Before Running an Essay Through an AI Detector: A Fair-Process Checklist for Teachers

On November 20, 2024, US Magistrate Judge Paul G. Levenson denied a preliminary injunction in Harris v. Adams (No. 24-cv-12437-PGL, D. Mass.), upholding Hingham High School's penalty on an AP US History project in Massachusetts (Gizmodo, Nov 22, 2024). Hingham won because the teacher set written Week-1 AI rules, the students copied uncredited prose from Grammarly's AI feature, and that text cited nonexistent books. On January 28, 2026, New York Supreme Court Justice Randy Sue Marber annulled Adelphi University's penalty against freshman Orion Newby as "devoid of reason" (Matter of Newby v. Adelphi University, Nassau County) and ordered expungement (GradPilot, Jul 10, 2026; Plagiarism Today, Oct 14, 2025). There, an instructor relied on a 100% Turnitin AI report score and a hunch that prose looked "too advanced" while ignoring 15 to 20 hours of tutored drafts in a campus neurodiversity program.

Grading 150 papers on Sunday night with zero paid investigation hours traps K-12 teachers, paraprofessionals, adjuncts, and department chairs. Schools demand integrity enforcement, LMS dashboards flash probability gauges, and state agencies omit investigation procedures. In Massachusetts, the Department of Elementary and Secondary Education (DESE) AI hub (updated August 26, 2026) promotes AI literacy yet sets zero rules on AI detectors, hearings, or paraprofessional duties (MA DESE; WBUR, Sep 19, 2025). Meanwhile, on September 2, 2025, the FTC finalized a 3-0 order barring detector vendor Workado from claiming unsubstantiated accuracy rates (MyChesCo, Sep 2, 2025).

This guide gives educators base-rate math, FERPA rules, a six-step inquiry protocol, and syllabus clauses as of October 29, 2026. Because ed.cafe at No Worker Left Behind is an education nonprofit rather than a law firm, use this as an instructional workflow rather than legal advice. For students responding to allegations, share our guide for students and parents accused of using AI.

Why Turnitin, GPTZero, and Originality.ai warn against solo-score discipline (and the base-rate math)

Every major AI-detection vendor warns schools never to use a detector score as standalone proof of misconduct because statistical classifiers estimate token predictability rather than matching copied lines to a source text.

Classic plagiarism software matches verbatim strings side by side against an indexed article. AI detectors cannot do that: model providers do not share private chat histories, and pasted plain text carries no cryptographic watermark. Instead, detectors score perplexity (word-choice predictability) and burstiness (sentence-length variation). Clean, rule-following academic prose lowers both metrics and inflates AI probability scores.

Treating an AI detector score like a plagiarism match confuses a hallway photoelectric smoke alarm with security-camera footage. Plagiarism matching hands you video of the exact library shelf; an AI detector is a ceiling sensor that rings equally loud for real fire, burnt toast, and locker-room shower steam. Where that smoke-alarm analogy breaks down is verification speed: a principal sees steam in a hallway within 30 seconds, whereas a submitted essay has no visible smoke, tempting tired graders at 10:30 PM to treat the alarm bell itself as the fire.

Read the vendors' own published limits:

• Turnitin: States its AI indicator "must not be used as the sole basis for adverse actions against a student," notes an approximate 4% sentence-level false-positive rate in vendor guides, and masks document scores from 1% to 19% behind an asterisk (*%) due to elevated low-range false positives (Turnitin FAQ; Turnitin Guide).
• GPTZero and Originality.ai: Both state in educator documentation that scores are probabilistic signals unsuitable as standalone proof for academic penalties.
• OpenAI: Shut down its own AI Text Classifier on July 20, 2023 after testing showed a 26% true-positive rate paired with a 9% false-positive rate on human essays (OpenAI Notice, Jul 20, 2023).

Check the base-rate math across a high school English department or community-college composition program grading 2,000 essays per semester (assumptions stated: 1,850 authentic student essays, 150 unauthorized AI essays [7.5% violation rate], and an 80% true-positive catch rate [120 AI essays caught]):

Even at a 1% false-positive rate across 1,850 authentic essays, nearly one in seven flagged students (13.4% in the 2,000-essay cohort) wrote their own work. At a 4% false-positive rate, 38.1% of flagged students are innocent, consuming 55.5 uncompensated faculty hours in tense hearings. Compare cross-industry error math in how often AI is wrong by task.

What Hingham High School and Adelphi University teach about syllabus rules and corroborating proof

Court rulings from 2024 through 2026 draw one bright operational boundary: schools win when instructors pair advance written AI rules with non-detector proof, and lose when a solo detector score replaces reading drafts.

• Why Hingham High School won (Harris v. Adams, No. 24-cv-12437-PGL, D. Mass., Nov 20, 2024): Judge Levenson denied an injunction over an AP US History penalty because three facts backed the teacher: (1) explicit first-week classroom rules barred uncredited AI generation; (2) the students admitted pasting text from Grammarly's AI feature; and (3) the pasted script cited nonexistent books, including a phantom volume by "Jane Doe" (Gizmodo, Nov 22, 2024).
• Why Adelphi University lost (Matter of Newby v. Adelphi University, N.Y. Sup. Ct., Jan 28, 2026): Relying on a 100% Turnitin AI report score and a claim that prose looked "too advanced" while ignoring 15 to 20 hours of Bridges Program tutoring drafts and 0% AI report checks from Grammarly and ZeroGPT was annulled under CPLR Article 78 as "devoid of reason" (Plagiarism Today, Oct 14, 2025).
• Why the University of Minnesota prevailed (Yang v. Regents of the University of Minnesota, affirmed Feb 2026): On a closed PhD exam banning ChatGPT, four professors documented un-taught seminar concepts and an unusual acronym (PCO) matching live ChatGPT test outputs (GradPilot Tracker, Jul 10, 2026).

Disparate false-positive risk for multilingual learners, IEP/504 accommodations, and Grammarly rewrites

Unchecked detector scans fall hardest on multilingual English learners, neurodivergent students taught structured writing formulas, and students using grammar extensions pushed onto school laptops by district IT.

In July 2023, Stanford researchers Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu, and James Zou published a study in Patterns (arXiv:2304.02819) testing seven GPT detectors on 91 human-written TOEFL essays alongside 88 US eighth-grade essays (Liang et al., arXiv:2304.02819):

• Across the seven detectors, 61.22% of the 91 TOEFL essays (often cited as 61.3%) were misclassified as AI-generated on average.
• 97.8% (89 out of 91 essays) drew at least one false AI flag, and 19.78% (18 out of 91 essays) were unanimously mislabeled as AI by all seven tools, compared with near-zero flags on eighth-grade essays.
• State both scientific caveats clearly: First, the TOEFL sample was small (n = 91 short test essays). Second, the study evaluated 2023 detector versions prior to 2024-2025 recalibrations. Cite Liang et al. for the linguistic mechanism (standardized transitions and high-frequency academic vocabulary lower perplexity), not as a blanket 2026 error figure.

Three everyday school practices trigger that exact low-perplexity signal:

1. IEP, Section 504, and neurodivergent scaffolds: Special educators and paraprofessionals teach Claim - Evidence - Reasoning templates that produce symmetrical sentences, the exact dynamic behind Newby v. Adelphi and Doe v. University of Michigan (GradPilot Tracker).
2. Approved dictation accommodations: Speech-to-text tools paired with grammar cleanup flatten sentence burstiness for students with dysgraphia or motor disabilities.
3. District-pushed Grammarly, Microsoft Editor, and Apple Writing Tools: Clicking "Rewrite for clarity" replaces student syntax with low-perplexity LLM clauses unless syllabi distinguish basic spellcheck from generative rewriting. At ed.cafe, where No Worker Left Behind is forming Peer Circle #1 in Boston (30 free chairs via the Massachusetts AI Hub; applications close December 1, 2026 at ed.cafe/start), adult multilingual learners and paraprofessionals see this confusion weekly.

FERPA, student privacy, and why teachers must never paste essays into free web detectors

Pasting a student's essay into a free web "AI checker" or a personal ChatGPT account ("Did AI write this essay?") risks violating federal privacy rules under FERPA (20 U.S.C. § 1232g) and undercuts any subsequent conduct hearing.

Under FERPA (34 C.F.R. § 99.31), identifiable student coursework is a protected education record (StudentPrivacy.ed.gov). Schools may share coursework with outside vendors without written parental consent only under the "school official" exception (34 C.F.R. § 99.31(a)(1)(i)(B)), requiring institutional control and contractual bans on commercial training or redisclosure. Review our rules on what not to paste into an AI tool at work and hold three lines:

• Never upload essays to free web checkers or personal chatbots: Unvetted sites lack a Data Privacy Agreement (DPA) and may retain student essays.
• Verify your district's Student Data Privacy Consortium (SDPC) list: Use only LMS-integrated tools covered by an active National Data Privacy Agreement (NDPA).
• Never ask ChatGPT if it wrote a paper: Chatbots cannot view other accounts' histories and routinely claim authorship of classic literature and human essays.

The ed.cafe 6-step educator fair-process inquiry and assignment-design rubric

Built from the court lessons in Harris v. Adams and Newby v. Adelphi, the ed.cafe 6-Step Educator Fair-Process Inquiry & Assignment-Design Rubric standardizes inquiry before any misconduct report is filed.

--- PART 1: 3-GATE PRE-SCAN CHECK (STOP IF ANY ANSWER IS "NO") ---
[ ] GATE 1 - ADVANCE WRITTEN RULE: Syllabus stated Green / Yellow / Red AI rules
    BEFORE students began drafting.
[ ] GATE 2 - FERPA-VETTED LMS TOOL ONLY: Institution-contracted tool with signed
    DPA (NEVER a free web checker or personal chatbot).
[ ] GATE 3 - ZERO SOLO-SCORE PENALTY: Score is an inquiry prompt only; zero
    grade cut or referral without non-detector corroboration.

--- PART 2: THE 6-STEP FAIR-PROCESS INQUIRY PROTOCOL ---
STEP 1 - CHECK TEXTUAL ANOMALIES FIRST (5 MINS): Ignore Turnitin *% (1%-19%).
  Check 3 citations/quotes against class texts for fake books, invented quotes,
  or un-taught acronyms; compare with 1-2 prior in-class samples.
STEP 2 - CHECK IEP / 504, ELL & EXTENSION CONTEXT: Check dictation, graphic
  organizers, campus tutoring, and school-pushed Grammarly/Editor rewriters.
STEP 3 - INSPECT CLOUD VERSION HISTORY: Review Google Docs or Word OneDrive
  edits (multi-session typing, deletions, rough notes, total editing minutes).
STEP 4 - SEND CALM 10-MINUTE CHECK-IN EMAIL:
  Subject: Quick 10-minute check-in on your [Assignment Name] draft
  Hi [Student Name], I am finishing feedback on [Assignment Name] and would like
  to meet for 10 minutes on [Option 1] or [Option 2] to walk through your draft
  and sources before posting grades. Please bring your live Google Doc or Word
  file (leave it unedited and un-saved beforehand so timestamps stay intact) and
  source PDFs. Nothing is finalized; this lets you walk through your work live.
STEP 5 - 3-QUESTION ORAL WALKTHROUGH (10 MINS; INVITE LIAISON IF IEP/504):
  (1) "Why did you choose [claim on p. X]?" (2) "Show me in [Source Y] where you
  found paragraph [Z]." (3) "Define [key term in essay] in your own words."
STEP 6 - MULTI-FACTOR OUTCOME: [ ] CLEAR (version log OR oral defense -> normal
  grade, zero conduct note)  [ ] GUIDED REWRITE (extension boundary confusion)
  [ ] FORMAL REFERRAL (2+ corroborators: fake cites + admitted paste or zero
  drafts + oral inability to explain thesis).

--- PART 3: COPY-READY SYLLABUS "TRAFFIC-LIGHT" AI CLAUSE ---
- GREEN (Allowed, No Note): Basic spellcheck and library searches.
- YELLOW (1-Line Footnote + Saved Chat Link): Pre-draft study questions when
  permitted on the prompt sheet.
- RED (Prohibited): ChatGPT, Claude, Gemini, Copilot, QuillBot, Apple Writing
  Tools, or Grammarly's "Rewrite" button to draft, rewrite, or cite-build prose.
- DRAFTING RULE: Draft start-to-finish inside one Google Doc or OneDrive Word
  file with Version History on. Detector scores never decide grades alone.

Three assignment-design shifts that cut Sunday-night detector policing in half

Pair the ed.cafe 6-Step Educator Fair-Process Inquiry & Assignment-Design Rubric with three upstream course tweaks that cut unpaid Sunday detective work:

1. Collect one live cloud link at outline and final submission: Have every student submit the same Google Doc or OneDrive link at Milestone 1 (thesis + three page-cited quotes) and Milestone 2 (final essay) so version logs are universal rather than punitive.
2. Require exact page citations to seminar handouts or local records: Chatbots summarize famous novels easily yet invent page numbers on closed readings. Run our 75-minute hallucination lab exercise before Paper 1 so students see phantom citations firsthand.
3. Shift 15% to 20% of essay weight to a five-minute in-class reflection: On due day, ask students on paper (or an accommodated keyboard) to summarize their main counterargument and one cut source in four sentences. Authentic writers finish in four minutes without a detector scan.

Where this stops being true

Four boundaries require stepping outside a classroom checklist:

• Union contracts and honor-board bylaws: Many K-12 contracts require an administrator or parent notice before questioning a minor on misconduct, while some college honor codes bar private faculty grade settlements outside the central integrity office.
• IDEA and Section 504 rules: If penalties accumulate toward suspension or placement changes for a K-12 student with an IEP or 504 plan, federal disability safeguards (20 U.S.C. § 1415(k)) govern.
• Proctored paper or locked-browser exams: Offline exam booklets have no cloud version logs and need no AI text detector.
• Attorney letters or threatened suits: Once a family copies counsel (Harris v. Adams; Newby v. Adelphi), route replies through your principal, dean, union rep, and school counsel.

Questions people ask

Key terms

• Base-rate trap: The probability rule under which even a 1% to 4% false-positive rate accuses dozens of innocent writers across thousands of authentic essays.
• Perplexity and burstiness: Measures of word predictability (perplexity) and sentence-length variation (burstiness) used by AI classifiers.
• Turnitin asterisk mask (*%): Turnitin's rule hiding 1% to 19% AI scores due to unreliable low-range false positives.
• School official exception (34 C.F.R. § 99.31): The FERPA rule allowing student record sharing only with contracted vendors barred from commercial data use.
• Traffic-light syllabus rule: Advance classroom rules separating Green (allowed), Yellow (footnote required), and Red (prohibited rewriting) AI tools.
• Corroborating textual anomaly: Objective non-detector proof such as fake cited books (Harris v. Adams) or un-taught acronyms (Yang v. Minnesota).

Sources

1. D. Mass., Harris v. Adams (No. 24-cv-12437-PGL, Hingham High School Order, Nov 20, 2024): Gizmodo Case Report.
2. GradPilot, "AI Cheating Lawsuits Tracker: Newby v. Adelphi, Harris v. Adams & Yang v. Minnesota" (Jul 10, 2026): GradPilot Tracker.
3. Plagiarism Today, "Adelphi University Sued Over AI Allegation (Orion Newby)" (Oct 14, 2025): Plagiarism Today.
4. Liang, Yuksekgonul, Mao, Wu, and Zou (Stanford), "GPT detectors are biased against non-native English writers" (Patterns, Jul 2023; n = 91 TOEFL essays, 2023 detector versions): arXiv:2304.02819.
5. Turnitin, "AI Writing Detection FAQ (Sole-Basis Warning & 1%-19% Asterisk Rule)" (current): Turnitin FAQ.
6. Turnitin Guides, "Using the AI Writing Report" (current): Turnitin Educator Guide.
7. OpenAI, "New AI classifier for indicating AI-written text (Discontinued Jul 20, 2023)" (Jul 20, 2023): OpenAI Notice.
8. MyChesCo, "FTC Finalizes Order Restricting Workado's AI Content Detection Claims" (Sep 2, 2025): FTC Enforcement Coverage.
9. Massachusetts DESE, "Artificial Intelligence in K-12 Schools Guidance" (Updated Aug 26, 2026): MA DESE Portal.
10. WBUR, "Massachusetts issues AI guidance for K-12 schools" (Sep 19, 2025): WBUR Report.
11. US Department of Education, "FERPA (20 U.S.C. § 1232g; 34 C.F.R. Part 99)" (current): StudentPrivacy.ed.gov.