Abstract
Designing an undergraduate systems thinking course in 2023 took months. Workshops, whiteboards, an academic drafting offline, iteration, constructive alignment, compliance, and endless box-moving on a digital canvas. One decision shaped the course: the learning designer judged 10 assessment submission points across 12 weeks excessive and argued to reduce them. The academic, who had taught the discipline for years, refused. Systems thinking concepts interlock, and a student who has not grasped one cannot proceed. The checkpoints were not overload; they were scaffolding. The design bent to the discipline, and the disciplinary expert was right.
Generative AI could now compress much of that. It can propose a structure, recommend staggered low-stakes checkpoints for interdependent concepts, and draft the content beneath them. It might well have reached that assessment design faster. Learning designers are tempted to insist something irreplaceable is lost. Much of it is not: a considerable portion of that time was drag, and automating drag is a good outcome. The argument here is not that slowness was valuable.
What mattered was less visible. Working on unfamiliar material, the learning designer frequently does not understand the subject. This is usually treated as a limitation. It is better understood as an instrument. The designer is an authentic novice, and their confusion is diagnostic: when an explanation fails for them, it will fail for a student (Wieman, 2007). This is the logic of usability testing: a designer cannot test an interface they already know (Nielsen, 1993). A language model can usefully simulate a novice reader. What it cannot replace is genuine uncertainty: having absorbed the discipline, it fills gaps silently rather than experiencing them.
But the instrument only works under a particular social condition. To say “I do not understand this” to an academic, repeatedly, about their own subject, requires standing, which no third space role description confers (Whitchurch, 2008). It is built through months of meetings, shared frustration, and accumulated trust, and it is what made the academic willing to re-explain rather than dismiss. The re-explanation is where the design insight came from.
This is the uncomfortable possibility. The inefficiency was not valuable in itself, but the relationship formed inside it, and the relationship is the precondition for the human contribution. Compress the process into an afternoon and the drag disappears, along with the conditions in which trust is made. We have no way to tell which parts of a slow process were waste and which were connection.
This Pecha Kucha presents a real course design, concedes what AI could do faster, and identifies what it cannot replace: genuine uncertainty, professional judgement, and trust. For institutions adopting AI-supported workflows, the practical implication is to protect time for co-design and explanation, treating relationship-building as work rather than overhead. It closes with a question beneath the debates about AI and the third space workforce (Bearman et al., 2023; Huijser et al., 2024): if the relationship was produced by the inefficiency we are now eliminating, can human-centred design survive its own optimisation?
Keywords
Human centred learning design, generative AI, trust, expert blind spot, usability, third space.
