AI Governance in Higher Education: A University CISO’s Playbook
Every AI governance guide I read assumes the same thing: one company, one org chart, one CEO who can issue a policy on Monday and expect compliance by Friday.
A university is not that place.
I am the CISO at Morgan State University, with 11,500 students and 3,000 staff. I govern AI across open networks, tenured faculty, 18-year-old students, a faculty senate that outranks my policy memos, and a general counsel who has seen every governance fad come and go. If you can govern AI here, you can govern it anywhere. This is the playbook I built, and it is the only one I know of written by a sitting university CISO.
- Map who actually decides before you write a single policy.
- Build a small steering committee that approves, not advises.
- Publish an allowed list instead of a blocklist, and make it the easy path.
- Hold the FERPA red lines without exception.
- Keep a human in every academic integrity decision.
Why higher ed is the hardest AI governance environment on earth
Corporations have hierarchy. Universities have shared governance.
My CEO equivalent is a president, but she does not rule by decree. Academic policy runs through the faculty senate, a body with constitutional authority over curriculum and teaching. I can recommend. I can warn. I cannot walk into a dean’s office and ban ChatGPT from the syllabus, because the syllabus is faculty territory, and faculty territory is protected by tenure and tradition.
Add the other complications. Our network is open by design, because research collaboration demands it. Our users include teenagers who grew up prompt-engineering before they could drive. Our data includes FERPA-protected education records, HIPAA-covered clinic data, and federally funded research subject to export controls. Our budget shrinks every year while the enrollment cliff keeps administrators up at night.
And the demand is real. Faculty want AI for grading and research. Students use it for everything, sanctioned or not. Admissions wants it for application screening. The helpdesk wants a chatbot. HR wants screening tools. Everyone is already using AI. The only question is whether anyone is governing it.
The stakeholder map: know who actually decides
Before you write a single policy, map the power. In higher ed, the org chart lies.
The provost
The provost controls academic affairs, which means the provost controls the faculty conversation. If the provost is not bought in, your governance program is a security memo that nobody reads. I brief our provost quarterly, in plain language, and I frame everything in terms of institutional risk and academic mission, never in terms of tools.
The faculty senate
You do not govern AI use in teaching without them. I learned this the expensive way: my first draft policy went to the senate and came back with forty comments. The second version, co-written with two faculty members, passed in one meeting. Lesson: governance by faculty, for faculty, with security as the enabler.
General counsel
Every AI decision I make runs past legal on two questions: what does FERPA say, and what does our state law say. Counsel does not write your governance framework, but counsel tells you where the red lines are. Find your red lines early.
Students
They are the primary users of AI on campus and the primary subjects of AI-driven decisions (admissions, proctoring, advising). If your governance program has no student voice, it will be resented and routed around. We have a student representative on our AI steering committee. It changed the conversation.
The board
How many AI systems, how many touch student data, what is the exposure, what are we spending. I report twelve decision-ready metrics to my board every quarter, and AI governance is now one of them.
What we actually built: a steering committee that works
Committees are where governance goes to die, so I designed ours to survive contact with reality.
Nine seats
CISO (chair), provost’s designee, faculty senate representative, general counsel, CIO, student representative, research office representative, HR representative, and one rotating seat for the department of the quarter. Nine is small enough to decide things and large enough that nobody can claim they were not consulted.
Monthly, 60 minutes
Agenda published a week ahead. No meeting runs over. Ever. The fastest way to kill a governance committee is to waste the time of busy people.
It approves
The committee does not advise. Tool approvals, policy changes, exception requests, all go through it. An advisory committee produces opinions. An approving committee produces governance.
Majority, with two vetoes
The CISO holds a security veto on data-protection grounds and the provost’s designee holds an academic-freedom veto on teaching grounds. In two years, neither veto has been used. The existence of the veto is what makes the votes honest.
Minutes in a week
Minutes published to the whole campus within a week. Nothing builds trust like showing your work. When faculty see that decisions are reasoned and documented, resistance drops.
We do not maintain a blocklist of banned AI tools. We maintain an allowed list of sanctioned tools, and we make the sanctioned path faster and easier than the shadow path.
The logic is simple. Blocklists are a game of whack-a-mole you will lose. Every semester brings new tools, and students adopt them faster than any committee can review them. A blocklist tells people what they cannot use and gives them no alternative. An allowed list tells people what they can use, with the university’s data protections already negotiated.
Our sandbox works like this:
Enterprise agreements first
We negotiated institutional licenses for two general-purpose AI platforms with data protection addenda: no training on our data, data residency guarantees, audit rights. These are the default. They cost money, but less than one incident.
Days, not months
Any department can request a new tool. The steering committee reviews it monthly. The review covers four questions: what data does it touch, where does the data go, what does the contract say, and who oversees it. Most requests clear in one meeting.
A 90-day lane
Unproven tools get a 90-day pilot with a defined user group, defined data boundaries, and a defined evaluation. Pilots that pass become sanctioned. Pilots that fail get a clear no, with reasons published. People accept “no” when they understand why.
The result: shadow AI did not disappear, but it shrank to the edges, and every department knows the sanctioned path is the path of least resistance. That is the whole game. Make governance the easy choice.
FERPA red lines: where I do not negotiate
FERPA, the Family Educational Rights and Privacy Act, protects student education records. In the AI era, “education records” covers more than most people think: grades, admissions files, advising notes, disciplinary records, and increasingly, the data exhaust of learning platforms.
My red lines, non-negotiable:
No student education records in AI tools without a signed data protection agreement
Not the free tier. Not “just this once for a pilot.” If a tool touches FERPA-protected data, it needs a contract with specific FERPA language, or it does not touch the data. I have killed three promising pilots on this rule. Each one hurt. Each one was correct.
No AI training on student data
Every vendor contract gets the same clause: our data is not used to train your models, period. If a vendor will not agree, we do not buy. This is a walk-away issue, and vendors know it, which is why most of them agree.
Directory information is not a loophole
FERPA lets institutions designate certain data as “directory information” (names, enrollment status), but AI systems that aggregate, profile, or make decisions about students go far beyond directory use. When in doubt, treat it as protected.
Research data gets its own lane
Federally funded research carries its own obligations: export controls, human-subjects protections, grant terms. Research AI use goes through the research office representative on the committee, not through the general IT path. Different rules, different reviewer.
Where I draw the line on proctoring AI
This deserves its own section, because it is where governance meets values.
AI proctoring tools promise to detect cheating through webcam monitoring, eye tracking, keystroke analysis, and room scans. They also produce false positives that disproportionately affect students with disabilities, students of color, and students in crowded housing. The research on bias in these systems is not ambiguous.
My position, stated publicly to our faculty senate: we do not deploy AI proctoring that makes automated academic integrity determinations about students. Human judgment stays in the loop, always. And any proctoring tool must clear a bias review before deployment, with the vendor’s testing data on the table, not in a marketing deck.
A tool can flag. A person decides.
Erdal Ozkaya
This cost me a vendor relationship and one very awkward meeting. I would make the same call again tomorrow. Some governance decisions are about risk. This one is about who we are.
The 6-month roadmap: steal this
If you are a university CISO starting from zero, here is the sequence that worked for me.
The bottom line
Higher education will not give you the clean org chart, the compliant users, or the big budget. What it gives you instead is a mission worth protecting: open inquiry, student futures, public trust. Govern AI like the mission depends on it, because it does.
Build the committee. Publish the allowed list. Hold the FERPA line. Decide the hard cases in the open. And remember that in a university, governance that faculty did not help write is governance that will not survive the semester.
About Author
