How Do We Write and Implement an Institutional AI Policy?

The draft has been circulating since the spring, and every operative sentence says "may." Or a vendor wants a site license signed before the end of the quarter, and nobody on the call can say what the institution has agreed to about student data. Or a detector flag has become an appeal, and the policy that was supposed to settle it turns out to say nothing about what evidence an allegation can rest on.

What a committee can defend is narrower than most drafts attempt, and publishing is the smallest part of the work. Four things turn a document into a working rule: a named executive owner, a scheduled review, a translation path into courses with someone accountable for it, and a graded assessment design that rewards the behavior you are asking for. Add a scope sentence that says which obligations the document carries and which belong to another policy, because one document absorbing data protection, research ethics, clinical regulation and academic conduct will be long, unread and unenforceable.

The short version

1. Declare what the document governs. If it is a risk-management instrument, say so, and put the teaching detail at department level.

2. Count the modal verbs. Each "may" where you mean "must" is a provision nobody can enforce.

3. Name an owner, a standing committee and a review date before publication, not after the first appeal.

4. Write down where FERPA, GDPR, HIPAA, IRB and research-integrity obligations live, and name the office that owns each one.

5. Fund staff training and student access in the same budget decision. Either one alone produces a rule some students cannot follow.

6. Ask how many courses detectably translate the policy, and read the answer as a support metric rather than a compliance metric.

Decide what the document governs

Before the wording, the scope. Manikonda and Outlaw (2026) crawled AI policies from 149 R1 and R2 universities in the United States and verified 130 university-level policies across 34 states. The register is advisory rather than directive: 95% of the policies carried a negative clarity-strength score, so weak language ("may," "can") dominates over directives ("must," "prohibited").

What those documents prioritize tells you what their authors thought the document was for. University-level policies cover data security, risk mitigation, procurement and legal compliance, with little pedagogical guidance. The school-level policies that exist concentrate on teaching and AI Literacy. Only eight business schools had a policy of their own, and those differed from their host framework in six of eight cases.

The remedy the authors propose is a layered structure: a university-wide risk-management layer, department-level policies, and inter-departmental committees with faculty and students. Adopt that and you get a scope sentence you can defend in the senate. This document manages institutional risk; the teaching decisions sit with the department that grades the work. Layering also answers the charge of vagueness without producing 40 pages.

Write provisions that bind, or take them out

Look at what programs publish when nothing forces the question. Eldredge et al. (2026) scanned all 48 CAHIIM-accredited health informatics and health information management master's programs in the US. Forty (83%) published at least one AI-related document. Of the analytic sample, 21 documents (53%) were guidelines, 9 (23%) informational, and only 7 (18%) formal policies.

The subject matter was narrower than the volume suggests. Academic integrity was the most frequent keyword (n = 139), ahead of citation (59), Assessment (50) and plagiarism (38). Disclosure of AI and contract cheating each appeared once. HIPAA appeared five times, FERPA eleven, and electronic health records not at all. Equity vocabulary was close to absent: inclusion (n = 9), accessibility (n = 9), equitable access (n = 2).

That is attention without a binding commitment behind it, which is the cheapest defect to fix. Count your modal verbs and write a consequence where you mean one. Define plagiarism and prohibited uses in words a first-year student can apply. Then score the result against the instrument Gutowski and Hurley (2025) built: five dimensions with explicit rubrics, covering prohibitiveness, permissiveness, educational integration, transparency and accountability, and depth. Most institutions in their sector took generally prohibitive positions while reserving discretion to individual instructors. No single approach was accepted, and clarity itself was treated as the precondition for defensible enforcement. Their recommendations are procedural: comprehensive guidelines whatever your stance, stakeholders involved in drafting, proactive training, governance designed to flex, and review cycles.

Decide who owns it and who sits on the committee

Two reviews put the weakness in capacity rather than intent. Ashiq (2026) synthesized more than seventy UK sources and found that only a minority of institutions maintain official AI governance plans. Strategic planning gaps produce policy drift and performative compliance. National guidance from the Office for Students and Jisc has been criticized for lacking specificity and enforcement power. Staff AI-training and infrastructure investment concentrate in research-intensive institutions, while teaching-led institutions face capacity constraints, widening a digital divide. Baroudi (2026) reviewed 19 sources and found that 7% of institutions had created senior AI leadership roles against 49% treating AI as a strategic priority. Weak policy frameworks and limited digital infrastructure drive that gap, particularly in the Global South.

Ownership, resourcing and revision are where the effort has to go, so settle the chair and the membership before drafting starts. Crompton et al. (2026) convened a Delphi panel from 22 countries and locations across six continents and produced a consensus framework with eight core areas: academic integrity, ethical and responsible use, privacy and protection, equitable access, GenAI literacy, integration strategy, human oversight and accountability, and institutional support and infrastructure. The panel also produced a six-part mechanism for keeping policy current: a multidisciplinary governance committee (backed by more than half the panel), scheduled review cycles (half), professional development, communication with all Stakeholders, evaluation of effectiveness, and monitoring of external developments.

Read the framework as a coverage checklist and the mechanism as an operating model. The panel's own weighting is the warning. Ethical and responsible use was referenced by two-thirds of panelists and privacy by over half, while equitable access drew only one-third. A consensus process under-weights equity first, and your committee will drift the same way unless someone holds that item.

Two membership decisions are worth making deliberately. Hingle and Johri (2026) had students co-design a GenAI course policy through guided inquiry. The priorities that emerged were training for students and instructors, standardized disclosure procedures, stronger institutional support rather than reliance on individual instructors, and a role in decisions about the rules governing their own learning. That argues for student membership rather than a consultation round. Baroudi found empowering and distributive leadership styles associated with higher faculty engagement and openness to change, and documented Change Management machinery (Valente's contagion model, Rieber and Williver's five-stage framework) as the support for adoption.

Tan et al. (2026) offer a six-dimension organizational framework whose propositions the authors explicitly leave for future empirical validation. Use it to structure what the document covers, and keep it out of the sentence where you claim something works.

Put the legal obligations where they can be enforced

Compliance with baseline law is not ethical data governance. Varadaraju and Vijayakumar (2026) argue that Learning Analytics governance in higher education has stayed compliance-first, centered on meeting FERPA and GDPR, and propose the LEAGUE framework: Lawfulness, Equity, Agency, Governance, Utility and Ethics by Design, demonstrated against an early-alert case study. The Delphi panel independently ranked privacy and data protection as critical for over half its members. The OECD Digital Education Outlook (2026) makes an enabling environment for trustworthy GenAI its third policy pillar, covering Privacy, safety, bias testing and transparency.

The health informatics corpus shows what scope discipline looks like when it works. Those programs barely mentioned HIPAA, FERPA or IRB guidance and never mentioned electronic health records, because those obligations live in broader compliance policy that already has an owner, a procedure and a penalty. Name where each obligation sits, name the office that holds it, and require students to meet it in coursework rather than in a footnote.

That paragraph also settles a procurement question. If bias testing and transparency are pillar conditions rather than aspirations, they belong in the purchase terms, and the policy is the place to say so.

Close the gap between the policy and the courses

This is the decision most policies fail, and it is measurable from both ends.

The faculty survey first. The AAC&U/Elon survey of 1,057 US faculty (Watson and Rainie, 2026) found 87% had created their own policies for students on generative AI use, while 48% said their institution had written guidelines and 35% said their department had. On structure, 55% reported a task force or oversight group, 37% new AI-focused classes, 17% an AI major or minor, 16% new academic leadership offices, and 13% had adopted AI literacy as a general education learning outcome. The sample is non-scientific and the authors say it is not generalizable. The asymmetry stands anyway: the rules students actually meet are written by the instructor in front of them.

Then the translation count. Ganguly et al. (2026) analyzed 116 institutional GenAI policies from 131 R1 universities alongside 98 computer-science course syllabi from 54 of those institutions. Institutions mostly encouraged GenAI use (73, 63%) while 31 (27%) discouraged it. At course level, 90 syllabi (92%) gave explicit guidelines, yet 49 (50%) prohibited GenAI outright, 40 (41%) permitted partial use for specified activities, and 7 (7%) communicated encouragement. Only 47 of the 131 institutions had both an institutional policy and a detectable course-level translation.

Put that question to your committee in those terms: how many of our courses detectably translate this document? A low number is a support failure, and the remedy is worked examples, faculty development and assessment support published alongside the rule. That support ecosystem is how Qian (2026) separates institutions from one another, and it is the difference between a policy students receive and a policy their instructor improvises around.

Coates, Croucher and Calderon (2025) supply the instrument for the accountability end: an academic integrity indicator framework built for institutional governors through research reviews, multi-institutional case studies, prototyping and expert confirmation, together with reforms to governance architectures, people and technologies. Indicators let a committee report upward without commissioning another literature review.

Write the evidence standard before you need it

Where enforcement fails, the failure is usually evidentiary, and this is where institutional exposure sits.

Munoz et al. (2026) analyzed actual generative AI misconduct allegation files. The evidentiary base was thinner than policy implies. System-recorded behavioral traces exist only in supervised assessment. Weaker process evidence, meaning drafts, supervision meetings and presentations, exists only where those practices were already routine. Natural justice requires that the student be told the allegation and given a chance to respond before any determination, usually in an investigative meeting or panel interview, with whatever they say entering the evidentiary record.

Detector output cannot carry that load. Hadra et al. (2026) measured accuracy of 0.69 and 0.61 for two widely used commercial tools, both failing on hybrid human-AI text, and both carrying a borderline misclassification risk for EFL student writing. Bassett et al. show that no threshold resolves the tradeoff between flagging honest work and missing concealed use. A finding that rests on a score nobody can interrogate is the case that generates the appeal, the complaint and occasionally the claim.

Three sentences belong in the document itself: what evidence a finding may rest on, that a detector score alone is not that evidence, and who decides when a case proceeds. Put the procedure beside the obligations collected in Legal Issues and Risks.

Say what you will measure before you publish

No study in this corpus measures whether an institution's AI policy changed student behavior, staff practice or learning outcomes. What exists is descriptive. Manikonda and Outlaw find policy language weak and misaligned. Ashiq documents drift. Baroudi names the absence of longitudinal and causal evidence as a gap. Eldredge et al. caution that analyzing public documents creates a visibility bias, so missing privacy or equity language is a finding about published guidance rather than about practice.

Two habits follow. Attach an evaluation plan to the policy: what you will measure, when, and who owns the finding. And when a colleague cites another institution's policy as proof that a measure works, check whether the claim is about a document or about an outcome. Tran, Liu and Nguyen (2026) reviewed 65 peer-reviewed papers on AI and social-emotional learning and found nearly three-quarters mentioned no policy implications at all. Of the minority that did, most lacked the actor-specific detail real policymaking needs: who should act, on what, why, when and how.

What institutions get wrong about this

Publication is the finish line. The estate Manikonda and Outlaw mapped is broad and shallow, advisory in tone, and thin on the problems that decide whether AI use goes well. A document with no owner and no review date is a press release with a policy number.

Vague wording is the safe wording. It is safer for the person who signs it, and the weakest thing to defend in an appeal. Gutowski and Hurley treat clarity as the precondition for defensible enforcement, and 95% of the policies in the R1 register score negative on clarity.

Academic integrity is the hard part, so cover that and the rest can wait. Integrity is the easiest provision to write, which is why it dominates keyword counts while equity vocabulary, access and data obligations sit near-absent. The provisions an integrity code cannot reach include procurement, disclosure procedure, accessibility, staff training and the assessment conditions under which integrity is achievable.

One document can hold every obligation. It can hold them and enforce none of them. HIPAA appeared five times across 48 programs' AI documents and electronic health records never, because those obligations already belong to compliance policy.

Faculty will follow once it is published. 87% of faculty had written their own rules before their institution published one. The counter-measure is professional development, worked examples and distributive leadership. A reminder email moves nothing.

Students ignore policy, so clarity is wasted on them. Braun and Khafizov (2026) surveyed 1,809 students, 250 faculty and 62 administrative staff at one teacher-education university. Students reported higher AI-use intensity and usefulness while faculty and staff reported stronger integrity concerns, and students reported higher perceived policy clarity than faculty did. What moves them is whether the tool is useful enough to be worth following a rule about.

What will come back at you

"We already have an academic integrity policy." It is necessary and not sufficient. What the evidence supports is division of labor: conduct rules stay where the integrity process can enforce them, and the AI policy carries procurement and data security, disclosure procedure, accessibility and access, staff training, and the assessment conditions under which integrity is achievable. The Delphi panel's insistence that significant GenAI outputs be reviewed by a human, and its preference for process-focused and oral assessment, has no home in an integrity code.

"Faculty will ignore it." Some will, and the reason is rarely will. Ashiq's evidence is capacity. Baroudi's is structure: 7% of institutions had created senior AI leadership roles against 49% treating AI as a strategic priority. Ganguly et al.'s is translation: 47 of 131 institutions had both a policy and detectable course-level implementation. Faculty are being left alone with the document.

"We cannot police it, so the policy is decoration." You cannot police it, and the evidence says you do not have to. Ogbo et al. (2026) model student AI use as a coordination problem in which cohort norms govern behavior, with threshold effects. Reflective assessment has to be rewarded above a critical level before responsible use displaces opportunistic practice, the reward has to be proportionate to the effort reflection demands, and peer sensitivity sets how fast a norm cascades. Students respond to assessment structures rather than to pronouncements. On Braun and Khafizov's cross-sectional self-report data, perceived usefulness is the strongest standardized association with Trust (β = 0.402) against institutional policy clarity (β = 0.223). The OECD adds that general-purpose chatbots improve output quality, that the advantage disappears and sometimes reverses in exams when AI access is removed, and that metacognitive engagement drops as work is offloaded. Put the requirement in the graded assessment.

"Students will not read it." They respond to it anyway, and the response is shaped by what their instructor says in week one. A policy only a committee has read has one reader.

"We cannot fund the support the framework asks for." Then cut the framework to what you can fund and say so in the document. The panel placed institutional support and infrastructure as the foundational enabler, and Adarkwah et al. (2026) analyzed 159 documents from 30 highly ranked universities in the ten countries best placed on the IMF AI Preparedness Index, four of which were excluded for lacking publicly accessible policies, leaving 26, and found many provisions declarative rather than operationally assured. A mandate to use AI without devices, connectivity, accommodations or training is an obligation the institution has assigned and cannot support.

What to do this week

1. Search the draft for "may" and "can," rewrite each instance where you mean "must" or "prohibited," and define plagiarism and prohibited uses in words a first-year student can apply.

2. Name one executive owner, one standing committee with faculty and student members, and one review date. Give the monitoring of external developments to a named role.

3. Decide the layering: which provisions are university-wide, which the department writes, and which obligations belong to FERPA, GDPR, HIPAA, IRB and research-integrity policy.

4. Write the evidence standard and the detector sentence, then read your own appeal procedure against them.

5. Ask for the translation number, and pair any new restriction with a worked example and a training date.

Where to go next

How Do I Write a Course AI Policy and Communicate It to Students? is the classroom counterpart, where the document becomes an assignment sheet and a spoken sentence. The research base sits in institutional governance research, the global Delphi framework and program-level policy audits. Administrators and Change Management cover the leadership and implementation dimensions, Educational AI Policy collects the sector-wide literature, and AI Governance covers how authority and accountability are distributed. On the enforcement side, Legal Issues and Risks carries the obligations a detector finding has to meet.