AI for Higher Education: Campus Assistants, Research Data Rules and When a University Runs Its Own Models

Campus-wide assistants, 2024 to 2026

Between 2024 and 2026, each of the large model providers released an education edition and started selling campus-wide licences. The details change often, so treat this as a snapshot of what the vendors’ own pages said on 25 October 2026.

These products solve the access problem well. A student or lecturer gets a strong model under institutional terms instead of a personal account with consumer settings. Some universities have gone further and cleared a campus assistant for sensitive data under their own contract; Stanford’s service page, for example, says its ChatGPT Edu deployment is approved for high-risk data and protected health information.

What a campus licence does not settle is whether a particular research contract, data use agreement or federal rule allows a dataset to leave the institution’s systems. That question is answered dataset by dataset, and for some of them the answer is no.

Where AI for higher education pays off: use cases by office

Teaching gets most of the attention, but the administrative and research offices are where the source material is already written down and the output can be checked line by line.

Student services. A chat assistant that answers from the registrar’s, bursar’s and financial aid pages, with links to the source, can handle routine questions at any hour and hand the rest to staff. It works when the content is current and the assistant says plainly when it does not know.

Advising knowledge bases. Advisers spend time looking up degree requirements, transfer credit rules, prerequisite chains and exceptions granted in past years. Retrieval over the catalogue, programme handbooks and advising notes lets an adviser ask a precise question and see the paragraph that answers it. Where advising notes are education records, access should follow the adviser’s caseload, not the whole department. The knowledge management use case hub shows how permission-aware search over internal documents is set up.

Admissions document review. Models can extract structured fields from transcripts, recommendation letters and portfolios, flag missing items and summarise a file for a reader. The admission decision stays with staff. In the EU, AI intended to determine access or admission is high-risk under Annex III point 3 of the AI Act, with obligations from 2 December 2027; our guide to the AI Act’s education categories covers that boundary in detail.

Research assistants. Literature search, summarising papers in a group’s own library, drafting methods sections and querying lab protocols are common requests. The private research assistant playbook describes a build that combines web crawl and retrieval on a group’s own corpus.

Grant writing support. Research development offices can use AI to check a draft against the funder’s solicitation, produce compliance checklists, draft biosketch and facilities boilerplate from approved text, and summarise reviewer comments from a previous round. Universities that run internal funding competitions can borrow the pattern in our grant application review workflow, where AI screens eligibility and completeness and panels receive cited summaries mapped to scoring criteria.

Administrative policy questions. Staff ask the same questions about travel, purchasing, leave and research compliance policies every week. An assistant that answers only from the current policy library, like the HR and policy helpdesk pattern, takes load off central offices and shows which policy and version it quoted.

Accessibility. AI can draft transcripts, captions, alt text and plain-language versions of course materials for a person to check and correct. That matters for US public institutions: the Justice Department’s Title II web and mobile accessibility rule now requires compliance by 26 April 2027 for public entities serving 50,000 people or more and by 26 April 2028 for smaller ones, after an interim final rule in April 2026 moved both dates back a year.

AI literacy. Faculty and staff need practical training on what each approved tool may be used for. An internal AI literacy training programme can be built on the university’s own policies and examples rather than generic material.

Why some universities run their own models

Research universities hold data with stricter rules than almost any company, and the rules often come attached to a specific contract or grant rather than to the institution as a whole.

Controlled unclassified information and the CMMC rule

Where a university holds a defence contract or subcontract that involves controlled unclassified information, the systems that process it must meet NIST SP 800-171; Revision 3 was published in May 2024 and supersedes Revision 2. The Department of Defense’s CMMC program rule is in 32 CFR Part 170, and the DFARS rule that puts CMMC requirements into contracts took effect on 10 November 2025. Fundamental research projects usually do not involve CUI, but engineering schools and applied research labs frequently hold contracts that do. Any AI tool that sees CUI is in scope. Our analysis of ChatGPT and Copilot under CMMC covers the cloud-side questions.

NIH controlled-access data

NIH notice NOT-OD-24-157 took effect on 25 January 2025. Approved Users of controlled-access human genomic data from NIH repositories must attest that their institution complies with NIST SP 800-171, and anyone using a third-party IT system or cloud service provider for analysis or storage must attest that the third-party system complies too. A genomics lab that wants to ask a model about dbGaP-derived data therefore needs either a provider it can attest for or a model running inside the institution’s compliant environment.

Export control

The Export Administration Regulations treat sharing controlled technology or source code with a foreign national on US soil as an export to that person’s country (the deemed export rule in 15 CFR 734.13). A shared AI assistant that indexes a lab’s export-controlled design files can make that release to anyone with access to the index. Collections for controlled projects need access limited to cleared project members, and the model provider’s own staff and locations become part of the export analysis if the data leaves campus.

Human-subjects data

Studies approved by an institutional review board come with consent forms and data management plans that say where identifiable data may be stored and who may see it. Many were written before generative AI and do not mention third-party model providers. Putting interview transcripts or clinical study data into an outside service may need an IRB amendment; keeping the analysis on the university’s systems usually fits the existing approval.

Student records under FERPA

FERPA allows disclosure of education records to a contractor without consent only when the contractor meets the school official conditions set out in the next section. That makes cloud assistants workable for many student services tasks, provided the contract is right, but some registrars and counsel prefer that bulk student records never leave institutional systems.

GDPR for EU universities

European universities process student and research data under the GDPR, including special category data in health and social science research. Transfers to providers outside the EEA need a legal basis under Chapter V, and research ethics boards often ask where processing physically happens. Running models in the university’s own data centre or a national research cloud keeps the question simple.

FERPA compliant AI tools: what the school official exception requires

Vendors often describe their products as FERPA compliant. The regulation does not certify products; it sets conditions on the institution’s disclosure.

Under 34 CFR 99.31(a)(1)(i)(B), a contractor, consultant or other outside party may be treated as a school official if it:

1. performs an institutional service or function for which the institution would otherwise use employees;
2. is under the direct control of the institution with respect to the use and maintenance of education records; and
3. is subject to the requirements of 34 CFR 99.33(a) on use and redisclosure of personally identifiable information.

The next paragraph of the same section adds a duty that matters a great deal for AI: the institution must use reasonable methods to ensure that school officials obtain access to only those education records in which they have legitimate educational interests. An assistant that indexes every advising note and lets any staff member query it fails that test regardless of where it runs.

For procurement, that translates into four checks: contract terms that put the vendor under the university’s direct control and bar secondary use, including model training; retention and deletion terms the registrar can enforce; access controls that follow roles and caseloads; and logs that show who retrieved which record.

Using the HPC and GPU clusters you already have

Many research universities already run shared high-performance computing clusters with GPU nodes, a batch scheduler and a research data storage service, often inside an enclave built for CUI or controlled-access data. That hardware can serve open-weight models.

A practical pattern is to reserve a small pool of GPU nodes for an inference service, serve models through an engine such as vLLM, and put a governed gateway in front of it so departments get an approved endpoint with quotas rather than raw cluster access. Batch jobs, such as extracting fields from ten thousand scanned forms or embedding a document archive, can still run through the scheduler at off-peak times.

Two constraints are worth planning for. Interactive assistants need GPUs that are always available, which sits awkwardly with a queue designed for long batch jobs, so most campuses separate the two. And the inference service inherits the security boundary of the cluster it runs on: a model in the CUI enclave can serve CUI projects, a model on the general cluster cannot.

AI policy in higher education: what to put in writing

Most universities published guidance on generative AI in teaching in 2023 and 2024. Fewer have an institutional policy that covers research data, administration and procurement as well. A complete policy usually addresses:

• Data classification by tool. A table that maps the university’s data risk levels to approved tools, so a researcher knows that a campus assistant is fine for public and internal data and that restricted data needs the private deployment.
• Approval of new tools. Who reviews an AI product before purchase, with security, privacy, accessibility and contract terms on the checklist.
• Teaching and assessment. A framework within which instructors set and publish course-level rules on AI use in assessed work, with assessment design and clear disclosure expectations at the centre.
• Research integrity. Disclosure of AI assistance in manuscripts and grant applications, in line with each funder’s and journal’s rules, and limits on uploading confidential peer review material.
• Research data agreements. A check that data use agreements, IRB approvals and export control plans permit the intended AI processing before a dataset is indexed.
• High-stakes uses. Human review and logging wherever AI touches admissions, grading, financial aid or student conduct, and for EU institutions a check against the AI Act’s high-risk categories.
• Staff literacy. Training for staff who use AI in their work. For European universities, the 2026 amendment to Article 4 of the AI Act keeps a duty on deployers to support the AI literacy of the people who operate AI on their behalf.

Which campus data can go where

A starting point for a data map, to be adjusted to each university’s own classification scheme and contracts.

How VDF AI fits a university

VDF AI is a platform a university installs in its own data centre, on a research cloud or inside an air-gapped enclave. Configured with local models only, prompts, documents, retrieval indexes, chat history and audit logs never leave the institution, which is the property research compliance offices need for CUI, NIH controlled-access data and export-controlled projects. VDF AI holds no certification or authorisation of its own; the deployment sits inside the university’s NIST SP 800-171 boundary, security plan and FERPA program like any other system the institution operates.

Each project, office or course can have its own knowledge collection. Retrieval respects per-collection access, so an advising assistant returns notes only for the adviser’s own students and a defence lab’s design files stay visible only to the project team. Multi-step jobs, for instance pulling fields out of admissions documents or testing a proposal against a funder’s solicitation, run as governed agents whose tools are granted per role and which wait for a staff approval before anything is sent or recorded. Every exchange is logged with the user, the model, the passages retrieved and any tool calls made.

On the university’s own GPU nodes, the VDF AI Router puts vLLM or Ollama model endpoints behind one gateway with departmental budgets, and it can add a cloud model for low-risk work where campus policy permits. Microsoft Entra ID sign-in works natively; a campus whose identity runs on Shibboleth, Okta or Keycloak plugs in via a reverse proxy that handles SSO, and every plan includes role-based access control. Research IT can trial the platform for three months on its own hardware before choosing a seat bundle.

The free AI readiness assessment helps a CIO or provost see whether data classification, infrastructure or governance will slow the first deployment, and the enterprise AI procurement guide lists the vendor questions that apply to any of the options in this article. This guide is general information, not legal advice on FERPA, export control or any specific agreement.