Before scaling AI, pause

Universities do not need to choose between banning artificial intelligence and adopting it everywhere. They need a disciplined way to decide where AI adds educational value, where it creates unacceptable risk, and what human capabilities must grow alongside it.

This challenge is especially important in Pakistan. Institutions are working with different levels of connectivity, staff preparedness, student access, data protection and administrative capacity. A policy copied from a well-resourced foreign university may therefore look impressive while remaining difficult to implement. Responsible adoption must begin with the institution’s actual students, teachers, systems and constraints.

UNESCO’s guidance on generative AI calls for a human-centred approach that protects privacy and supports ethical, safe, equitable and meaningful use in education. Its emphasis is not merely on controlling tools, but on building coherent policy, human capacity and pedagogical design. The NIST AI Risk Management Framework offers a complementary organisational logic: govern AI use, map its context, measure risk and manage it over time. NIST also makes clear that its framework is voluntary and is currently being revised, so it should be used as a practical reference rather than treated as law or a fixed compliance certificate.

PAUSE-AI translates these ideas into six questions that a university leadership team can use before approving an AI initiative. It is an original decision framework, not a formal standard or accreditation instrument.

P: Purpose and educational value

Question: What human or educational purpose will this use of AI serve?

Begin with the problem, not the product. “We should use AI” is not a purpose. Reducing repetitive administrative work, giving students more opportunities for formative feedback, helping faculty develop differentiated learning resources, or improving access for learners with disabilities are clearer purposes.

For every proposed use, write a one-sentence value hypothesis:

If we use this system for [specific task], it should improve [defined outcome] for [named users], without weakening [essential human capability or right].

Then ask whether AI is necessary. A simpler workflow, better training or a conventional digital tool may solve the problem with less cost and risk.

Evidence to record: the problem statement, intended users, desired learning or service outcome, non-AI alternatives considered and the capability that must remain human.

A: Accountability and human judgement

Question: Who is answerable for the decision, the output and any harm?

An AI system cannot own academic responsibility. A named person or role must remain accountable for approving the use case, reviewing outputs, responding to errors and stopping the system when necessary. “Human in the loop” is too vague unless the institution defines what the person checks, when intervention is required and whether they have the authority and time to act.

Accountability should be particularly explicit when AI influences admissions, assessment, student support, employee evaluation, research or the allocation of opportunities. People affected by a consequential decision should know when AI has materially shaped it and how to seek review.

Evidence to record: accountable owner, review procedure, escalation route, appeal mechanism, disclosure requirement and shutdown authority.

U: Users, participation and equity

Question: Have the people most affected helped shape the decision?

Responsible adoption is not something an IT office does to a university. Faculty, students, professional staff and relevant specialists should help define the problem, test the system and identify harms that a procurement team may miss.

In Pakistan, equity cannot be reduced to whether a tool is technically available. Institutions should examine device access, bandwidth, language, disability, digital confidence, paid versus free accounts and the hidden cost of requiring students to use commercial services. A system that works well for confident English-speaking users on reliable broadband may perform very differently across the actual student population.

Evidence to record: stakeholder groups consulted, accessibility test, language test, low-bandwidth option, non-AI alternative and support for students who cannot or do not consent to use the tool.

S: Safety, privacy and knowledge integrity

Question: What information enters the system, where does it go, and how will errors be detected?

Do not place confidential student records, unpublished research, examination material, personal health information or commercially sensitive data into a public AI tool without an approved basis and appropriate safeguards. The institution should understand retention, model-training, access-control and deletion arrangements rather than relying on a generic assurance that a platform is “secure”.

Safety also includes knowledge integrity. Generative systems can produce confident but incorrect information, fabricated references and biased representations. Verification must therefore be part of the workflow, not an optional final glance. The level of verification should rise with the consequence of the task.

Evidence to record: data classification, approved inputs, vendor terms, retention settings, access controls, required verification method, incident-reporting route and prohibited uses.

E: Evidence, evaluation and evolution

Question: What would demonstrate that the initiative works, and what would make us stop it?

A successful demonstration is not evidence of educational value. Before a pilot begins, establish a baseline and select a small number of meaningful measures. Depending on the use case, these might include learning quality, staff time, error rates, student experience, accessibility, cost and the frequency of human corrections.

Pilot with a bounded group and a defined end date. Compare the result with the previous practice or a credible alternative. Record failures and unintended effects, not only success stories. Because AI tools, terms and capabilities change quickly, approval should expire unless it is reviewed.

Evidence to record: baseline, success indicators, harm indicators, pilot population, evaluation date, decision threshold, review cycle and exit plan.

A: Alternative futures and second-order effects

Question: If this use scales, what futures might it create for learning, work and institutional trust?

Most AI decisions are assessed only for immediate efficiency. Strategic foresight adds a longer view. Leadership teams should consider at least three plausible outcomes:

  • Assisted capability: AI removes low-value friction while people gain stronger judgement and domain expertise.
  • Uneven advantage: benefits concentrate among already well-equipped students or departments, widening gaps inside the institution.
  • Capability erosion: dependence on automated outputs weakens writing, reasoning, teaching expertise or institutional memory.

These are not predictions. They are lenses for noticing consequences early. For each scenario, identify signals that would show it is beginning to emerge. A rise in output speed may look positive, for example, while falling ability to explain or defend that output may indicate capability erosion.

Evidence to record: three plausible scenarios, early indicators, affected capabilities, distributional effects and a response option for each scenario.

The one-page PAUSE-AI decision checklist

Before approving an AI pilot, the leadership team should be able to answer yes to all twelve statements:

Purpose

  • We can name the specific educational or organisational problem.
  • We have compared AI with simpler, lower-risk alternatives.

Accountability

  • A named human role owns the decision and its consequences.
  • People affected by consequential decisions can obtain explanation and review.

Users and equity

  • Relevant students, faculty and staff have participated in design or testing.
  • An accessible, affordable and workable alternative exists for excluded users.

Safety and integrity

  • Data inputs, retention, access and prohibited uses are documented.
  • Outputs are verified in proportion to their academic or human consequence.

Evidence and evolution

  • The pilot has a baseline, measures, end date and stop criteria.
  • Approval will be reviewed when the tool, terms or context changes.

Alternative futures

  • We have examined benefits, unequal effects and possible capability erosion.
  • We know which early signals will trigger adaptation or withdrawal.

If any answer is “no”, the initiative is not necessarily rejected. It is simply not ready to scale.

A 90-minute leadership exercise

Universities can use PAUSE-AI in one working session:

  1. Choose one real use case, 10 minutes. Keep the scope narrow, such as AI-assisted formative feedback in one course.
  2. Complete the six PAUSE-AI questions, 35 minutes. Require evidence, not assumptions.
  3. Create three future snapshots, 20 minutes. Describe the assisted-capability, uneven-advantage and capability-erosion outcomes one year after scale-up.
  4. Define the pilot, 15 minutes. Set its users, duration, measures, safeguards and stop criteria.
  5. Assign ownership, 10 minutes. Name the accountable lead, reviewer and next decision date.

The output is not another broad AI policy. It is a defensible decision record for one real use.

The leadership principle

The most future-ready university will not be the one that adopts the greatest number of AI tools. It will be the one that improves learning and institutional capability while keeping responsibility, inclusion and human judgement visible.

That is the purpose of the pause: not to delay innovation, but to make innovation worth scaling.

Continue reading

A companion article addresses the learning side of the same question: AI literacy is more than prompting: what Pakistan’s universities need to build. Related topic pages cover AI literacy in Pakistan and strategic foresight in Pakistan.

Call to action

Dr. Salman Khatani and Fiker Futures Academy work with universities, schools and organisations on AI literacy, strategic foresight and responsible, human-centred transformation. To adapt PAUSE-AI into a leadership briefing, faculty workshop or institutional readiness lab, contact Fiker Futures Academy or review the dedicated Dr. Salman Khatani profile.

Authoritative references