Opinion
Higher education is responding to generative artificial intelligence with understandable concern, but also with growing overcorrection. Across the sector, more invigilated exams, tighter controls, handwritten tasks and revived oral checks are re-emerging as the preferred answer to assessment assurance. The message is clear: if AI makes unsupervised work difficult to trust, universities must return to more secure formats.
This instinct is understandable. It is also educationally risky.
For decades, higher education worked hard to move beyond examination-dominated assessment. That shift happened for good reason. Traditional exams often privilege recall under pressure over deeper understanding. They can disadvantage students with anxiety, language barriers and diverse learning needs. They rarely reflect how graduates actually work in professional settings, where complex problems are solved through consultation, evidence use, revision, collaboration and increasingly, intelligent tools.
Yet in the space of two years, the arrival of generative AI has pushed many institutions back toward the very methods assessment reform was trying to move beyond. Not because universities suddenly discovered that exams are better for learning, but because they appear harder to game. Assurance is being driven by fear of AI rather than by a clear theory of how students learn.
That is the mistake.
The real challenge is not how to recreate a pre-AI classroom. It is how to ensure that students can still demonstrate judgment, ownership and competence in an AI-rich environment. Graduates are not heading into workplaces where artificial intelligence will be absent. They are heading into workplaces where they will be expected to use it, interrogate it, challenge it and remain accountable for work produced with its assistance.
This changes what higher education should be verifying.
A student who can produce polished text with AI has not demonstrated much. A student who can identify where AI is wrong, test its claims against authoritative sources, revise its output, and then justify those changes under supervision has demonstrated something far more important. The competency at stake is evaluative judgment: the ability to decide whether an AI-assisted output is accurate, appropriate and fit for purpose in a disciplinary or professional context.
That is why the answer cannot be a blanket return to exams.
A more credible alternative is distributed assurance across a program. Instead of treating every unit as though it needs the same secure assessment format, units can be designed with context-appropriate assurance tasks that reflect disciplinary needs. In one subject, that might be a viva. In another, a code walkthrough. Elsewhere, a live demonstration, presentation, practical defence or timed application task may make more sense. The point is not that one format is universally superior. The point is that students must, at meaningful points across a course, demonstrate under supervised conditions that they can explain, justify and apply their reasoning.
This is precisely where practical pedagogy matters. Most universities now have some form of AI policy telling students whether they may use generative AI in an assessment. But policy only sets permission boundaries. It does not show students how to use AI responsibly, nor does it show educators how to design assessment that makes student reasoning visible.
That gap is what the SAGE framework was designed to address. Developed through empirical work at CQUniversity, SAGE structures AI-assisted assessment into six steps: Generate, Evaluate, Refine, AI Critic, Reflect and Defend. Students first work with AI openly, but not passively. They must evaluate the AI response against real sources, revise it, critique its limitations and reflect on its strengths and risks. Only then do they reach the final Defend step: a short supervised checkpoint where they must demonstrate ownership of the reasoning behind the work. To support adoption, a freely available Defend planning tool (sage-framework.com/defend_tool) accompanies the framework, enabling educators to design supervised checkpoints suited to disciplinary context, class scale and learning outcomes. The framework is presented not as a policy but as a pedagogy, and its evidence base currently spans more than 800 students across five Australian campuses and eight peer-reviewed studies.
The significance of this is broader than one framework. The key principle is that assurance should be designed, not improvised through a return to old habits. Universities do need secure moments of verification. But that does not mean rebuilding an entire assessment system around closed-book exams. It means integrating fit-for-purpose assurance points across a degree in ways that align with disciplinary learning outcomes.
If generative AI has exposed anything, it is not simply that students can cheat. It is that many existing assessments were already too easy to complete without demonstrating real understanding. The answer is not to retreat to the past. It is to build assessment systems that reflect the conditions students will actually face after graduation.
The future of assurance in higher education should not be a blanket return to the exam hall. It should be a more intelligent model in which students are allowed to work with AI, but are also required, at the right moments and in the right ways, to prove that the thinking is genuinely theirs.
Dr Mahmoud Elkhodr is a Senior Lecturer in ICT at CQUniversity Australia whose work has contributed to the development and empirical validation of the Structured AI-Guided Education (SAGE) framework.