AI Literacy Is Not Prompt Engineering: Universities Need to Teach Students How to Evidence Their Thinking

Standfirst: In an AI-rich university, the final submission is no longer enough. Students need to show how they thought, selected, checked, revised and took responsibility.

AI-assisted work is already happening

The question for universities is no longer whether students will use generative AI.

They already are.

In 2025, reporting on the Higher Education Policy Institute and Kortext student survey stated that use of any AI tool among UK undergraduates had risen from 66% in 2024 to 92% in 2025, with 88% saying they used tools such as ChatGPT for assessments (Guardian, 26 February 2025). The precise numbers will continue to change, but the direction is clear: AI-assisted study is no longer marginal.

That changes the assessment problem.

For the last few years, much of the institutional conversation has been dominated by detection, misconduct, bans, tool access and generic guidance. Those questions matter, but they are not sufficient. The deeper educational issue is this:

If students use AI, can they evidence the human thinking inside the work?

That question moves the debate away from a narrow concern with whether AI was used and towards a more useful concern with judgement, authorship, verification and responsibility.

What the supplied AI-in-HE papers show

The papers provided in the AI in HE folder do not all make the same argument. That is useful. Taken together, they show a field moving between enthusiasm, caution and practical implementation.

Several papers present AI as a potential support for personalised learning, administrative efficiency, feedback, creativity and future-facing digital capability. Dhupal (2025), Bholane (2025), Sarumi (n.d.) and Slimi and Villarejo-Carballido (2024) all describe AI as having significant potential for teaching, learning, administration or assessment, while also acknowledging risks around privacy, bias, academic integrity, institutional readiness and ethical governance.

Other papers are more sceptical of transformation rhetoric. Murgatroyd (2024) warns that AI is often presented as a market-ready transformation of teaching, learning and assessment, while the practical reality may be slower, more incremental and more burdensome. Ahmed (2024) explicitly argues for careful and rational evaluation of AI integration rather than accepting global pressure or persuasive technology discourse. Xiao and Lim (2026) go further by asking whether AI is actually a solution to higher education's underlying problems, proposing a technology-agnostic decision rule: adopt technology only where it enables educators to do what they otherwise cannot, does something demonstrably better at affordable cost, or does the same work while reducing cost without undermining education.

Govil (2025) offers a particularly important bridge for this article. The paper argues that AI's impact on learning depends less on the technology itself than on how it is used: judicious engagement may scaffold learning and cultivate new literacies, while indiscriminate reliance risks cognitive offloading, demotivation and loss of critical thinking. Weimann-Sandig (2024) similarly frames AI as changing the role of university educators rather than simply removing that role, with future skills and digital literacy becoming part of the educational task.

Raymond's Higher Education AIDE Framework (2026) is especially relevant because it moves from general policy anxiety to programme-level design. It argues that higher education adoption is often fragmented, either too narrowly focused on plagiarism or too broadly limited to policy statements, and proposes a structured programme model: Awareness, Integration, Development and Evaluation. That aligns closely with the local action framework developed in this folder: institutions set the conditions, programmes sequence and coordinate, and modules enact practice.

The combined message is not "AI will transform everything" or "AI should be resisted". It is more practical: universities need purpose-led adoption, assessment redesign, staff and student AI literacy, ethical governance, equity awareness and evidence of human judgement.

The final product is no longer enough

Traditional assessment often assumes that the submitted artefact is a reliable proxy for the student's learning. An essay, report, presentation, reflection or design may be marked as if it shows the student's knowledge, reasoning, choices and authorship.

Generative AI weakens that assumption.

This does not mean every AI-assisted submission is dishonest. It means the final product alone may tell us less than it used to. A polished answer can now be produced with limited understanding. A fluent paragraph can hide weak source handling. A confident explanation may contain fabricated claims. A plausible reflection may say little about what the student actually thought.

The answer is not simply to retreat into detection. AI detection is uncertain, uneven and often adversarial. Nor is the answer to ban AI across the board. Blanket prohibition may be appropriate in some assessment contexts, but as a general institutional strategy it ignores the reality that graduates will enter workplaces where AI-assisted writing, analysis, coding, design and communication are increasingly common.

This is where the supplied literature converges. Grzybowski (2025) treats AI as a significant challenge to university education and assessment because it unsettles the relationship between authorship, written work and academic verification. Raymond (2026) argues that higher education needs to move beyond reactive, fear-based responses and embed responsible AI into curricula. Govil (2025) warns that over-reliance can weaken the very capacities universities are supposed to cultivate. These papers support the same conclusion: final-product assessment is fragile when the process behind the product is invisible.

The more durable response is assessment design that asks students to evidence process, not only product.

What process evidence means

Process evidence does not mean adding a long portfolio to every assignment. It means deciding, deliberately, where the student's route to the final artefact matters.

Students might be asked to show:

• the original task, question, source material or problem;
• where AI was used and where it was not used;
• examples of prompts or instructions;
• outputs they accepted, rejected or revised;
• drafts, plans, notes or decision logs;
• source-checking and fact-checking trails;
• how they dealt with bias, hallucination, privacy or data concerns;
• what human judgement shaped the final submission;
• a short authorship statement explaining what they are taking responsibility for.

The aim is not to burden students with bureaucracy. The aim is to make learning visible.

AI literacy should therefore mean more than prompt engineering. It should include the ability to explain, verify, critique, contextualise and take responsibility for AI-assisted work.

This also responds to the equity concerns in the supplied literature. Ahmed (2024), Sarumi (n.d.) and Xiao and Lim (2026) all raise questions about uneven access, institutional pressure, bias or inequity. If AI use is left implicit, confident and well-resourced students are likely to benefit first, while more cautious students may avoid tools or use them anxiously. Process evidence can help make expectations explicit and teachable.

Storybook as a small example

A recent Storybook Film Studio project offers a useful illustration, although it is not the main evidence for this article.

The project began with a child's source story and rapidly expanded into illustrated pages, music, films, premiere pages, research synthesis, public communication and audit documents. The interesting educational point is not simply that AI-assisted production was fast. It is that the process could be traced.

There was source material. There were visual outputs. There were accepted and rejected directions. There were research folders, planning documents, webpages, timelines and reflective writing. The finished artefacts mattered, but they were not the only evidence. The folder showed how the work developed.

That is the principle universities need to take seriously. In an AI-rich environment, students should not only submit the thing. They should be able to show how the thing came to be.

Policy is not enough

The local Cardiff Met AI materials reviewed for this article show a familiar institutional pattern. A university position statement and AI principles can set a constructive tone: responsible use, transparency, ethics, academic integrity, critical evaluation, attribution and AI literacy for staff and students. That matters.

But a position statement does not automatically change practice.

The Cardiff Met stance review makes this distinction clearly. The principles establish a stronger and more structured framework for responsible AI use, but they still need operational translation: clear implementation phases, practical use categories, programme-level expectations, module-level assessment design, templates, guidance sheets and ownership.

This is the key lesson for higher education more broadly. AI policy has to move through the institution.

Institutional guidance must become programme design. Programme design must become module practice. Module practice must become assessment briefs, classroom activities, student guidance and marking criteria.

This is consistent with Sarumi's emphasis on institutional readiness, faculty adaptation and robust policy, and with Slimi and Villarejo-Carballido's expert-perspective study, which treats AI implementation as affecting administration, teaching, learning, assessment, integrity and ethics rather than one isolated domain. It is also consistent with Raymond's argument that programme-level integration is often the missing layer between broad policy and classroom use.

Three connected levels of action

The strongest local framework is the three-level model developed in the AI action documents:

Institutional level = set, define, revise, audit, monitor, protect
Programme level = translate, sequence, coordinate, check, evaluate, preserve
Module level = state, teach, design, scaffold, reflect, embed

This is useful because it avoids the mistake of asking individual lecturers to solve the whole AI problem alone.

Institutional level: set the conditions

Institutions need to decide what AI is for.

The first question should not be: which tools should we buy? It should be: what educational purposes should AI support, and what must it not undermine?

At institutional level, universities should:

• establish a clear AI position linked to learning, assessment, research, operations and graduate capability;
• define acceptable, restricted and prohibited uses;
• govern privacy, data protection, bias, procurement, transparency and accountability;
• revise assessment policy away from detection-first assumptions and towards validity in an AI-rich environment;
• define baseline AI literacy expectations for staff and students;
• audit equity, access, disability implications and uneven digital confidence;
• evaluate administrative AI against educational value, not novelty alone.

The institutional role is not to write every module brief. It is to set the conditions in which responsible practice can happen.

This institutional layer is supported most directly by Sarumi (n.d.), Ahmed (2024), Slimi and Villarejo-Carballido (2024), and Xiao and Lim (2026): AI adoption is not only a teaching choice but a governance, ethics, equity and institutional-readiness issue.

Programme level: create coherence

Programme teams then need to translate broad policy into disciplinary practice.

This is where many AI strategies will succeed or fail. Students experience a programme, not an institution-wide policy. If one module encourages AI, another bans it, a third ignores it and a fourth assumes everyone knows how to disclose it, students receive mixed messages.

At programme level, teams should:

• map where AI is relevant, useful or risky across the programme;
• sequence AI literacy development across levels of study;
• coordinate the assessment diet so it is not over-reliant on product-only tasks;
• define what responsible AI use looks like in the discipline;
• align AI use with graduate attributes, employability and professional standards;
• check whether expectations are fair and realistic for the cohort;
• review student experience and assessment patterns regularly.

The programme role is coherence. It turns institutional principles into a learning journey.

This is where Raymond's AIDE Framework is especially useful. Its sequence of Awareness, Integration, Development and Evaluation gives programme teams a way to move from general ethical awareness to scaffolded capability and review. The local programme-level action document reaches the same conclusion in different language: AI literacy should be sequenced, assessment should be coordinated across the programme, and institutional policy must be translated into disciplinary expectations.

Module level: make it real

At module level, AI guidance has to become concrete.

Students should not have to guess whether AI use is encouraged, allowed with disclosure, restricted or prohibited. Assessment briefs should say what is permitted, what is not, whether disclosure is required and what kind of process evidence is expected.

Module leaders can:

• clarify AI expectations in every module;
• embed AI-use guidance directly into assessment briefs;
• design tasks where AI supports inquiry, drafting, feedback, comparison or critique rather than replacing thinking;
• collect process evidence where the final product could plausibly be AI-generated;
• teach AI literacy in discipline-specific ways;
• preserve tasks that require interpretation, explanation, defence and judgement;
• use low-stakes activities to build confidence and fairness;
• review practice after each delivery cycle.

The module role is enactment. It is where students learn what responsible AI use looks like in practice.

This module-level emphasis is supported by Dhupal's claim that AI integration needs careful design and teamwork, Govil's warning that AI's value depends on how it is used, and Weimann-Sandig's argument that educators' roles change rather than disappear. The lecturer's role becomes less about pretending AI is absent and more about designing learning situations in which students practise judgement.

Assessment should be the first operational lever

If a university changes only one thing first, it should review assessment.

Assessment is where the AI issue becomes unavoidable. It touches academic integrity, student confidence, staff workload, learning design, employability and quality assurance. It is also where abstract policy becomes real for students.

The local implementation pathway puts this well: the first meaningful operational lever is programme-level assessment review, because that is where institutional guidance becomes curricular reality.

That review should not ask only, "Can AI complete this assessment?" It should ask:

• What learning is this assessment supposed to evidence?
• Is the final product enough evidence of that learning?
• Where should process evidence be included?
• What AI use is educationally valuable here?
• What AI use would undermine the task?
• How will students know the difference?
• What does responsible disclosure look like?
• How will staff judge the quality of human thinking?

This is not about making every assessment AI-proof. Some tasks may still need controlled conditions. Others may deliberately include AI. The point is to make the assessment decision explicit.

The supplied papers support this balanced position. Murgatroyd (2024) cautions against overblown transformation claims and suggests incremental adoption is more likely. Xiao and Lim (2026) argue for selective use rather than technology-led reform. Raymond (2026) argues for structured programme integration rather than one-off AI activities. Together, they point towards assessment review as a design task, not a panic response.

Detection cannot carry the educational burden

Detection may have a limited role in academic integrity processes, but it cannot be the centre of an AI strategy.

It is reactive. It tends to create suspicion. It does not teach students how to use AI responsibly. It can also distract from the more important design question: why was the assessment so vulnerable to outsourcing in the first place?

An AI-era assessment strategy should be design-first, not detection-first.

That does not mean ignoring misconduct. It means building tasks, briefs and learning activities that make human thinking more visible and harder to fake. Drafts, rationales, oral explanations, annotated prompts, source checks, decision logs and reflective authorship statements can all help. None is perfect. Together, they shift attention from policing the tool to evidencing the learning.

AI literacy should be documented judgement

The phrase AI literacy is often used loosely. Sometimes it means knowing how to write better prompts. Sometimes it means understanding what large language models are. Sometimes it means knowing the university rules.

All of that may be useful, but it is not enough.

For higher education, AI literacy should include documented judgement:

• knowing when AI is appropriate and when it is not;
• using AI to support learning rather than replace it;
• checking claims, sources and outputs;
• recognising bias, fabrication and overconfidence;
• understanding privacy and data boundaries;
• explaining how AI shaped the work;
• taking responsibility for the final submission.

This connects directly to graduate capability. Employers may value graduates who can use AI tools. But they will need more than tool familiarity. They will need people who can judge outputs, verify information, communicate uncertainty, protect sensitive data and decide when human expertise matters.

That is the deeper meaning of AI literacy in the supplied literature. Dhupal (2025) links AI to digital literacy, coding and computational thinking. Weimann-Sandig (2024) links AI to future skills and the changing role of educators. Raymond (2026) links AI capability to ethics, judgement and critical reasoning. The argument here extends those points: students should not merely learn to use AI; they should learn to evidence responsible use.

Conclusion: from detection to responsibility

Universities should stop asking only whether students used AI.

They should ask whether students can explain how they used judgement.

The future of AI in higher education is not avoidance, detection or prompt engineering alone. It is documented judgement: the ability to show how human thinking, ethical responsibility and disciplinary standards shaped AI-assisted work.

That means building AI literacy across three connected levels. Institutions set the conditions. Programmes create coherence. Modules make responsible use real in practice.

In an AI-rich world, the final product is no longer enough evidence of learning. Students need to evidence the thinking inside it.

Source Map

The argument above draws on the supplied AI in HE folder as follows:

• AI potential and learning support: Dhupal (2025), Bholane (2025), Slimi and Villarejo-Carballido (2024), Sarumi (n.d.).
• Caution against hype and uncontrolled adoption: Murgatroyd (2024), Ahmed (2024), Xiao and Lim (2026), Govil (2025).
• Programme-level integration and AI literacy development: Raymond (2026), Weimann-Sandig (2024), the local programme-level and connected-level action documents.
• Assessment redesign and process evidence: Grzybowski (2025), Raymond (2026), the local module-level action document and implementation pathway.
• Governance, ethics, privacy, bias and equity: Ahmed (2024), Sarumi (n.d.), Slimi and Villarejo-Carballido (2024), Xiao and Lim (2026), Cardiff Met AI stance review and Cardiff Met AI guidance.
• Institutional implementation: Cardiff Met stance review, institutional-level action document, implementation pathway, connected-level framework and AI strategy spreadsheet.

References

Ahmed, S.A. (2024) 'The Politics of Integrating Artificial Intelligence into Higher Education: Benefits <> Risks', Occasional Papers, 88.

Bholane, K.P. (2025) 'Artificial Intelligence Revolutionizing Higher Education: Through the Viewpoints of Faculties, Students and Parents'. Paper presented at International Conference on AI in Education: Navigating Challenges and Embracing Opportunities, 19 January 2025.

Dhupal, S. (2025) 'AI in Higher Education: Shaping the Future of Teaching & Learning Process', Scholarly Research Journal for Interdisciplinary Studies, 14(90). doi: 10.5281/zenodo.17045571.

Govil, R. (2025) 'Higher Education in the Age of Artificial Intelligence', Economic & Political Weekly, 60(45), pp. 35-39.

Grzybowski, J. (2025) 'Artificial Intelligence (AI) and the University: How AI Will Change Our Approach to Education', Studia Gilsoniana, 14(2), pp. 297-333. doi: 10.26385/SG.140211.

Murgatroyd, S. (2024) 'Artificial Intelligence and future of higher education', Revista Paraguaya de Educación a Distancia, 5(1), pp. 4-11. doi: 10.56152/reped2024-vol5num1-art1.

Raymond, C.J. (2026) The Higher Education AIDE Framework: Embedding Responsible AI into University Curricula. White paper.

Sarumi, O.O. (n.d.) 'Reimagining Higher Education with AI: Pedagogical, Ethical, and Institutional Perspectives'.

Slimi, Z. and Villarejo-Carballido, B. (2024) 'Unveiling the Potential: Experts' Perspectives on Artificial Intelligence Integration in Higher Education', European Journal of Educational Research, 13(4), pp. 1477-1492. doi: 10.12973/eu-jer.13.4.1477.

Weimann-Sandig, N. (2024) 'How the use of AI is changing the role of educators at universities - and why this is by no means a bad thing', INTED2024 Proceedings, pp. 394-400.

Xiao, J. and Lim, D.C.L. (2026) 'Is AI the solution to the problems that make higher education "ill" in the first place? Towards a technology-agnostic, future-proof approach', Journal of Applied Learning & Teaching, 9(1). doi: 10.37074/jalt.2026.9.1.14.