Responsible AI adoption in higher education: none, then one

Responsible AI adoption in higher education starts with every AI capability off, then one module for one programme, the decision record read, extend or stop.

Responsible AI adoption in higher education starts with being able to say no: on NUADU, AI is modular and opt-in per capability, and everything on the platform works with every AI capability switched off. That turns the question round. A QA office is not deciding whether to adopt AI across the board; it is deciding which module to switch on, for which programme, and what it needs to read before going further. Here is a four-step sequence a cautious institution can follow, what stays fixed whichever modules run, and where the DPO comes in.

What does responsible AI adoption in higher education look like?

A sequence, not a switch: run with everything off, switch on one module for one programme, read the decision record, then extend or stop. Each step can be explained to an academic board in a sentence, and none of them commits the institution to the next.

Policy, training and a task force are the usual starting points for AI in a university. Assessment needs an operational layer as well, because a grade, a flag or a credential is a decision about a named learner. The sequence is that layer.

  1. Run with everything off: author, deliver, conduct, mark, report and certify with no AI module enabled, the marker marking and a named educator signing every result.
  2. Switch on one module for one programme: the standards mapper on one programme's items, with a named human approving each proposed alignment, or AI-assisted marking on one course, with the marker still deciding the grade.
  3. Read the decision record before widening the scope: inputs, outputs, overrides, approvals and decisions are logged, and agreement rates between AI and markers are visible to QA.
  4. Extend or stop: extend to a second programme, or leave the module off entirely, and the rest of the lifecycle carries on either way.

The order is the safeguard. "What did it do, and who decided" is answered from the file rather than from a demo, and a module that cannot pass that reading for one programme has no business running for the whole institution. Step one has a single exception: the integrity defaults at exam start, lockdown browser and identity verification, which are configurable like every other integrity layer and supported by a DPIA template.

AI where you switch it on. Logged wherever it runs.

Who decides which AI modules run?

The institution, at a scope it chooses. Each AI module can be enabled per assessment, per programme, or per institution — or left off entirely. Run all of them, some, or none. The scope sits where assessment policy is already decided:

  • Per assessment — one exam, one module, one decision.
  • Per programme — a department's policy, applied across its assessments.
  • Per institution — one setting for everyone.
  • Per region or use case — for organisations running across jurisdictions.

The fourth scope comes from the enterprise side of the site, where AI scope is under the organisation's control: each AI module can be enabled per region, per use case, per programme — or disabled entirely. Its cross-region example is concrete, with integrity layers enabled in one jurisdiction and not in another, and it matters more for competence assessment for employees than for a single-campus university.

Because the inventory is named — eight capabilities and seven workflow assistants on the AI assessment tools page — a configuration decision can name the module it switches on. Who decides inside each of those modules is written down as well, and that is what the decision record will be read against.

What stays the same with AI switched on?

The named human, the rubric and the record. No AI output reaches a learner without a named educator's approval, every override is logged, and no grade, flag or credential is decided by AI. The modules that draft or mark apply your rubric, not a generic model's opinion.

Whichever modules run, the same audit trail records inputs, outputs, overrides, approvals and decisions for the same readers. That constancy is what makes the DPO's part manageable. A module switched on for one programme is, in data-protection terms, a processing activity with a defined scope, which is the shape a data protection impact assessment needs: under Article 35(1) of the GDPR, the controller assesses the impact before any processing likely to pose a high risk to people's rights and freedoms. The documentation pack, DPIA template included, is available on request to the reviewers who need it — QA, DPO, procurement, academic board — and the page on GDPR compliance for assessment covers the DPA, data residency and sub-processors.

Two threads lead on from here. The regulatory frame sits underneath the sequence rather than inside it, and why the EU AI Act treats assessment in education as high-risk is a question of its own. The learner side has a second switch: a module a learner can opt into, such as a practice route after a result, is switched on by the institution first and by the learner second — which is what an AI tutor for higher education should look like.

Frequently asked questions

What is opt-in AI?

AI that does nothing until someone with the authority to switch it on does so. On NUADU that someone is the institution, per capability and per scope. The integrity defaults at exam start are the one exception the AI page lists, and they are equally configurable.

How do you implement AI in higher education?

In assessment, one module and one programme at a time, with the decision record read before the scope widens. A first full deployment takes 2–3 weeks, depending on integration with your LMS, SSO and learner records, and no AI module has to be part of it.

Can you turn off AI features?

Yes. AI is modular and opt-in per capability, so a module can be switched off for one assessment, one programme or the whole institution, and the platform works with every AI capability switched off.

→ Walk through the AI with your compliance lead in the room (30 min)