Modality does not equal risk
Liz Smith, Dawn Gilmore, Chinh Nguyen, Paul Wiseman, Caitlin Yolland, and Jim Frederickson
In the rush to respond to generative AI, universities risk reviving an old and damaging assumption: that online learning is inherently lower in quality than on-campus offerings.
It isn’t, and framing it that way risks undoing some important advances higher education has made in the equitable provision of quality study opportunities.
Nearly one in five students now study online with online learning now the preferred study mode of postgraduate students in Australia, not because it is easier, but because it works. In the postgraduate space, it provides working professionals with a flexible mode of study that fits into their lives and connects learning to their day-to-day work. In short, online learning provides equitable access and authentic learning for a growing cohort for whom on-campus study simply doesn’t work.
Yet online learning continues to fight for legitimacy alongside on-campus delivery. As Tim Fawns suggested recently, this is problematic. The Higher Education Standards Framework was deliberately designed to be modality-neutral. Quality is not determined by where learning happens, but by how well it is designed, how effectively it is implemented, and how convincingly it demonstrates learning.
This is where the current debate has lost its way.
This was on full display in the TEQSA webinar on credible assurance earlier this month, where nearly 2,000 of us dialled in hoping for answers. What emerged most clearly was not a solution, but a confirmation of the scale of the challenge, and that no one has solved it yet. There is still a significant gap between high-level discussion about assurance of learning and assessment security and what is actually feasible in everyday teaching practice, regardless of mode.
The problem has never been online assessment. The problem is assessment design.
That said, the immediate integrity problem cannot be ignored. There is an acute risk now: in some assessment contexts, it is increasingly difficult to assure that submitted work is a student’s own or that course learning outcomes have been met. This matters for students, universities and employers, and for HESF 1.4.4. The sector cannot simply wait for programmatic assessment to mature before acting.
But the acute problem should not be mistaken for the whole problem. Short-term controls may be needed to stem the bleeding, but they should support, not replace, the longer-term redesign of assessment so that assurance is cumulative, layered and educationally meaningful.
Generative AI has made these problems harder to ignore: the acute risk to academic integrity and the chronic weakness of assessment systems that rely too heavily on isolated tasks, insufficient evidence of learning, and the illusion that tighter controls alone can secure learning.
The sector broadly agrees that online invigilation is an unwinnable arms race. Controlled environments still have a place where students must demonstrate safety-critical knowledge or real-time judgement independently, such as making accurate dosage calculations in clinical practice. However, treating them as the default response to AI risk constrains both what we assess and how students can demonstrate their knowledge and capabilities.
What is encouraging is that this moment has prompted the sector to enter a broader conversation about program-based assessment. A clear, cumulative path of assurance across a program, rather than a collection of disconnected, one-off “secure” tasks, is both more defensible and more educationally sound.
What we can’t ignore is that credible assurance of learning and secure assessment require both purposeful design and skilled facilitation, supported by adequate resources. While online education offers scale, it has never been the cheaper option, particularly when the goal is to produce valid, trustworthy evidence of learning.
Where to next? Don’t try to solve the AI problem by giving up educational advances
The higher education sector is still working out how best to respond to generative AI. Therefore, it would be a serious mistake to respond to this threat by abandoning hard-won advances in online assessment design, authenticity and flexibility. This is not an argument against standards, quality, or the assurance of learning. But if the price of “assurance” is dragging online education back to forms of assessment we already know to be limited, exclusionary or pedagogically weak, the cost is too high.
The return to in-person exam centres is one such example in the current discussion. The greater confidence this may offer in verifying student identity comes at a cost. It reduces access for the very learners online education is designed to support, and it pushes assessment away from authentic, practice-based demonstrations of learning. This is not a neutral shift; it narrows who can participate and what kinds of learning are valued. And it does so without solving the underlying issue.
Academic misconduct did not begin with AI, and it will not be solved by physical presence alone. No system, online or on campus, has ever been entirely secure. The goal is not to eliminate risk, but to design for validity and to build assessment systems that generate trustworthy evidence of learning despite that risk.
Programmatic, layered and authenticity-focused approaches are not compromises. They are exactly the kinds of approaches that contemporary assessment research and emerging AI frameworks are pointing towards.
There is, of course, a familiar refrain here. When plagiarism and essay mills dominated the conversation, many of us argued for a shift away from detection and towards authentic assessment. By contrast, in the world of AI, we are now seeing a push back towards detection. But this is not the same problem. AI is more prevalent, more pervasive, and increasingly embedded in the workplaces our graduates are entering.
Data from an Australian multi-university survey found that 83% of university students are using AI in their studies, with 44% using generative AI daily. This intensifies the challenge and underscores the need for solutions that lean even further into authentic assessment.
Banning AI is not the answer when it is already woven into both learning and work. Our task is not to get better at detecting cheating, but at detecting learning. That means designing assessment on the assumption that students will use AI, while requiring forms of evaluative judgement that go beyond effective prompting. Students must be accountable not only for the final product, but for the chain of micro-decisions that produced it. And if we are serious about assessing that, we need better ways to make the learning process visible.
We still need to rethink assessment regardless of mode. To balance assurance of learning with the need to graduate students who can thrive in a workforce where AI is among the fastest-growing skills, and to design authentic tasks that use controls only where they are genuinely needed.
Many Australian higher education institutions have revised assessment policies in recognition of this shift. The move away from course-by-course responses and towards program-based assurance of learning is positive, mode-neutral progress. It is a solution that resists pulling the sector backwards, towards narrower, more rigid forms of assessment that are easier to control but harder to justify.
The task ahead is not to make online learning look more like the past, or even like progressive on-campus equivalents. It is to ensure that all learning, regardless of mode, is designed for the realities of the present.
If AI has exposed a weakness, it is not online education, but our continued reliance on thin, disconnected forms of assessment across all modes of delivery.
As Phillip Dawson discussed, the ‘Swiss Cheese’ approach to assessment design is elegantly simple, but powerful. It begins with the recognition that every assessment task has holes and that no single task is “AI-proof”. Once we let go of the search for perfect ‘solutions’ and instead layer a sequence of rich, imperfect, but meaningful tasks over time, including carefully placed secure points, we create an assurance model that is defensible, trustworthy and, crucially, aligned to real evidence of learning rather than the illusion of control.
If we respond to AI by privileging surveillance over design, or physical presence over validity, we risk solving the wrong problem, and in doing so, diminish both access and quality of education to many.
Modality does not equal risk.
Poor design does.
Liz Smith is the Director of the Learner Experience in Lifelong Learning at the University of New South Wales
Dawn Gilmore is Director, Academic Products & Student Experience, Melbourne Online, Office of the Provost at the University of Melbourne
Chinh Nguyen is Senior Advisor, Student Experience, Melbourne Online, Office of the Provost at the University of Melbourne
Paul Wiseman leads innovation and quality in teaching at Monash Online, Monash University
Caitlin Yolland is Associate Dean, Education, Behavioural and Social Sciences, Adelaide University Online
James R. Frederickson is the Associate Dean, Online MBA at Melbourne Business School at the University of Melbourne