Interlude
Theory: The Moment We Are In | Template: The Contemporary Essay | Words: ~1,810
# Why This Moment in AI Is Different
Every decade since the 1960s has had its "this changes everything" moment in educational technology. Programmed instruction was going to change everything. Multimedia CD-ROMs were going to change everything. The LMS was going to change everything. MOOCs were going to change everything. Each wave left real improvements behind, and each wave failed to deliver on the maximal version of its promise. It is reasonable, on that track record, to be sceptical when a new wave arrives.
This act argues, cautiously, that the current wave is different. Not because the technology is smarter than its predecessors, but because the structural position it occupies in an institution is new.
What We Used to Believe
The standard EdTech narrative treated each new technology as a better delivery channel for the existing model. Video delivered lectures. LMS delivered courses. MOOCs delivered courses at scale. The model underneath — a human expert designs a curriculum, learners move through it, assessment at the end — stayed intact. The technology changed the packaging.
Generative AI breaks that pattern in a specific way. For the first time, the technology can participate in the act of instruction itself. It can answer a question in natural language. It can produce an example tuned to a learner's context. It can mark an open-ended response. It can hold something that looks like a conversation. Whether it does any of these things well on a given day is a separate question. The point is that the role it is stepping into is not delivery. It is instruction.
The Research That Changed Everything
Three empirical shifts in the last three years matter.
First, scale of capability has outpaced the governance conversation. Systems that could not pass an undergraduate exam in 2021 now pass professional licensing exams with room to spare (OpenAI, 2023; Kung et al., 2023). The curve is not flattening the way earlier technology curves flattened.
Second, learner behaviour has already changed, regardless of what institutions have decided. Surveys of undergraduate and secondary learners consistently show majority use of generative AI for homework, revision, and writing (Tyton Partners, 2024; HEPI, 2024). Whatever policy exists on paper, the population has moved. Institutions that design as if this were not true are designing for a cohort that no longer exists.
Third, governance research has begun to catch up, with frameworks emerging from UNESCO (2023), the European Union AI Act (2024), and a growing professional literature on AI in assessment (Bearman et al., 2023; Lodge et al., 2023). The shape of a credible regime is becoming visible, but the gap between the best guidance and average institutional capacity is wide.
The combination — capability moving fast, behaviour already changed, governance partially formed — produces the situation Act III documents. Adaptive systems are being deployed into institutions that do not yet have the professional, ethical, or architectural infrastructure to govern them. The failures that result are not algorithmic. They are institutional.
The Framework
Act III walks through those failures in sequence: recommender systems without pedagogical framing, data-driven personalisation without student agency, automated assessment without construct validity, governance frameworks that cover inputs but not outputs, and the accumulating evidence that adaptive platforms, on average, are under-delivering on their stated outcomes (Major et al., 2021; Escueta et al., 2020).
The act is not an argument against AI in education. It is an argument that the current generation of products has inherited the content-delivery equation from the era before it, and is now trying to deliver content faster and more personally without having reconsidered whether faster and more personally is the right goal.
The theories in this act — including the APLF framework that closes the series — are attempts to describe what a governable, auditable, whole-person adaptive system would actually look like, and what it would cost to build one.
The Invitation
If you are designing, procuring, or regulating any form of AI in education right now, the practical question is not "is this a good product." It is "is this a product whose failure we could actually detect, explain, and correct." Most current systems do not pass that test. The act that follows is, in part, a map for what it would take to change that.
If an AI system quietly failed a cohort of learners in your institution this term, how would you know — and who would be accountable?