Implementing a defensible academic integrity program: a complete institutional guide

The problem with detector-only programs

The current generation of academic integrity programs at most institutions follows a single-tool logic: route assessed submissions through an AI detector, and open proceedings when the score crosses a threshold. The problem is not that these detectors are useless — as triage instruments they have value — but that they have been promoted to a role they cannot perform. They are being used as primary evidence when they were only ever appropriate as screening signals.

The legal and regulatory record is now clear enough that this is not a judgment call. In February 2026, a court ruled against Adelphi University after it penalized a student based on an AI detection finding, finding that the institution had acted "without valid basis and devoid of reason." The court did not dispute that the detector scored high. It found that a detector score — an inference drawn by a probabilistic model — did not meet the evidentiary standard required to impose significant consequences on a student.

The UK's Office of the Independent Adjudicator (OIA) has said the same thing in regulatory guidance. Its casework note on AI and academic misconduct is direct: a detector score alone should not be treated as sufficient evidence of misconduct, and the institution bears the burden of proof. OIA casework also documents a recurring pattern of procedurally defensible findings being overturned on substantive grounds — the institution simply did not have enough evidence to support the conclusion it reached.

The statistical problem compounds the legal one. Liang et al. (arXiv:2304.02819) tested seven widely used AI detectors against authentic essays written by non-native English speakers and measured an average false positive rate of 61.3%. For native English speakers writing genuine essays, the same detectors produced near-zero false positive rates. This is a peer-reviewed finding about a systematic disparity in how these tools perform across a population most institutions serve in substantial numbers.

Vanderbilt University, when it disabled Turnitin's AI detector in 2023, made part of this arithmetic explicit in its published reasoning: even at the vendor's claimed 1% false positive rate on a general population, a large institution would wrongly flag hundreds of genuine submissions per year with no way to identify which ones. Turnitin's own published guidance on false positives acknowledges the limitation and cautions against using the score as a sole basis for adverse action.

The transition to a defensible program is not about adding more tools. It is about changing what kind of evidence the program produces before any dispute arises — moving from probabilistic inference about the final product to behavioral observation of how the work was created.

---

Program architecture: the three-tier model

A defensible integrity program operates across three tiers that function in sequence. Tier 1 reduces the likelihood of misconduct through clear policy and student education. Tier 2 creates prospective evidence for assessed work before any dispute arises. Tier 3 provides the structured process for when a concern does arise. Each tier depends on the others: good Tier 3 process cannot compensate for the absence of Tier 2 evidence; strong Tier 2 capture is weakened if the Tier 1 policy framework does not give students clear expectations.

%%{init: {'theme': 'base', 'themeVariables': {'primaryColor': '#e8f4ee', 'primaryTextColor': '#16130c', 'primaryBorderColor': '#1f5a3a', 'lineColor': '#6f5618', 'secondaryColor': '#fffdf8', 'background': '#fffdf8', 'mainBkg': '#fffdf8', 'clusterBkg': '#f0ede4', 'titleColor': '#16130c', 'edgeLabelBackground': '#fffdf8'}}}%%
flowchart TB
    subgraph T1["Tier 1 — Deterrence and Education"]
        P["Policy: clear syllabus language\nper-assignment AI rules"]
        SE["Student education: what records\nto keep, how process capture works"]
        ST["Staff training: detector limitations,\nevidentiary standards"]
    end
    subgraph T2["Tier 2 — Prospective Evidence"]
        CA["Process capture enabled for\nassessed writing"]
        CO["Consent obtained before\neach session begins"]
        SE2["Sealed, tamper-evident record\ncreated at time of writing"]
    end
    subgraph T3["Tier 3 — Investigation and Adjudication"]
        TR["Triage: concern raised,\ncorroborating evidence assessed"]
        GE["Evidence gathered:\nprocess records, prior work, metadata"]
        PA["Student conversation\nand panel if warranted"]
    end
    T1 -->|"Clear expectations\nreduce disputes"| T2
    T2 -->|"Prospective records\nfund investigations"| T3
    T3 -->|"Outcomes inform\npolicy calibration"| T1

The key insight is the direction of dependency. Tier 2 evidence is created before any concern arises — this is what makes it prospective. It cannot be retrofitted after a flag. An institution that has not deployed Tier 2 for a given assignment will conduct any Tier 3 investigation with a material evidence gap that a Tier 1 policy alone cannot fill.

---

Tier 1: Policy and education

What the research says about deterrence

The QAA's guidance on contract cheating is instructive: institutions with clear, specific policies produce fewer misconduct disputes than institutions with vague or inconsistent language. This is not because students read policies carefully; it is because clear policies give instructors a consistent enforcement reference, reduce the ambiguity that generates borderline cases, and establish the shared understanding that makes early conversations with students more productive.

OIA casework shows a corresponding pattern. Ambiguous policies — policies that say AI use "may not be appropriate" without specifying what is permitted, or that require "your own work" without defining how AI assistance interacts with that requirement — generate more formal disputes and more successful student challenges. When a student can plausibly argue they did not understand the rule they allegedly broke, the institution is on weaker ground.

Syllabus language

Every syllabus covering assessed writing should specify three things in plain terms: what AI tool use is and is not permitted for this assignment, what disclosure is required for any permitted use, and what evidence a student is expected to be able to produce if asked. The three-tier model from the syllabus AI policy language article — prohibited, permitted with disclosure, and encouraged within limits — provides the decision framework, but whichever tier an instructor chooses, the policy should name the evidence expectation. A student who knows in advance that a sealed process record is accepted as proof of genuine authorship is in a much better position than one who is told after a concern is raised that they should have kept records.

Student education

Students need three things before the first assessed submission. First, a clear explanation of what the AI policy for this course is and why it exists. Second, a practical explanation of what records they can keep on their own behalf — that sealed process records belong to the student, are verifiable by anyone, and can be produced in any dispute. Third, an honest account of how disputes are resolved: that the institution must prove misconduct, not the student disprove it, but that process evidence makes that resolution faster and fairer for everyone.

The goal is not to make students anxious about surveillance. It is to give honest students a clear path to being believed when they are wrongly flagged — which, as the Liang et al. data shows, happens at a rate that is not trivial.

Staff training

Every instructor who submits integrity concerns and every administrator who manages proceedings needs to understand two things that vendor marketing often obscures. First, AI detector scores are not evidence; they are triage triggers. A score alone is not sufficient to open formal proceedings. Second, the institution bears the burden of proof. The investigation's job is to find out what happened — not to confirm a suspicion. Staff who understand these two constraints will initiate fewer unfounded proceedings and handle the ones they do initiate more defensibly.

---

Tier 2: Prospective process capture

Prospective process capture means deploying a tool that records the writing session — keystrokes, paste events, timing, composition sequence — at the time the work is being produced, with student consent obtained before work begins. The record is sealed at session close, producing a tamper-evident log that can be examined by the institution, the student, and any review body, without requiring trust in the party that produced it.

Which assignments to capture

Not every assignment needs process capture. The cost-benefit calculation favors deployment on assessed writing that carries significant weight toward a grade, writing where the assessment is primarily about the student's own analytical process rather than a polished product, and writing in subject areas or at assessment points where AI generation concerns are most prevalent. Draft submissions and low-stakes tasks may not warrant the overhead. The criteria for including or excluding an assignment type should be set at institutional level and applied consistently — not decided assignment-by-assignment in ways that could appear selective in retrospect.

Consent design

Consent should be per-session and explicit. The student should be told, before each session begins, that the session is being recorded for academic integrity purposes, what data is being collected, who can access the record, how long it is retained, and that they may decline — with the consequence that the assignment must be completed outside the captured environment. The consent record itself should be sealed as part of the process record. As the FERPA and student writing data privacy article explains, a consent-gathered record is both legally stronger and epistemically stronger: a student who consented to and acknowledged the capture in real time has no credible later claim that the record was fabricated.

The consent screen should not be buried in a terms-of-service flow. It should be the first thing a student sees when beginning a captured session, and it should say — in plain language, not legalese — what is being recorded and why.

Data retention

The institution's data retention schedule for process records should match its appeal window. If a student can appeal an integrity finding up to two years after a decision, process records for assessed work should be retained for at least that period after the final decision is made (not after submission). Many institutions set a default retention period of two to three years for academic records. Process records should fall within that schedule, not outside it. Records retained beyond the appeal window create privacy obligations without evidentiary benefit. Records deleted before the appeal window closes create evidence gaps at exactly the moment they are most likely to be needed.

Privacy and regulatory obligations

For US institutions, writing process data collected through an institution-deployed tool is almost certainly an education record under FERPA and carries the corresponding obligations: student inspection rights, disclosure restrictions, and data minimization obligations. For UK and EU institutions, GDPR applies, with additional obligations around lawful basis, erasure rights, and restrictions on secondary use. The full analysis is in the FERPA and student writing data privacy article. The practical upshot: the vendor contract must give the institution the ability to produce student records on request, delete them on request, and prohibit the vendor from using student data for any purpose other than providing the contracted service.

The capture-to-seal workflow

%%{init: {'theme': 'base', 'themeVariables': {'primaryColor': '#e8f4ee', 'primaryTextColor': '#16130c', 'primaryBorderColor': '#1f5a3a', 'lineColor': '#6f5618', 'secondaryColor': '#fffdf8', 'background': '#fffdf8', 'mainBkg': '#fffdf8', 'clusterBkg': '#f0ede4', 'titleColor': '#16130c', 'edgeLabelBackground': '#fffdf8'}}}%%
flowchart LR
    A["Student opens\nassignment"] --> B["Consent screen\ndisplayed"]
    B -->|"Decline"| C["Assignment completed\noutside capture\nenvironment"]
    B -->|"Accept"| D["Session begins\nCapture active"]
    D --> E["Keystroke, paste\nand timing data\nrecorded in real time"]
    E --> F["Session ends\n(submission or close)"]
    F --> G["Record sealed\nhash-chained,\ntamper-evident"]
    G --> H["Student receives\naccess to their\nown record"]
    G --> I["Institution holds\nsealed record in\nappeal-window retention"]

Authmarks implements this workflow with per-session consent, cryptographic sealing at session close, and student access to their own record. Both the institution and the student receive access to the same sealed record — removing the asymmetry that makes process records legally vulnerable.

---

Tier 3: Investigation process

When a concern arises, the investigation exists to find out what happened — not to confirm a suspicion. The six-step framework from the investigating AI-writing concerns article maps onto the following flow. The full investigation flowchart is:

%%{init: {'theme': 'base', 'themeVariables': {'primaryColor': '#e8f4ee', 'primaryTextColor': '#16130c', 'primaryBorderColor': '#1f5a3a', 'lineColor': '#6f5618', 'secondaryColor': '#fffdf8', 'background': '#fffdf8', 'mainBkg': '#fffdf8', 'clusterBkg': '#f0ede4', 'titleColor': '#16130c', 'edgeLabelBackground': '#fffdf8'}}}%%
flowchart TD
    A["Concern raised\n(detector flag, instructor observation,\npeer report, anomaly)"] --> B{"Triage:\ncorroborating evidence\nindependent of detector?"}
    B -->|"No"| C["Document and close\nRecord: concern, triage\nreasoning, no further action"]
    B -->|"Yes — proceed"| D["Gather evidence before\ncontacting student\n(process records, prior work,\nsubmission metadata, timeline)"]
    D --> E["Population assessment\n(non-native speaker? short text?\nformal register writer?)"]
    E --> F["Student conversation\n(information gathering —\nnot a disciplinary hearing)"]
    F --> G["Student may provide\nown evidence\n(process record, drafts,\nresearch trail)"]
    G --> H{"Assess all evidence\nAgainst burden of proof:\ncan the institution articulate\nwhat supports the finding?"}
    H -->|"Yes — evidence is sufficient"| I["Formal panel\nwith documented reasoning"]
    H -->|"No — evidence insufficient"| J["Close with record\nNote: concern raised,\nevidence insufficient\nto proceed"]
    I --> K["Outcome recorded\nwith full reasoning\n(not just conclusion)"]

The triage step is the most important gate in the process. A concern that cannot produce corroborating evidence independent of the detector should not proceed to a formal investigation. This is not a low bar to protect students at the expense of institutional authority — it is the minimum standard required by the OIA and borne out by the Adelphi ruling. Opening formal investigations on score-only triggers produces the procedural record that leads to successful appeals.

At the population assessment step, the investigator should explicitly consider whether the student under review belongs to a group for which the detector is known to over-flag. Non-native English speakers, students who write in dense academic register, and submissions under roughly 300 words all produce elevated false positive rates on current detectors. If the student is in one of these groups, the evidentiary threshold for proceeding should be correspondingly higher. This is not a separate policy rule — it follows directly from correctly applying the burden of proof.

---

Evidence documentation standards

The panel file is what survives the investigation. In a successful appeal, the weakest link is almost always the documentation — not because the underlying finding was wrong, but because the reasoning that produced it was not written down. The following contents are necessary for a panel file that can withstand formal challenge:

1. The initial concern and trigger. What was observed, when, by whom. If a detection tool was run: which tool, which version, the score produced, and the caveats the vendor itself attaches to that score. Do not summarize the score as a finding — record it as a trigger. Date-stamp everything.

2. The triage decision. What corroborating evidence was identified at triage. If the decision was to proceed, document why the corroborating evidence was considered sufficient to warrant a formal investigation. If the decision was to close, document that reasoning too.

3. Evidence gathered, with provenance. Each piece of evidence should record its source, when it was collected, what tool or method was used to obtain it, and any material limitations. For process records: the tool and version that produced them, whether they are sealed and hash-verified, and who has access. For prior work comparisons: which submissions were compared and by what method. For metadata: what system produced the metadata and what its chain of custody is.

4. Population assessment. A brief note on whether the student belongs to a population with elevated false positive rates, and what effect, if any, this had on the evidentiary threshold applied.

5. The student's response and any evidence they provided. The student's account of how the work was produced. Any evidence the student submitted — drafts, research notes, their own process record, library access records. Each piece of student-provided evidence should be assessed, not dismissed. If the panel did not find student evidence persuasive, the file should say why.

6. The panel's reasoning. Not just its conclusion. The panel should identify which evidence it relied on, what weight it gave conflicting evidence, and how it applied the institution's burden of proof. A file that records "the panel finds AI was used" without this reasoning will not survive a formal challenge and signals that the panel may not have examined the evidence rigorously.

7. Outcome and penalty. The finding, the penalty applied, the appeal rights communicated to the student, and confirmation that procedural timelines were met.

---

Common failure modes

The following table covers the most frequently recurring failures in AI-writing integrity proceedings, drawn from OIA casework patterns and the procedural findings in cases that have been successfully appealed or litigated.

Failure mode	Why it leads to failed appeals	Mitigation
Treating the detector score as a finding	The OIA and the Adelphi court both found this insufficient. The score is probabilistic inference; a finding requires evidence of behavior.	Use scores only for triage. Document the corroborating evidence that actually justifies proceeding.
Not disclosing which tool was used	A student cannot challenge evidence they do not know exists. Concealing the tool or its version is a procedural failure that is independent of the merits.	Disclose the tool name, version, score, and the vendor's own caveats in the formal notification.
Running multiple detectors until one flags	This is p-hacking — not investigation. It inflates the apparent confidence of the finding and does not produce independent corroboration.	Run one detector as a triage tool and treat its score as a single data point, not as a finding.
Absence of version history treated as evidence	Not all students use version-history-enabled tools. Absence of a record the student was never told to keep is not affirmative evidence of fraud.	Document the absence as an evidence gap — it applies equally to both sides — not as a finding.
Not giving adequate response time	Procedural timelines are enforceable as institutional commitments. A student who cannot respond meaningfully before the decision is made has a due-process argument independent of the merits.	Audit your notification-to-response windows against your published policy. Backlogs are a common source of this failure.
Not documenting the panel's reasoning	A decision that records only its conclusion cannot be reviewed on appeal. Review bodies require the reasoning, not just the outcome.	Train panel members to document which evidence they relied on and why they weighted it as they did.
Failing to assess student counter-evidence	If the student provides a process record, drafts, or a research trail and the panel dismisses it without documented reasoning, the appeal will focus on that dismissal.	Each piece of student-provided evidence should appear in the panel file with a brief assessment of what weight was given it and why.

---

Measuring program health

A well-designed program should produce measurable outcomes. The following metrics track program health at the institutional level and flag when processes are not working as intended.

Concerns raised per semester. Track the volume of concerns at the triage stage — the point where an instructor or administrator decides something warrants a formal look. A rising trend without a corresponding rise in formal findings may indicate that instructors are uncertain about the triage threshold. A falling trend may indicate under-reporting or detector fatigue.

Rate of concerns closed at triage. The proportion of raised concerns that do not proceed to formal investigation because corroborating evidence was insufficient. This rate should not be zero — it reflects the triage gate working as designed. A rate near zero means that either very few concerns are raised (selective reporting), or that concerns are routinely proceeding without adequate corroboration (the triage gate is not functioning).

Rate of findings upheld on appeal. This is the most important program health metric. A high upheld rate means the program is producing defensible findings — findings that rest on sufficient evidence and followed correct process. A high rate of successful appeals — where the student wins on appeal — means the program is not producing defensible findings. That is not primarily a fairness problem (though it is that); it is a signal that the evidence and process standards being applied at first instance are not sufficient to survive review.

Time from concern raised to resolution. Long delays are a source of procedural vulnerability and are unfair to both students and staff. Track time from initial concern to triage decision, from triage to formal notification, and from notification to outcome. Publish internal targets and audit compliance against them.

Cases resolved by process record. Track how many cases were resolved — in either direction — by the existence of a sealed process record. Cases where a process record exonerated a student who was wrongly flagged are as important to count as cases where a process record supported a finding of misconduct. Both demonstrate that the Tier 2 investment is producing value.

---

Rollout sequence

Implementing this program requires sequencing. Deploying Tier 2 capture without Tier 1 policy clarity means students encounter process recording without understanding why it exists. Deploying Tier 3 process changes without Tier 2 capture means investigations continue to rest on the same thin evidence base. The sequence below is a practical 16-week path from current state to a full three-tier program, with a calibration phase after the first complete semester.

%%{init: {'theme': 'base', 'themeVariables': {'primaryColor': '#e8f4ee', 'primaryTextColor': '#16130c', 'primaryBorderColor': '#1f5a3a', 'lineColor': '#6f5618', 'secondaryColor': '#fffdf8', 'background': '#fffdf8', 'mainBkg': '#fffdf8', 'clusterBkg': '#f0ede4', 'titleColor': '#16130c', 'edgeLabelBackground': '#fffdf8'}}}%%
gantt
    title Integrity Program Implementation Sequence
    dateFormat  YYYY-MM-DD
    axisFormat  Week %W

    section Phase 1 — Policy and Training
    Review and update AI policy documents        :p1a, 2026-01-05, 7d
    Update syllabus template language            :p1b, after p1a, 7d
    Staff training — detector limitations        :p1c, after p1a, 7d
    Staff training — investigation process       :p1d, after p1b, 7d

    section Phase 2 — Pilot Capture
    Select volunteer cohort and assignments      :p2a, 2026-02-02, 7d
    Deploy capture tool, test consent flow       :p2b, after p2a, 7d
    Pilot runs — process records collected       :p2c, after p2b, 14d
    Review pilot findings with instructors       :p2d, after p2c, 7d

    section Phase 3 — Full Deployment
    Consent architecture reviewed with DPO       :p3a, 2026-03-16, 7d
    Retention schedule confirmed                 :p3b, after p3a, 7d
    Full deployment for assessed writing         :p3c, after p3b, 14d
    Student orientation updated                  :p3d, after p3b, 7d

    section Phase 4 — Review and Calibration
    First-semester metrics review                :p4a, 2026-05-18, 7d
    Appeals analysis and findings audit          :p4b, after p4a, 7d
    Policy and process updates                   :p4c, after p4b, 7d

Phase 1 (weeks 1–4): Policy update and staff training. Audit existing AI policy documents for gaps and internal inconsistencies. Update the institutional syllabus template and any faculty guidance to reflect the three-tier evidence expectation. Conduct staff training on detector limitations and the six-step investigation framework. This phase costs nothing beyond staff time and reduces immediate exposure.

Phase 2 (weeks 5–12): Pilot process capture with a volunteer cohort. Select a small group of instructors who are willing to deploy process capture on one or two assessed assignments. The pilot should include the full consent flow, test that records are accessible to both staff and students, and confirm that retention and deletion processes work as intended. Review findings with participating instructors before full deployment — expect to surface practical friction that is better resolved in a controlled pilot than at scale.

Phase 3 (semester 2): Full deployment with consent architecture. Before scaling, have the institution's data protection officer or legal counsel sign off on the consent language, the retention schedule, and the vendor contract against the FERPA/GDPR checklist from the privacy article. Then deploy for all assessed writing in scope. Update student orientation materials to explain the program before they encounter it.

Phase 4 (after first full semester): Review and calibration. Pull the program health metrics described above. Audit a sample of completed cases against the documentation standards. If appeal rates are elevated, identify where in the process the documentation is weakest and fix it before the next semester. This review cycle is what converts a policy document into a functioning program.

---

The honest caveat

This program does not eliminate misconduct. It changes what the institution can prove. That shift matters because it runs in both directions: a defensible program correctly finds against students who used AI tools, and it correctly exonerates students who did not. The exonerating use case is more common — most students who are flagged by a detector did not use AI, and the harm from a wrongful finding to a student's academic record can be severe and lasting.

An integrity program that produces findings upheld on appeal at a high rate is doing both jobs. It is catching misconduct accurately enough that the evidence sustains review. It is also — necessarily — exonerating genuine students at triage, because those cases do not proceed. That combination is what a defensible program looks like. The detector-only model produces neither outcome reliably.

---

Tier 1 deep dive: building assignment types that reduce AI misconduct risk

The three-tier model places policy and education in Tier 1, but this tier contains more than documentation. It includes the upstream decisions about what kinds of assessments the institution asks students to produce. Assignment design is among the most effective — and most underused — levers in academic integrity programs, because it operates before any misconduct occurs and reduces the conditions that make misconduct attractive.

What the QAA research shows about assignment type and risk

The QAA's research on contract cheating produced a finding that has direct implications for AI-assisted misconduct: assignment types strongly predict the prevalence of attempts to outsource the work. Anonymous take-home essays — where the student submits a document with no observed process, no in-person component, and no personalizing constraint — are consistently the highest-risk format. In-class timed assessments under supervision are the lowest risk, but they carry significant pedagogical trade-offs: they do not allow deep analytical work, disadvantage students with test anxiety or time-management challenges, and are impractical for many types of writing assessment.

The practical implication is not to replace all essays with in-class tests. It is to understand the risk spectrum and apply design mitigations to assignments that occupy the high-risk region. Take-home assessed writing remains a legitimate and often the most appropriate format for many learning outcomes — but it benefits from design features that make it harder to outsource effectively.

Assignment redesign taxonomy

Assignment formats can be mapped onto two axes: how much of the writing process is observable by the institution, and how much of the content requires personalized knowledge that cannot be fabricated without authentic experience or engagement.

High-surveillance, lower personalization assessments include standard timed in-class essays. Risk is low because the process is observed, but the trade-off is limited depth and significant accessibility concerns.

Low-surveillance, lower personalization assessments are the canonical risk case: anonymous take-home essays on general topics with no personal dimension. A student submitting on "the ethical implications of artificial intelligence in healthcare" has a format that is maximally amenable to AI generation.

Low-surveillance, higher personalization assessments include personal reflections, situated analyses tied to specific experiences, and iterative draft portfolios. The format is not supervised, but the content requires authentic engagement that an AI system cannot fabricate — because the experience being described belongs uniquely to the student.

Process-visible, any personalization level assessments are those where process capture is deployed: the writing process itself is recorded, regardless of where or when the student writes. This format preserves the full pedagogical flexibility of take-home assessment while creating a prospective behavioral record.

Specific assignment design tactics

Several design choices significantly reduce AI misconduct risk without requiring in-person supervision.

Personal experience requirement. Requiring students to ground their analysis in specific personal or professional experiences they can name and describe creates a content constraint that AI systems cannot fulfill convincingly. An essay analyzing a supply chain disruption that the student directly worked on, or reflecting on how a historical event connects to their own family's experience, requires the student to supply information no model has access to. Instructors reviewing for AI use should note that implausible personal narratives — generic, underdeveloped "experiences" that read as invented rather than recalled — may themselves be a corroborating signal worth noting at triage.

Scaffolded process with checkpoints. Breaking assessed writing into stages — initial annotation, thesis draft, evidence outline, first complete draft, final submission, each due at separate points — forces iterative engagement with the work over time. The progression of drafts, when captured and timestamped, is itself a form of prospective evidence. A student who produces a mature final submission but cannot produce an annotated bibliography or thesis draft that predates it by the required interval has a gap in their documented process.

Oral defense component. An end-of-submission conversation with the student about the content they submitted — five to ten minutes, focused on the reasoning behind specific choices in the text — tests comprehension in a way that no static detection tool can approximate. A student who can discuss the work intelligently has demonstrated engagement with it. A student who cannot explain their own argument has surfaced a concern that can be properly investigated. This approach does not require elaborate examination procedure; a brief post-submission discussion structured around two or three specific questions about the paper achieves the purpose.

Unusual constraint. Assigning a topic or format constraint that is specific enough to be rare in training data — write a policy brief responding to a specific local regulation published this semester, or analyze a case using a framework drawn from the specific readings assigned in this module — raises the difficulty of generating plausible AI output and simultaneously creates an implicit test of whether the student engaged with the course material. Generic prompts invite generic output.

The limitation of assignment redesign

Assignment redesign addresses the assessment condition but not the underlying pressure that drives misconduct. Research on academic integrity consistently identifies time pressure, fear of failure, and unclear task expectations as the primary reported drivers of contract cheating and AI-assisted work. Students who are under conditions of extreme load — multiple high-stakes deadlines converging, personal or financial crises, or genuine comprehension gaps in the subject matter — are more likely to use any available shortcut regardless of the assignment format.

Programs that add assignment redesign without also examining support structures are addressing the symptom. The most effective deterrence combines assignment design that is inherently harder to outsource with genuine academic support — accessible writing consultants, reasonable workload distribution, clearly communicated expectations for each assessment, and a culture in which students who are struggling can ask for help before they are in the position of submitting work they did not produce.

Communicating assignment design rationale to students

How the design rationale is communicated matters as much as the design itself. Instructors who frame redesigned assignments as "we changed this to catch cheaters" create an adversarial starting position that is pedagogically counterproductive. The same assignment can be framed honestly as "designed to demonstrate your specific thinking and your engagement with this material" — which is accurate, not euphemistic, and tells honest students what they will demonstrate by doing the work well.

When students understand that personal experience requirements and oral defense components exist to give them a stronger evidentiary basis for their own authorship — not only to prevent fraud — the framing shifts from surveillance to protection. An honest student in a well-designed assignment benefits from features that make their genuine work harder to dismiss.

Assignment design decision tree

%%{init: {'theme': 'base', 'themeVariables': {'primaryColor': '#e8f4ee', 'primaryTextColor': '#16130c', 'primaryBorderColor': '#1f5a3a', 'lineColor': '#6f5618', 'secondaryColor': '#fffdf8', 'background': '#fffdf8', 'mainBkg': '#fffdf8', 'clusterBkg': '#f0ede4', 'titleColor': '#16130c', 'edgeLabelBackground': '#fffdf8'}}}%%
flowchart TD
    A["New assessed\nwriting assignment"] --> B{"Carries significant\ngrade weight?"}
    B -->|"No"| C["Standard assignment\nNo special design\nrequired"]
    B -->|"Yes"| D{"In-class supervised\nformat feasible?"}
    D -->|"Yes — feasible\nand pedagogically\nappropriate"| E["In-class format\nLowest risk\nNo capture needed"]
    D -->|"No — take-home\nformat required"| F{"Topic allows\npersonal experience\nor situated constraint?"}
    F -->|"Yes"| G["Add personal\nexperience requirement\nor unusual constraint"]
    F -->|"No — general\ntopic required"| H["Add scaffolded\nprocess checkpoints"]
    G --> I{"High-stakes\nassessment?"}
    H --> I
    I -->|"Yes"| J["Deploy process\ncapture for session"]
    I -->|"No"| K["Consider oral\ndefense component"]
    J --> L["Add oral defense\ncomponent if feasible"]
    K --> L
    L --> M["Document design\nrationale in\ncourse materials"]

Assignment type risk table

Assignment type	AI misconduct risk	Mitigation options	Key trade-off
Anonymous take-home essay, general topic	High	Process capture; add personal experience requirement; unusual constraint	None with capture; personalization may limit some topics
Timed in-class essay	Low	Supervision is inherent mitigation	Depth and accessibility limitations
Personal reflection or situated analysis	Medium-low	Scaffold checkpoints; oral defense	Authenticity verification requires reading engagement
Iterative draft portfolio with checkpoints	Medium	Process capture on key drafts; timestamp verification	Workload for instructor feedback at each stage
Project report with specific local case	Medium-low	Unusual constraint limits model output quality	Topic specificity may limit student choice
Literature review, general topic	High	Process capture; oral defense; scaffolded annotation step	Depth requires time; oral defense overhead

---

Communicating the program to different stakeholder groups

A three-tier integrity program will fail in implementation if the communication strategy treats all stakeholders as a single audience. The arguments that will persuade an administrator differ materially from those that will persuade a student. The framing that is effective for legal and compliance teams is not the framing that creates faculty buy-in. Each group has specific concerns, specific risks they are trying to manage, and specific vocabulary through which they receive information about institutional programs. A communication plan that ignores these differences will encounter resistance that could have been anticipated and addressed.

Students: protection, not surveillance

The core message for students is honest and accurate: this program exists to protect genuine work from unfair challenge. If a student wrote their essay themselves, the process record proves it — quickly, verifiably, without requiring the student to argue against a probabilistic score they cannot interrogate.

Students are already aware that AI detectors exist and that their work may be run through them. Many are anxious about being wrongly flagged, particularly students who write in a formal academic register, who are non-native English speakers, or who have been told by instructors that their writing "sounds like AI." The message that resonates with this population is not "we are watching you" but "we are building the infrastructure to make sure that if you are ever wrongly accused, the accusation does not stand."

Students should also be told, explicitly, that sealed process records belong to them and can be produced by them in any dispute. The asymmetry of detector-based programs — where the institution holds probabilistic evidence and the student has no corresponding record — is something this program explicitly corrects.

Faculty: role clarity and reduced burden

The message for faculty is one of role clarification. Under a detector-only program, faculty are implicitly being asked to make both the triage judgment (this looks suspicious) and the substantive finding (this student cheated). Those are different roles, and conflating them creates legal exposure for the institution and an uncomfortable burden for instructors who are not trained as adjudicators.

Under the three-tier model, faculty do one job: raise a concern if they observe one. The formal investigation process exists precisely so that faculty are not put in the position of both instructor and judge. Faculty training should focus on four things: how to raise a concern through the triage process; what the triage gate checks before a formal investigation opens; what documentation to preserve about the concern itself; and what they should not do — specifically, do not communicate a finding to the student before the formal process has run.

Faculty buy-in is typically higher after the first case in which a formal process, rather than an informal instructor judgment, resolves a concern cleanly and defensibly. The program reduces the risk that an instructor's legitimate concern becomes their personal legal exposure.

Administrators: risk management and institutional liability

For academic administrators, deans, and provosts, the argument is one of risk management arithmetic. The cost of a wrongful finding — staff time on appeals, legal costs if the matter escalates to external challenge or litigation, reputational damage from public cases, and the ongoing resource drain of a high appeal volume — typically exceeds the per-student cost of process capture infrastructure by a significant margin.

The OIA casework record and the Adelphi ruling are the concrete evidence for this argument. These are not hypothetical risks; they are documented outcomes of programs that deployed detector-only models and found that the evidence base could not sustain challenge. The question for administrators is not whether to invest in defensible evidence infrastructure — it is whether to invest before or after a high-profile failure.

Legal and compliance teams: standards alignment

Legal and compliance teams need the program framed in standards terms. The program is designed to produce records that meet FRE 901 (US) or equivalent Evidence Act authentication requirements. The consent framework is designed to comply with FERPA (US) and GDPR (UK/EU) obligations including student inspection rights, deletion rights, and restrictions on secondary use. The process documentation standards are designed to enable the institution to demonstrate procedural fairness to any review body — OIA, accreditor, or court.

The specific technical claims that are worth making to legal teams: that the process record is cryptographically sealed at session close (meaning it cannot be retroactively altered by the institution, the student, or the vendor), that the student received the same sealed record as the institution (removing the asymmetry that generates authenticity challenges), and that the consent obtained at session start is itself part of the sealed record.

Accreditors: academic quality and QA frameworks

For accreditation bodies, academic integrity is an academic quality concern. An institution that produces a high volume of integrity findings that are then overturned on appeal is demonstrating a quality gap in its processes — not in its students' behavior. The institution is finding things it cannot prove.

The framing for accreditors is: this program represents a quality assurance upgrade to the institution's integrity processes. It replaces probabilistic triage tools promoted to the role of evidence with a prospective behavioral record that meets evidentiary standards. The outputs — appeal upheld rates, time to resolution, cases resolved by process record — are measurable quality indicators that the institution can track and report.

Communications planning table

Stakeholder	Key message	Key concern to address	Primary channel
Students	Process records protect your honest work	Fear of wrongful flagging; surveillance anxiety	Orientation, syllabus, first class session
Faculty	Raise the concern; the process does the rest	Burden of being both instructor and adjudicator	Staff training sessions; faculty handbook
Administrators	One successful appeal costs more than a year of capture infrastructure	Institutional liability; resource allocation	Executive briefing; risk register
Legal/compliance	Program meets FRE 901 / FERPA / GDPR standards	Authentication requirements; data obligations	Technical briefing with documentation
Accreditors	Program is a QA upgrade, not just an enforcement tool	Appeal rates as quality indicator	Annual quality report; self-assessment

---

Managing the vendor relationship

The vendor relationship for a process capture tool is a data relationship as much as a service relationship. The institution is not simply purchasing software; it is delegating the collection and custody of student education records to a third party. The contract terms that govern that relationship determine whether the institution can fulfill its legal obligations to students and whether the records produced by the tool have the evidentiary properties the program depends on.

Contract requirements

The institution's contract with any process capture vendor should contain, at minimum, the following provisions.

Data use restriction. Student data collected through the tool may be used only for the purpose of providing the contracted service to the institution. The vendor may not use student writing, behavioral data, or process records for any secondary purpose — including but not limited to training machine learning models, benchmarking detection tools, or conducting research — without explicit separate consent from both the institution and the affected students.

Student record production. The institution must retain the right to produce student records in response to student inspection requests under FERPA, subject access requests under GDPR, or orders from any regulatory or judicial body. The vendor must be able to fulfill production requests within a timeframe that meets the institution's regulatory obligations (typically 30 days under GDPR; 45 days under FERPA).

Deletion rights. The institution must retain the right to delete student records on request — either as part of a GDPR erasure right exercise or as part of the institution's own data retention schedule. The vendor must confirm deletion and must not retain copies in any backup or training dataset.

Portable export. On contract termination or institution request, the vendor must provide all student records in a portable, machine-readable format that the institution can store independently. Lock-in to a vendor-specific format is a chain-of-custody problem: if the vendor ceases to operate, the institution must be able to access records it is legally obligated to retain.

Service level agreement with evidence of uptime. The SLA should include not just uptime commitments but evidence of uptime — audit logs or independent monitoring that demonstrate the system was available at the time records claim to have been created. Chain-of-custody documentation requires a reliable timestamp infrastructure.

FERPA School Official exception

The most common lawful basis for US institutions deploying third-party tools under FERPA is the School Official exception: a vendor may access education records without student consent if the institution has determined that the vendor has a legitimate educational interest and if the vendor is under the direct control of the institution with respect to the use and maintenance of those records.

The direct-control requirement is not met simply by including contract language that says the vendor is under direct control. The institution should review whether the vendor's actual terms and operational practices constrain the vendor from using student records for any purpose other than the contracted service. If the vendor's terms of service or privacy policy reserve rights to use data for model training or research, the School Official exception may not be available, and the institution has a compliance gap it needs to address before deployment.

The verification independence test

The single most important technical question to ask any process capture vendor is whether the sealed record can be verified without the vendor's involvement. This question distinguishes between two architecturally different approaches.

In the first approach, verification runs through the vendor: the institution submits a record to the vendor's platform, the vendor's systems confirm it is authentic, and the institution relies on that confirmation. This means the chain of trust runs through the vendor. If the vendor has the technical capacity to modify records, or if the vendor ceases to operate, the verification chain is broken.

In the second approach, verification runs through a cryptographic public standard: the record contains a hash computed from the record content, the hash is verifiable against the record using public cryptographic functions, and any party can independently confirm that the record content matches the hash without communicating with the vendor. This is the architecture that produces a genuinely tamper-evident record — one where tampering would change the hash and invalidate verification, and where the vendor cannot unilaterally change the record without detection.

Institutions should ask vendors to demonstrate the verification step without any access to the vendor's platform. If that demonstration is not possible, the record is not independently verifiable, and its chain-of-custody properties are weaker than a true cryptographic seal.

Vendor evaluation table

Capability	Questions to ask	Red flags	What a good answer looks like
Data ownership	Who controls the sealed records — institution or vendor? Can the institution access records if the vendor relationship ends?	Vendor retains exclusive custody; no export path on termination	Institution-controlled storage; portable export on request
Verification independence	Can verification be performed without accessing the vendor's platform? Can a third party verify a record independently?	Verification requires vendor login or API access	Public hash verification; cryptographic seal verifiable offline
Deletion rights	Can the institution delete records on student GDPR request? Does the vendor confirm deletion across all storage?	Vendor retains backups; deletion is partial	Written deletion confirmation; vendor certifies removal from all systems
Export format	Can all records be exported in a portable format on contract termination?	Proprietary format; no export available without vendor tools	Standard format (JSON, PDF); exportable without vendor software
Training data use	Does the vendor use student writing or process data to train models?	Terms of service reserve training rights	Explicit prohibition on training use; auditable

---

International student considerations

The population-specific false positive problem

Non-native English speakers represent a significant — and at most research universities, a majority of graduate student — population. The Liang et al. finding that AI detectors produce a 61.3% average false positive rate for non-native writers is not a footnote for this population: it means that nearly two in three authentic essays written by students in this group will be flagged as AI-generated by current detection tools. This is not a modeling edge case; it is the expected behavior of tools calibrated primarily on native-English training corpora.

Institutions that are aware of this disparity and continue to use detector scores as primary evidence — without explicit population-adjusted thresholds — are applying a triage instrument to a population for which it is known to systematically fail. This creates both a fairness problem and a legal exposure: a student who can demonstrate that they belong to a population with a documented 61.3% false positive rate, and that the institution used a score from one of those tools as its primary evidence, has a strong procedural challenge regardless of the substantive facts.

Immigration status and due-process stakes

For international students on student visas, an academic misconduct finding carries consequences that extend materially beyond academic penalty. Depending on the jurisdiction and the institution's reporting obligations, a serious academic misconduct finding may trigger immigration reporting requirements, may affect visa renewal, or may constitute grounds for exclusion from the institution with resulting immigration consequences. These are not consequences that a domestic student typically faces.

The materially higher stakes for this population are relevant to how the institution applies its burden of proof. This does not mean maintaining a separate, lower evidentiary standard for domestic students — it means applying the stated standard rigorously, because for international students the cost of an erroneous finding is disproportionately severe.

Language and procedural fairness

A student who does not fully understand the formal proceedings they are party to has a legitimate procedural fairness challenge. This is relevant to international students in two ways. First, the formal notification of allegation must be comprehensible to the student receiving it: legal and procedural language that is impenetrable to a student who is not a native English speaker is not adequate notice. Second, students should be able to participate in proceedings in a language they understand, either through proceedings conducted in their language where that is feasible, or through access to interpretation.

An international student who received a formal notice they did not fully understand, and who participated in a hearing without adequate interpretation, has a procedural fairness challenge that is independent of the merits of the underlying concern. Institutions should audit their notice templates and hearing procedures against this standard.

Multilingual assignments and process capture

An institution with large international student populations may have students completing assessed writing in languages other than English. The question of whether process capture works across languages is worth examining. The behavioral composition signature that prospective capture relies on — the distribution of pause events, keystroke rhythms, and editing patterns — is grounded in cognitive processing at linguistic boundaries, not in English-specific patterns. The underlying theoretical basis of this approach (drawing on Crossley et al.'s work on writing process features) is that the composition rhythm reflects cognitive effort at the level of idea formation, which is a cross-linguistic feature.

However, calibration for specific non-English language populations is less well-studied, and institutions deploying capture for non-English assessed writing should note this as a limitation in any proceeding that relies on the behavioral record. The tamper-evidence property of a sealed record — that the record has not been altered since it was created — is language-independent; the behavioral interpretation of what the record shows about the composition process is where calibration questions arise.

Applying a higher corroborating threshold

The practical implication for institutions with large international student populations is not to exempt this group from integrity processes. It is to apply a higher corroborating evidence threshold before proceeding formally with a case that is triggered by a detector score alone. This follows directly from correctly applying the burden of proof: if the tool used for triage is known to have a 61.3% false positive rate for the population the student belongs to, then the triage stage should require stronger corroboration before a formal investigation opens.

This is not a preferential standard — it is the logical consequence of using evidentiary tools whose population-specific error rates are documented. An institution that fails to account for known instrument bias in its triage threshold is not applying the burden of proof rigorously; it is applying the tool mechanically, which is precisely the failure mode the OIA's casework guidance identifies.

---

Program evolution: what the second and third year look like

A successfully implemented integrity program does not look the same in year three as it did in year one. The initial deployment is necessarily cautious: piloting capture, training staff on new investigation standards, and collecting the first round of data. The program that exists after two to three years of operation should reflect everything learned from that data — including which assignment types are generating the most concerns, which triage decisions are holding up on review, and whether the cultural signals suggest that students and faculty understand what the program is for.

Year 1: foundation and first data

Year 1 is about getting the infrastructure in place and collecting baseline data. The pilot capture cohort provides the first real experience of how consent flow, record access, and retention schedules work in practice. Staff training on the investigation framework produces the first generation of investigators who understand the triage gate and the documentation requirements.

The metrics collected in year 1 provide the baseline against which progress in years 2 and 3 is measured. What proportion of concerns are closed at triage? How long does the average case take from initial concern to resolution? How many cases are resolved — in either direction — by a process record? A program that does not collect these numbers in year 1 cannot demonstrate improvement later.

Year 2: calibration and assignment-level redesign

Year 2 is where the data from year 1 becomes actionable. The appeal upheld rate from year 1 identifies where the program's evidentiary foundation is weakest. If a significant proportion of formal findings are being challenged successfully, the year 2 response is to identify whether the failure is in triage (concerns proceeding without adequate corroboration), documentation (panels not recording their reasoning), or assignment design (some assignment types are generating a disproportionate share of difficult cases).

Assignment redesign for the highest-risk formats — typically anonymous take-home essays on general topics — can be sequenced in year 2 based on what year 1 data shows. Process capture coverage can be expanded based on pilot experience. Staff who participated in the first cycle of cases now have firsthand knowledge of what the documentation standard requires, and training in year 2 can be more targeted than the initial general training.

Year 2 is also typically when the first cases resolved by prospective process records appear — both exonerations (where a process record shows a flagged student composed their work genuinely) and findings (where the absence of a plausible process record corroborates other evidence of concern). Both types of cases are valuable: they demonstrate to faculty and students that the Tier 2 investment is not theoretical.

Year 3: the cultural shift

The leading indicator of a genuine cultural shift in a functioning integrity program is what students do voluntarily. In a program that is working as intended, students who are flagged in year 3 will often produce their own sealed process records unprompted — because they know from experience, or from their peers' experience, that producing such a record resolves the concern faster than any other available response. When students are proactively maintaining their own process records before any concern is raised, the program has achieved the deterrence and protection goal simultaneously.

The faculty-side cultural shift is visible in how detection tools are used. In year 1, many instructors treat a high detector score as a finding and bring concerns to the triage process at a high rate. In year 3, instructors who have been through the triage process several times understand that a score alone is a trigger, not a conclusion — and the proportion of concerns that proceed past triage without adequate corroboration falls as a consequence.

The year 2–3 failure mode: reverting to detector-first under workload pressure

The most common failure mode in the second and third year of a program is reverting to detector-first triage under workload pressure. This typically happens in periods of low-concern volume — when no cases have been formally investigated for several months — when the institutional memory of why the triage gate was implemented begins to fade. Staff who were trained on the rationale for the gate may have moved on; new staff may not have received the same grounding. Without regular reinforcement, the path of least resistance in a busy period is to skip the corroboration step and proceed on the score alone.

The structural solution is an annual training refresher tied explicitly to the documented case for the triage gate — and specifically to the metrics from the institution's own casework. An institution that can point to concrete cases where the triage gate prevented a wrongful formal investigation, or where the appeal upheld rate declined after the gate was consistently applied, has better evidence for the gate's necessity than abstract principle alone.

Program maturity trajectory

%%{init: {'theme': 'base', 'themeVariables': {'primaryColor': '#e8f4ee', 'primaryTextColor': '#16130c', 'primaryBorderColor': '#1f5a3a', 'lineColor': '#6f5618', 'secondaryColor': '#fffdf8', 'background': '#fffdf8', 'mainBkg': '#fffdf8', 'clusterBkg': '#f0ede4', 'titleColor': '#16130c', 'edgeLabelBackground': '#fffdf8'}}}%%
flowchart LR
    subgraph Y1["Year 1 — Foundation"]
        A1["Policy update\nand staff training"]
        A2["Pilot process\ncapture"]
        A3["Baseline metrics\ncollected"]
    end
    subgraph Y2["Year 2 — Calibration"]
        B1["Year 1 data\nreviewed"]
        B2["Assignment redesign\nfor highest-risk types"]
        B3["Capture coverage\nexpanded"]
        B4["First process-record\nresolutions"]
    end
    subgraph Y3["Year 3 — Cultural Shift"]
        C1["Students produce\nrecords proactively"]
        C2["Faculty treat scores\nas triage only"]
        C3["Appeal rate\ndeclines"]
        C4["Annual training\nrefresher embedded"]
    end
    Y1 -->|"Data informs\ncalibration"| Y2
    Y2 -->|"Calibration produces\ncultural change"| Y3
    Y3 -->|"Outcomes inform\nnext cycle"| Y1

---

Cost and resource planning

A common obstacle to implementing a three-tier integrity program is the absence of a clear resource argument. This section provides the framework for making that argument — both the direct costs of the program and the comparison against the cost of the failure mode it is designed to prevent.

Direct costs

Process capture tool licensing. Per-student or per-seat licensing costs vary by vendor and institution size. The cost should be calculated per assessed submission (not per enrolled student), since not all assignments will be in scope for capture. For most institutions, this is the largest direct cost item.

Staff training. The initial training investment for the three-tier model is a one-time cost covering detector limitations, the six-step investigation framework, and documentation standards. Annual refreshers are a smaller ongoing investment. Both are primarily staff time costs rather than external spend.

Policy documentation. The cost of reviewing and updating institutional AI policy documents, syllabus templates, and student orientation materials is a staff time cost concentrated in year 1. Subsequent annual updates are marginal.

Data retention infrastructure. If process records are held by the vendor, this cost is embedded in the licensing fee. If the institution retains records independently, it requires storage infrastructure sized to the retention period and the record volume.

Indirect costs

Investigations. The largest indirect cost in any integrity program is staff time for investigations. A well-designed triage gate reduces the number of cases that proceed to formal investigation; cases that proceed should be ones with adequate evidence, which reduces the cost of appeals. The net effect of a functioning three-tier program should be a reduction in total investigation cost relative to a detector-only program that produces a higher rate of formal proceedings without adequate evidence support.

Appeal management. External appeals to bodies like the OIA typically generate significant staff costs in document preparation, representation, and response. A program that reduces the appeal upheld rate reduces this cost directly.

The ROI argument

The cost comparison that makes the ROI case for process capture infrastructure is not process capture versus nothing. It is process capture versus the cost of a successful external appeal. A single successful external appeal — where a student's finding is overturned — typically involves internal staff time for documentation and response, potential legal costs if the matter escalates to litigation (as in the Adelphi case), reputational costs from public coverage, and the resource cost of reconsidering a case that should not have proceeded to a formal finding in the first place.

The per-student cost of process capture infrastructure is typically a fraction of the per-case cost of one successful external appeal. The argument is not that capture tools are free; it is that the evidentiary failure rate of detector-only programs generates a downstream cost — in appeals, in legal exposure, and in institutional reputation — that significantly exceeds the infrastructure investment.

Staffing

A process capture program requires a designated administrator who understands the system, can produce records when needed, manages the retention schedule, and oversees vendor compliance. At most institutions, this is approximately a quarter-FTE ongoing role — not a dedicated position. The program administration function can typically be absorbed into an existing academic integrity coordinator or registrar role with the addition of training on the specific system deployed. It is not a new hire; it is a defined responsibility within an existing role.

Cost planning table

Cost item	Rough magnitude	Notes
Process capture licensing	Per assessed submission or per enrolled student	Primary variable cost; varies significantly by vendor and scale
Initial staff training	1–2 days per cohort	One-time investment; covers detector limitations, investigation framework, documentation
Policy documentation update	10–20 staff-hours	Year 1 only; primarily existing staff time
Annual training refresher	0.5 days per cohort	Ongoing; protects against triage gate erosion
Data retention infrastructure	Minimal if vendor-hosted	Significant only if institution-side retention required
Program administrator time	Approx. quarter-FTE	Ongoing; can be absorbed into existing role with clear scope
Cost of one successful external appeal	Significant — multiple staff-days plus potential legal costs	Comparison point for ROI argument; not a program cost

---

Frequently asked questions

How should a university implement an AI writing integrity program?

A university should implement an AI writing integrity program as a three-tier framework. Tier 1 covers policy and student education: updating syllabus language to specify what AI use is and is not permitted for each assignment, training staff on detector limitations, and ensuring students understand both the rules and the evidence they can keep on their own behalf. Tier 2 deploys prospective process capture for assessed writing, with per-session student consent and cryptographically sealed records that belong to both the student and the institution. Tier 3 defines the structured investigation process, with a triage gate requiring corroborating evidence before formal proceedings open. The tiers must be implemented in sequence — evidence capture cannot be retrofitted after a concern arises.

What is the three-tier academic integrity model?

The three-tier model is a framework for defensible academic integrity programs that separates the program's functions into deterrence and education (Tier 1), prospective evidence creation (Tier 2), and structured investigation and adjudication (Tier 3). Each tier operates in a specific sequence and depends on the others. Tier 1 sets the policy expectations that give Tier 2 its rationale. Tier 2 creates the prospective behavioral record that gives Tier 3 its evidence base. Tier 3 generates outcomes and data that calibrate Tier 1 policy going forward. The model addresses the core failure of detector-only programs, which promote triage instruments into the role of primary evidence.

Why are AI detectors alone not sufficient for student misconduct findings?

AI detectors are not sufficient for academic misconduct findings for three independent reasons. First, the legal and regulatory record is explicit: both the Adelphi court ruling (February 2026) and the OIA's published casework guidance state that a detector score alone does not meet the evidentiary standard required for adverse action against a student. Second, the statistical problem is documented: Liang et al. measured a 61.3% average false positive rate for non-native English writers across seven widely used detectors — the tools systematically over-flag a large and legally protected student population. Third, detectors are probabilistic inference about a finished product; they produce no information about how the work was created. A finding requires evidence of behavior, not a score about the output.

What is prospective process capture in academic integrity?

Prospective process capture is the deployment of a tool that records the writing session — keystrokes, paste events, timing distributions, composition sequence — at the time the work is being produced, before any concern has arisen. The record is created with the student's explicit informed consent and sealed at session close using cryptographic techniques that make the record tamper-evident. The key word is prospective: the record is created before any dispute, which makes it independent of any incentive to fabricate. Both the institution and the student receive access to the same sealed record. This is the evidentiary infrastructure that allows integrity findings to rest on behavioral observation rather than probabilistic inference about a final product.

What evidence does an institution need to prove AI plagiarism?

An institution needs corroborating evidence that is independent of the detector score. The detector score may trigger an investigation but cannot constitute the finding. Corroborating evidence may include a sealed process record showing that the composition session does not reflect the typing pattern, timing distributions, or editing sequence of genuine human writing; significant stylistic inconsistency with a body of prior verified work from the same student; metadata indicating the document was created or modified in a timeframe inconsistent with the student's account; or, in stronger cases, a student who is unable to discuss the content of their submission in a post-submission conversation. Each piece of evidence should be documented with its source, how it was obtained, and any material limitations.

How should universities handle academic integrity appeals?

Universities should treat appeals as a quality assurance signal, not just an adversarial process. A high rate of successful appeals — where students are winning on appeal — indicates that the institution is not producing defensible findings at first instance. The appeal handling process should confirm that the panel file is complete and documents the panel's reasoning (not just its conclusion); that the student had adequate notice of the allegation and adequate opportunity to respond; that all evidence the student provided was assessed and that the assessment was documented; and that the burden of proof was correctly applied — the institution must prove misconduct, not the student disprove it. Appeals that succeed because the reasoning was not documented are preventable through training; appeals that succeed because the evidence was insufficient indicate a triage gate failure.

What are the most common reasons AI plagiarism findings fail on appeal?

Based on OIA casework patterns and the Adelphi ruling, the most common reasons AI misconduct findings fail on appeal are: treating the detector score as the finding rather than as a triage trigger; failing to document the panel's reasoning (recording only the conclusion); failing to assess evidence provided by the student — process records, drafts, research trails — and to document why it was or was not persuasive; running multiple detectors until one flagged (inflating apparent confidence without producing independent corroboration); failing to give the student adequate time to respond to the formal allegation; and not disclosing which detection tool was used, its version, and the vendor's own published caveats about its reliability. Most of these failures are procedural rather than substantive — the underlying concern may have been legitimate, but the process did not meet the standard required to sustain the finding.

What metrics should an academic integrity program track?

An academic integrity program should track five key metrics. First, concerns raised per semester — the volume at the triage stage, which tracks reporting patterns and triage threshold consistency. Second, rate of concerns closed at triage — the proportion not proceeding to formal investigation, which should not be zero; a zero rate indicates the triage gate is not functioning. Third, rate of findings upheld on appeal — the most important program health indicator; a high upheld rate means findings are defensible, a high overturn rate signals an evidence or process problem. Fourth, time from concern to resolution — tracking procedural timeline compliance. Fifth, cases resolved by process record — in either direction, exonerations and findings both; this tracks the Tier 2 investment's operational value.

How do FERPA and GDPR apply to writing process capture?

Under FERPA, writing process data collected through an institution-deployed tool is almost certainly an education record, which means students have inspection rights, disclosure is restricted, and the institution has data minimization obligations. The most common lawful basis for using a third-party vendor is the School Official exception, which requires that the vendor be under the institution's direct control with respect to the use and maintenance of student records — meaning the vendor cannot use the data for secondary purposes. Under GDPR, applicable to UK and EU institutions, the requirements additionally include a documented lawful basis for processing, student erasure rights, and restrictions on secondary use including training. The vendor contract must give the institution the ability to produce student records on request, delete them on request, and prohibit the vendor from using student data for any purpose other than the contracted service.

What does a defensible academic integrity finding require?

A defensible academic integrity finding requires: a triage decision that identified corroborating evidence independent of the detector score; a complete panel file that documents the initial concern and trigger, the triage reasoning, each piece of evidence gathered (with source and limitations), any population-specific considerations that affected the evidentiary threshold, the student's response and any evidence they provided, the panel's reasoning for how it weighted each piece of evidence, and the outcome and penalty with appeal rights. The finding must rest on the institution's affirmative evidence — not on the student's failure to disprove a suspicion. The process must have given the student adequate notice, adequate time to respond, and the opportunity to provide their own evidence. Any piece of student-provided evidence that was not persuasive must be documented with a reason. A finding that meets all of these conditions will survive external review. A finding that fails on any of them may not.

---

Related reading:

• Investigating AI-writing concerns: a six-step framework
• What forensic-grade means in academic integrity
• FERPA, GDPR, and student writing process data
• Syllabus AI policy language: the three-tier model
• Academic integrity evidence framework