Making AI Policy Enforceable In Higher Education

ai policy in higher education

Ask a university where it stands on AI and it will likely point to a policy. That’s a fair answer. In the 2025 EDUCAUSE AI Landscape Study, respondents reporting AI-related acceptable-use policies rose from 23% to 39% in a year, and K-12 adoption has climbed sharply too (GovTech; EDUCAUSE).

It’s just not the question that matters.

Ask instead what evidence shows that campus tools follow the policy, and the answers thin out. Verification and Validation (V&V) closes that gap: it is the evidence that a system meets its requirements and serves its intended use. This blog sets out a risk-based V&V framework for higher education technology leaders, aligned to the NIST AI Risk Management Framework (RMF), with a short note for K-12.

What this means for institutions:

  • Policies are spreading quickly. The open question is how institutions confirm they’re being followed.
  • Institutions and providers already share good questions (HECVAT) and contract terms (data privacy agreements). Shared, testable evidence is the next step, and both sides benefit.
  • A V&V framework aligned to the voluntary NIST AI RMF 1.0, with testing depth scaled to risk, turns policy statements into requirements that can be tested and evidenced.

This blog doesn’t intend to share any legal advice. Laws and frameworks are described as of September 2026.

Policy is spreading. Proof isn’t.

A policy states intent. Enforcement needs evidence of when practice drifts from it.

AI policy in education
Source: EDUCAUSE 2024 and 2025 AI Landscape Studies (via GovTech, 2025); RAND, "AI Use in Schools Is Quickly Increasing but Guidance Lags Behind" (September 30,2025); RAND Policy Minded podcast (August 12, 2026).
‍

Yet policy may not resolve real decisions. A 2026 preprint tested five Western Australian universities’ published governance against 15 scenarios. Only 6 of 75 cases were fully resolved; authority was often named only generically, without decision criteria (Poudyal). The study is small and outside the US, but the gap is familiar.‍

Source: Poudyal, "Stress-testing university AI governance," arXiv:2608.28925 (August28, 2026).

Even strong frame works show it. UT Austin’s “AI-Responsible, AI-Forward” framework sets out eight teaching-and-learning principles, including agency and stewardship, and is reviewed annually. Operational controls sit in separate security guidance. What’s missing is evidence that a tool, in a particular course, behaves consistently with those principles.

That missing link is testability. “Use AI ethically” cannot be tested. “A human must confirm every grade before release” can.

The rules are narrow, and everyone asks for proof

The US has no broad higher-education AI governance mandate. Instead, narrower rules require institutions or providers to show specific outcomes:

Rule What it covers What must be shown Timing
Colorado SB 26-189) Automated decision-making technology used in "consequential decisions," which the statute defines to include access to or eligibility for education Notice to the individual; a plain-language explanation of adverse outcomes; meaningful human review; records kept for 3 years Signed May 14, 2026; takes effect January 1, 2027 (Colorado General Assembly)
California AB 1159 Student data used in AI systems; creates the Higher Education Student Information Protection Act (HESIPA) Operators may not use student information to train or develop AI systems Signed September 10, 2026; higher-ed provisions operative July 1, 2027 (Privacy Rights Clearinghouse; F3 Law)
ADA Title II web and mobile rule Digital content and apps of public entities, including public universities Conformance with WCAG 2.1 Level AA April 26, 2027, for entities serving 50,000+; April 26, 2028, for smaller ones, after an April 2026 extension (UPCEA)

‍
Each rule depends on evidence: an accurate explanation, meaningful human review, data kept out of model training, or an interface that works with assistive technology. These are V&V questions.

Questionnaires are not enough

Institutions and providers already agree on many questions and contract terms. HECVAT shows how far that collaboration has come. Since2016 it has helped institutions assess providers against policy and risk appetite, and HECVAT 4 (February 2025) added community-developed AI questions and lets providers complete one annual assessment for multiple institutions (EDUCAUSE Review). The missing step is evidence that the important answers hold in each institution’s own deployment.

Existing mechanism What it establishes What V&V adds
HECVAT 4 (with AI questions) A common, reusable description of a provider's security, privacy, and AI practices Confirmation of the answers that matter most for high-risk uses, in the institution's own configuration
Accessibility Conformance Reports (VPAT) A documented accessibility baseline for the product Verification as deployed, with the institution's content and integrations Section508.gov
Student data privacy agreements (e.g., the SDPC National Data Privacy Agreement) Common contract language on data use Confirmation that actual data flows match the terms (SDPC)

‍
Neither side can close the gap alone. Interviews with 13 edtech leaders at seven institutions found that staff struggle to hold providers accountable because they lack visibility into provider systems (Kelso et al.). Providers, in turn, cannot see every customer’s configuration and use. Conformance, and its evidence, must come from both sides.

Reusable evidence can also shorten procurement and prepare providers for emerging documentation duties, including Colorado’s.

What a NIST-aligned V&V framework looks like

No new standard is needed. NIST AI RMF 1.0 expects testing, evaluation, verification and validation throughout the AI lifecycle, with documented results (NIST AI 100-1). Because the RMF is voluntary and has no certification, institutions should claim alignment, not compliance. NIST is revising the framework.

Verification asks whether a tool and its configuration meet institutional requirements. Validation asks whether it serves its intended educational use. These definitions are adapted from IEEE 1012-2024. In NIST terms, Govern and Map define policy, intended use and risk; Measure is where V&V happens; Manage determines when to repeat it.

V&V inside the NIST AI RMF

Three ideas make this practical.

  • Check both the product and the process. Product V&V tests whether the configured platform meets policy on data flows, fairness, explanations, accessibility and safe output. This is where AI assurance does its work. Process V&V tests whether people use the tool as required. “Faculty make the final decision” is a workflow requirement, not a software feature.
  • Scale the depth to the risk. Scrutiny should rise with consequence and likelihood, and independence matters most at the top tier: NIST recommends that V&V be performed by people distinct from test and evaluation teams. A Level 4 tool, such as admissions scoring, financial aid, grading or discipline flags, warrants independent testing, fairness checks and proof that human override works. A Level 1 back-office tool may need only provider attestation and spot checks. Keeping low-risk reviews light supports adoption.
an illustrative of scaling V&V depth to risk
Illustrative tiering. Each institution sets its own risk tolerance.
‍
  • Re-verify on change, not just on the calendar. Annual review is not enough when models change between cycles. Provider release notes should trigger re-verification, making updates a routine shared workflow.

From policy clause to evidence

Convert each policy statement into a requirement, a pass condition and an evidence artifact. The table below uses published policies and laws.

Policy statement (source) Testable requirement Evidence
"Agency" (UT Austin); the educator of record retains sole authority over final grades (California model policy) No grade is finalized without a named human confirming it Workflow test results and override logs
"Stewardship" (UT Austin); no training on student data (California AB 1159) Student data doesn't flow to model training or undisclosed third parties Data-flow test report plus the provider's attestation
"Pre-deployment evaluation" for bias and equitable outcomes (California model policy) Performance is comparable across student groups, within a documented tolerance Fairness test report
Plain-language explanation of adverse decisions (Colorado SB 26-189) Explanations are accurate, specific to the case, and readable Explanation review results
WCAG 2.1 AA (ADA Title II) AI interfaces conform as deployed Assistive-technology test report


Take a hypothetical Level 4 admissions-prioritization tool. Under an “agency” principle, every application must reach a human reader, and reviewers must be able to override scores. Tests sample low-ranked applications to confirm human review, then try to finalize a decision without a reviewer and confirm the system blocks it. The evidence is a test report and override log, both rerun after a model update. For the “stewardship” principle, the equivalent test is a data-flow check, and data assurance is what confirms that student data doesn’t reach model training or undisclosed third parties.

AI in higher education policy clause to evidence

Structured this way, a policy clause or HECVAT answer becomes reusable evidence for both institution and provider.

Start with the inventory, not another policy

Five steps get an institution moving:

  1. Inventory AI in use, including features added to the LMS, CRM and student information system.
  2. Assign each tool a risk tier.
  3. Turn high-tier policy clauses into testable requirements.
  4. Agree evidence terms with providers, including model-change notice and test access.
  5. Run baseline V&V on selected high-risk tools and repeat it after changes.

Few institutions have the capacity to build this alone, which is where advisory and consulting support can shorten the path from policy to first baseline.

A note for K-12

K-12 faces the same evidence gap with less capacity and younger users. RAND reports that 80% of principals said in spring 2026 that their school or district had AI policy or guidance, up from45% in 2025. Yet interviewed teachers often said they had no policy or found it confusing.

Requirements are also becoming more specific. Maryland, Oklahoma, Idaho and Virginia require district AI policies; Maryland requires an AI coordinator, and Oklahoma bars AI from primary use in grading, discipline and other high-stakes decisions. Each requirement can be tested. Because most districts cannot run deep V&V alone, state rubrics, consortium testing and reusable evidence matter even more.

Conclusion

Higher education has written strong AI principles. What is missing is a mechanism to demonstrate them. Risk-scaled V&V aligned to the NIST AI RMF provides it, and because institutions and providers already share questions and contract terms, shared evidence is the natural next step.

Begin with one high-risk campus AI system and ask: what evidence shows that it follows our policy today?

Related articles

No items found.
Education
Text Link
Artificial Intelligence
Text Link