred to December 2027 — see the new AI-admissions laws in Colorado, California, and the EU). NIST's AI Risk Management Framework, which the US Department of Education tells institutions to adopt, names transparency, human oversight, and bias management as core. The ladder just sequences those recognized principles into rungs you can actually place a school on.

Score your institution

Five yes/no questions. Each "no" caps you at the rung below it.

1. Disclosed: Does your written policy tell students a detector may be used, and name it? (No → Rung 0.)
2. Human-Reviewed: Is it impossible to receive a penalty on a detector score alone — is corroborating evidence and a human judgment required? (No → capped at Rung 1.)
3. Contestable: Does the accused get notice, see the evidence, and get an appeal to someone other than the accuser? (No → capped at Rung 2.)
4. Bias-aware: Are there accommodations and an audit for ESL, disabled, and neurodivergent writers who are false-flagged at higher rates? (No → capped at Rung 3.)
5. Validated: Would your institution publicly stand behind the specific tool's reliability for your student population? (No → you have an FM-6 problem at any rung.)

If you can't answer "yes" to #2, you are one motivated student and one bad flag away from the fact pattern in half the cases above.

The market already moved — the policies haven't

Two things are true at once, and the gap between them is the story.

The vendors have pivoted from "detect" to "defensible process." Turnitin launched Clarity, a drafting environment that captures writing process rather than just scoring output. GPTZero was absorbed into an "authenticity layer." Packback markets "detection to prevention"; Cadmus sells "authentic assessment." The commercial center of gravity has moved to human-in-the-loop, provenance, and defensibility — Rung 2-and-up language. (There's a real privacy trade-off in keystroke/draft monitoring worth watching, but the direction is clear.)

Institutional policy hasn't kept up. Only about 23% of institutions report having any AI acceptable-use policy at all (EDUCAUSE, 2024), and detector use often runs silently inside the LMS with no student-facing disclosure of the threshold or the appeal path. That vacuum — documented tools, undocumented process — is precisely the enforcement gap that turns a flag into a filing.

A handful of institutions show what the top rungs look like in practice: Washington State University's policy states that suspicion of AI use "is not sufficient for a finding of student responsibility" and builds in notice, a meeting, and appeal (Rung 3). Vanderbilt, MIT, and UT Austin stopped or declined to endorse detection outright, citing unreliability and bias (a Rung-4 posture) — part of a wave of 60+ colleges that turned detectors off. They are not anti-integrity; they moved the burden off the number. For the fuller picture of who uses what, see do colleges actually use AI detectors.

Use this framework

The Due-Process Ladder and the Score-as-Verdict taxonomy are released under Creative Commons Attribution 4.0 (CC BY 4.0) — reprint them, adapt them, drop the table into your deck or policy memo. Attribution to this page is all we ask. We'll revise as the case law develops; if you know of a case, policy, or ruling we should fold in, tell us.

---

GradPilot is an independent essay-feedback platform. This is a descriptive governance framework and journalism, not legal advice; institutions should consult counsel on their own policies. We are vendor-neutral: accuracy figures here are drawn from independent and peer-reviewed research and institutions' own disclosures, not from detector vendors' marketing. Case details are dated to mid-2026 and several matters remain pending; verify specifics against the linked primary sources before relying on them.

Sources

• Vanderbilt University, "Guidance on AI Detection and Why We're Disabling Turnitin's AI Detector" (Aug 16, 2023) — the 750/75,000 math.
• Liang, Zou et al., "GPT detectors are biased against non-native English writers," Patterns (Cell Press), 2023 — the 61.3% ESL false-positive figure.
• Scarfe et al., PLOS ONE (University of Reading), 2024 — 94% of AI exams undetected.
• EDUCAUSE 2024 AI Landscape Study — ~23% of institutions have an AI acceptable-use policy.
• Common Sense Media, "The Dawn of the AI Era" (2024) — Black teens ~2× more likely to report a false AI accusation.
• Turnitin, "Understanding false positives" and AI Writing Report guidance — "start a review rather than end one"; document/sentence false-positive rates.
• NIST AI Risk Management Framework (2023); US Dept of Education, OET AI toolkit (2024).
• EU AI Act, Annex III(3) and Articles 14 & 26 — education/exam-monitoring as high-risk; human oversight; disclosure.
• ATIXA — Title IX complaint-resolution model (investigator/decision-maker separation).
• Washington State University — AI academic-integrity policy ("not sufficient for a finding").
• Case record: Matter of Newby v. Adelphi (NY Sup. Ct., Article 78, 2026); Yang v. Univ. of Minnesota (D. Minn. / MN Ct. App.); Harris (RNH) v. Hingham (D. Mass.); Rignol v. Yale (D. Conn.); Doe v. Univ. of Michigan (E.D. Mich.); Kato v. Palo Alto USD (N.D. Cal.); UK OIA case summaries (2025). See our lawsuits tracker for dockets and outcomes.