We're still asking the wrong question about assessment in the age of AI.

The usual question is 'how do we prove students are learning?' A more useful one is: what kinds of evidence are good enough for the claim a qualification makes, and who will collect, interpret, and stand behind that evidence? It's a small shift in wording, but it changes what counts as a solution.

Framed that way, the options in front of us become a lot clearer. In this context, we can allow AI use and assess whether students can deploy it critically and appropriately. We can design tasks where AI engagement doesn't invalidate the outcome. Or we can create supervised moments where students demonstrate capability without unauthorised assistance. Each of these has a place. What none of these approaches can do is cover everything a credential claims, on its own. Pretending one can is where the trouble starts, and it's the trap a lot of sector commentary keeps falling into.

What TEQSA's review surfaced

TEQSA recently examined exactly this question: what does credible assurance of learning actually require in fully online programs? It was an important conversation, and, like much of the current sector discussion, it surfaced the problem more sharply than it resolved it. If anyone in this space is selling you a single fix, that's worth treating with some caution.

I've found it more useful to think in terms of this formulation:

credible assessment = observed performance + known conditions + expert judgment + time

Each of those four elements is doing real work.

Observed performance is the 'what', the artefact or behaviour we're actually looking at.

Known conditions tell us how much weight that observation can bear (was this open book, supervised, collaborative?).

Expert judgment is the human interpretive layer that no rubric fully replaces.

And time, repeated evidence gathered across a program rather than a single high-stakes moment, is what lets a developmental story emerge.

Making learning visible along all four of those dimensions costs human time. That's not a design flaw to be engineered away; it's the actual price of assurance. You can reduce the human time in that equation, but only by accepting one of three trade-offs: lower assurance, narrower assessment, or more surveillance and automation. There is no fourth option that avoids all three. Anyone proposing one is quietly making that trade somewhere in the system, whether they've named it or not.

Not a case for locking students in exam rooms

None of this is an argument for locking students in rooms and calling the problem solved. Exams, vivas, oral defences and supervised practical demonstrations have real value for particular outcomes, foundational knowledge, professional readiness under pressure. But over-reliance on secure, exam-style tasks narrows what we can actually assess, and it fragments the learning experience into a series of disconnected checkpoints rather than a coherent story of growth.

The better direction, I think, is to secure the right things: program-level design with multiple points of evidence gathered over time, a coherent developmental progression across a degree, and shared responsibility across teaching teams rather than each course carrying the assurance burden alone.

Think of it less like a single lock on the front door and more like a building with several load-bearing walls, no one wall has to do all the work, and the structure is more resilient for it.

Where this is taking us at Charles Sturt

This is the direction Charles Sturt has been moving in practice. We've shifted from discussion to implementation: supervised assessment embedded across the institution, designed to generate genuine evidence of capability, rather than simply collecting submitted work.

That shift is also the foundation for something larger we're building course by course, a whole-of-degree picture of what our graduates actually know and can do, assembled from evidence gathered across a program rather than inferred from any single task.

There's more to share as this work develops, and I'd be glad to hear how others in the HERDSA community are navigating the same trade-offs, particularly anyone working through this in fully online contexts, where the TEQSA review suggests the hardest version of the problem is playing out right now.

So, after reading this post, how could you increase the ‘credibility of your assessment in the age of AI?

Banner image source: Jackie Cooper (20219)