I’m quite proud of this section of The AI Crisis in Higher Education, largely because it’s an intentional sketch of a wider approach to the political economy of (digital) pedagogy which I’ve been gradually stumbling towards over recent years:
The bigger problem with Pratschke’s (2024) advocacy lies at the level of political economy. If pedagogy becomes embedded in interaction with the machine, what happens when that machine is no longer as accessible, engaging or helpful as it currently is? It takes an early phase in the rollout of a technology and advocates rebuilding education around an accurate and plausible account of its affordances, without considering how those affordances may change over time or be supplemented by constraints that reshape what we can do with it. In doing so, it risks reproducing the very dynamic that Pratschke, rightly, warns against. While she’s right to warn against LLMs being used “to revert to outdated models of delivery”, she fails to consider how her own approach creates the groundwork for “increasing automation in education” (Pratschke 2024: loc 2231). If we see AI integration as the insertion of LLMs into every aspect of teaching and learning, we create the conditions in which instructors come to appear as costly irrelevances within an otherwise machinic pedagogy.
This does not mean they will be dispensed with immediately or everywhere. Rather, the human role is likely to be gradually eroded within a transformed political economy. By definition, human-in-the-loop instruction will tend to be cheaper than human-centric instruction. This creates the conditions in which pedagogical arguments for human-in-the-loop instruction blur into economic arguments for it. If “outdated models of delivery” are understood as the result of insufficiently enlightened learning design, it becomes easy to present more technologically intensive approaches as the obvious solution, justified in pedagogical terms but driven by economic incentives. Pratschke (2024), to her credit, points to exactly these concerns but does so in a way that leaves the causal relationship underexplored. The risk of “increasing automation in education” follows logically from generativism rather than being a failure of it. We shouldn’t dismiss Pratschke’s pedagogical argument on this basis, but we should qualify endorsement of it by considering what the institutional implications would be if this agenda were pursued in full.
The problem is that models of delivery cannot be understood in isolation from their political economy. What Weller (2020) calls the “infinite lecture hall” model of MOOCs was pushed so forcefully in the early 2010s because it appeared to free prestigious universities from the economics of scarcity. If one Harvard professor could teach the world’s philosophy students, then what was effectively a luxury good could be distributed globally. MOOCs did offer real benefits, most obviously flexibility for lifelong learners, but to explain their rise and fall in purely pedagogical terms is to ignore the commercial pressures shaping them. The same is true of large language models. While learning designers may not be able to transform the context in which they work, this does not justify ignoring how that context shapes and is shaped by learning design over time. In this sense, my argument is less a critique of Pratschke (2024) than a call for approaches to pedagogy and learning design which are institutionally and economically sensitive about their own context.
This is particularly important given the economic uncertainty outlined in the previous chapter. It cannot be assumed that current systems will remain stable in terms of capability, reliability or pricing. At the same time, they cannot be ignored given their widespread use by staff and students. Configuring learning entirely around currently available systems is therefore risky. It risks deepening platform lock-in, incentivising automation as the human role is gradually eroded under resource constraints, and leaving students anchored in human-in-the-loop learning when the aim should be to “shift left”, making LLM use one component within a broader learning process. This is the only way to address the pedagogical challenges posed by existing use while supporting the development of meaningful practice. We can support students in learning to use language models creatively and carefully, without risking them coming to be dependent on commercial software with an uncertain future.
This is going to be crucial in the coming years. Assuming no financial rescue for higher education there’s going to be a rapid push towards reducing the cost per student of delivering teaching and learning. What happens when pedagogical logical and financial logic meet in a cash strapped university? How do we make claims about pedagogical value and pedagogical redlines in ways that register with the financial logic that will otherwise be dominant?
