A university lecturer receives two excellent reports. Both are well structured, clearly written and supported by appropriate arguments. One reflects weeks of reading, analysis and revision by the student. The other was produced largely through a sequence of prompts to a generative AI system.
If the final documents look equally convincing, the problem is no longer simply whether students are using artificial intelligence. The more important question is whether the assessment still provides credible evidence of learning.
This is becoming one of the defining challenges for higher education in Asia. Universities cannot reasonably design every assignment around the assumption that generative AI will disappear. Nor is it sufficient to replace educational judgement with a search for AI-generated language.
The task is harder but more useful: redesign assessment so that students have to make their thinking, decisions and application of knowledge visible.
AI Has Changed What a Finished Assignment Can Prove
For decades, many university assessments have treated the submitted product as evidence of the learning that produced it. An essay demonstrated analysis. A report demonstrated research. A computer program demonstrated programming. A presentation demonstrated preparation and understanding.
Generative AI weakens that assumption because sophisticated outputs can increasingly be produced with substantial machine assistance.
The OECD Digital Education Outlook 2026 makes an important distinction. Generative AI can improve performance on a task without necessarily producing equivalent learning gains. Used with clear pedagogical intent, however, it can also support learning, feedback and skill development.
That distinction should change the assessment question.
Instead of asking only, “Is this submission good?”, lecturers increasingly need to ask, “What evidence shows that the student can personally explain, apply, evaluate and defend the thinking represented in this submission?”
This does not mean returning every course to closed-book examinations. It means widening the evidence used to judge learning.
Detection Is Not an Assessment Strategy
Universities need academic integrity procedures, and deliberate misrepresentation still requires appropriate institutional responses. But AI detection cannot carry the entire assessment system.
The University of Hong Kong’s 2026 work on assessment redesign describes reliance on detection technologies as potentially insufficient and unsustainable, placing greater emphasis on authentic application, higher-order thinking, transparent AI use and making learning processes visible.
This is a more durable direction because assessment design remains useful even as technology changes.
A detection-centred model asks whether a machine probably contributed to a document. A learning-centred model asks whether the student has demonstrated the intended capability.
The second question is the one a university ultimately needs to answer.
The Learning Evidence Stack
Universities can make AI assessment in higher education more robust by collecting several layers of evidence rather than relying on one final product.
| Evidence layer | What it reveals | Possible assessment method |
|---|---|---|
| Product | What the student ultimately produced | Report, prototype, analysis, presentation or code |
| Process | How the work developed | Drafts, design notes, research log or version history |
| Decisions | Why particular choices were made | Commentary, design rationale or methodological memo |
| Application | Whether knowledge transfers to a new situation | Case variation, practical task or unseen problem |
| Defence | Whether the student can explain and challenge the work | Short viva, demonstration, questioning or presentation |
| Reflection | What the student understands about strengths, weaknesses and AI use | Reflective statement or AI-use disclosure |
No course needs all six layers in every assignment. The principle is redundancy. When several forms of evidence point to the same capability, confidence in the assessment becomes stronger.
A polished report may be ambiguous. A polished report combined with documented decisions, an altered case problem and a five-minute defence tells the lecturer considerably more.
Redesign the Task Before Adding More Policing
Some assignments are especially easy to outsource because the task mainly asks students to reproduce information that is widely available.
Consider how a conventional assignment can be changed.
| Conventional task | Stronger redesign |
|---|---|
| Write 2,000 words explaining a management theory | Apply two competing theories to a local organisational case and defend which explains the outcome better |
| Produce a market analysis | Analyse a specified dataset, document assumptions and respond to a changed scenario during class |
| Write a programming solution | Submit the program, demonstrate it live and explain two design decisions and one limitation |
| Summarise academic literature | Compare conflicting findings, identify an unresolved question and justify a proposed research approach |
| Create a business proposal | Create the proposal, then defend financial and operational assumptions under questioning |
The objective is not to create assignments that AI cannot touch. That target is likely to become increasingly unrealistic.
The objective is to design assignments in which successful completion requires capabilities the course genuinely intends to develop.
Make the Learning Process Visible
One useful change is to assess selected points in the journey rather than only the destination.
A research assignment might include a research question, source-selection rationale, preliminary interpretation, final analysis and short oral defence. A design project might preserve sketches, rejected approaches and testing decisions. A software module might assess requirement interpretation and debugging alongside the final application.
This does not require lecturers to grade every draft.
Some process evidence can simply function as verification. Students might submit a short decision log identifying three important choices, what alternatives they considered and why they rejected them.
This creates something generative AI makes unusually valuable: a reason to assess judgement rather than presentation quality alone.
Use AI Disclosure to Improve Assessment, Not Merely Police It
Where AI use is permitted, students should know what must be disclosed.
A useful disclosure can answer four questions:
- Which AI tools were used?
- What were they used for?
- Which parts of the final work remain substantially the student’s own judgement?
- How were AI outputs checked, changed or rejected?
This turns disclosure into educational evidence.
A student who can identify an incorrect AI recommendation, explain why it failed and show the stronger alternative demonstrates critical judgement. That may reveal more about learning than pretending the technology was never present.
UNESCO’s guidance on generative AI similarly argues that education systems need to reconsider learning outcomes and assessment as human and machine capabilities increasingly overlap.
Preserve Some Independent Performance
Universities should also identify which capabilities students must demonstrate without substantial external assistance.
A nursing student cannot outsource every clinical judgement. An engineering graduate needs some independent quantitative reasoning. A programmer needs to understand code well enough to diagnose failure. A journalism student needs to verify information. A finance graduate needs to recognise when an apparently plausible calculation is wrong.
The appropriate balance will differ by discipline.
Independent performance might be demonstrated through laboratories, practical demonstrations, supervised problem solving, oral questioning, simulations or short controlled tasks rather than relying only on traditional examinations.
The key is to decide deliberately which capabilities may be AI-assisted and which must remain independently demonstrable.
Do Not Confuse AI Resistance With Good Assessment
An assessment can be difficult for AI and still be educationally poor.
Adding obscure instructions, requiring unusual formatting or creating artificial complexity may make automated completion harder without improving learning. Assessment redesign should begin with the intended outcome, not with techniques for defeating software.
A useful test is:
- Capability: What should the student be able to do?
- Evidence: What would convince us they can actually do it?
- Assistance: Which forms of help are acceptable while producing that evidence?
- Verification: Where do we need independent confirmation?
- Transfer: Can the student perform when the problem changes?
If those questions are answered clearly, the AI policy becomes easier to write because it follows the educational purpose rather than attempting to define every possible technology.
Assessment Redesign Is an Institutional Project
Individual lecturers can improve individual assignments, but inconsistent rules across a university create a different problem.
One lecturer may prohibit AI completely. Another may encourage it. A third may permit brainstorming but not drafting. Students moving between these courses can easily misunderstand the boundaries unless expectations are explicit.
Research involving lecturers at Malaysian private universities has similarly highlighted the importance of clearer institutional policies, professional development and alignment between technology, pedagogy and learning outcomes.
Universities therefore need several layers of coordination:
- a university-level statement describing responsible AI use;
- programme-level agreement about capabilities graduates must demonstrate independently;
- course-level rules describing permitted assistance;
- assessment-level instructions explaining disclosure requirements;
- faculty development focused on redesign rather than software policing.
This also connects to technology procurement. As Asia Education Journal has previously examined in its guide to evaluating EdTech evidence, institutions should distinguish vendor claims from evidence of actual educational impact.
Measure Whether the Redesign Works
Universities should avoid declaring an assessment innovation successful merely because lecturers implemented it.
Evaluate the redesign itself.
Possible evidence includes whether students can explain their work more effectively, whether performance transfers to unfamiliar problems, whether academic-integrity disputes change, whether marking becomes more consistent, whether students understand AI-use rules and whether workload remains sustainable for teaching staff.
Graduate capability also matters. Assessment should ultimately connect with what students need beyond university. The journal’s analysis of graduate employability in Asia makes a related point: one headline outcome rarely captures the full capability of a graduate or institution.
When Assessment Innovation Becomes an Institutional Achievement
Occasionally, an assessment initiative develops beyond ordinary curriculum improvement into a measurable institutional milestone. A university might implement a documented assessment model across an unusually large number of programmes, achieve a clearly defined participation milestone or establish another distinctive result that can be independently verified.
That is separate from deciding whether individual students have learned.
The Asia Record official website describes recognition around clearly defined and verifiable achievements. Institutions considering an Asia Record application should therefore separate the specific measurable achievement from broader claims about educational quality. Becoming an Asia Record holder documents an approved achievement; it does not replace programme accreditation, regulatory approval or academic quality assurance.
Institutions asking how to get an Asia Record, or whether they should apply for Asia Record recognition, should first establish precisely what was achieved, how it was measured and what evidence allows an outside party to verify the claim. That distinction matters for credible record recognition in Asia and for any form of record certification in Asia.
Universities Need Better Evidence of Learning, Not Just Better Student Outputs
Generative AI has exposed a weakness that already existed in some forms of assessment: universities have sometimes treated the finished product as though it were identical to the learning behind it.
Those two things can no longer safely be assumed to be the same.
The strongest response is not to make every assessment AI-free. It is to make learning harder to hide.
Ask students to show their decisions. Give them unfamiliar problems. Require them to explain assumptions. Test whether knowledge transfers. Use practical demonstrations where appropriate. Make permitted AI assistance explicit. Preserve independent performance where professional capability requires it.
Universities that do this will be better positioned not merely to control AI use, but to answer the question assessment was always supposed to answer: what can this student actually understand, judge and do?