Foundations for AI Educators
A course in the theory an AI educator actually works from — how language models work, where their knowledge comes from, and how people learn. Six pillars from LLM internals to the EU AI Act, plus a two-lecture research-methods module. Its through-line is epistemology as an engineering problem. In production now.
Colleagues keep asking me which AI courses are actually worth taking in 2026, now that courses on AI and vibecoding are sold from every direction. The most reliable test I have found is a simple one: ask whoever teaches the course what a language model is. The clearer and more precise the answer, the lower the chance that everything after it is noise.
It is fair to ask why this matters once ‘AI’ has become a kind of electricity — you do not need to understand voltage to boil a kettle. The reason is that any serious work with agentic systems runs straight into the engineering parts a language model is made of. Without a working picture of those parts, the core ideas stay out of reach: what context actually is in the engineering sense, and therefore how to manage it; why the model holds no memory of you at all — every inference call is new — and how the illusion of memory is assembled only from the prefix of tokens you feed back into the context window each time; and what tokenization, retrieval and RAG, evaluation, and navigating a file system really amount to.
Not seeing these abstractions is what produces the most common early problems with agents: making a process run the same way every time, shaping how an agent makes decisions, getting it to remember, and deciding what it should remember versus what it should simply know where to find. A teacher — like the producer or project manager on this kind of work — does not need every answer, but does need to understand what is being asked. Even working researchers now carry a real gap between using these models and understanding what they are.
This course is built to close that gap. It sets out what a language model is and where its knowledge comes from, how people learn, and what all of this asks of anyone who teaches with it. The through-line is epistemology treated as an engineering problem; the structure is six pillars and a companion pair of lectures on research method, each module below a single lecture.
The six pillars
The first three pillars are the core — the technology, its epistemology, and how people learn. The last three set the frame: literacy as a construct, ethics and regulation, and intellectual history.
How language models work
How language models work below the level of metaphor. The lecture builds up from tokens and embeddings to attention and the context window, then covers how a model is trained and adapted, why it hallucinates, and what calibration and scaling laws actually describe. Throughout, a behaviour is traced to a concrete part of the architecture rather than treated as a property of ‘the AI’.
The epistemology of machine knowledge
When a model’s output counts as knowledge rather than a confident guess. The lecture separates fluency from truth and treats a claim as knowledge only when it carries a verifiable source and an exact location. It reads model output through the philosophy of testimony, and uses Frankfurt and Popper to mark the line between justified belief and plausible noise.
Cognitive and learning science
How people learn, and what that requires of any tool placed between a student and a subject. The lecture covers cognitive load, how understanding is built rather than transmitted, scaffolding that has to fade, and difficulties that aid learning rather than hinder it — leading to the central question of which cognitive operations a learner may hand to a model and which must stay with them.
AI literacy as a construct
AI literacy as a measurable set of competencies rather than fluency with one interface. The lecture works from the established definitions and competency frameworks, names over-reliance and automation bias as distinct failure modes, and argues for teaching mental models — tokenization, provenance, calibration, offloading — that outlast any particular tool.
Ethics, society, and regulation
The ethical and regulatory questions a model raises, taken one at a time. The lecture covers representational harm, the data and human labour a model is built on, copyright and privacy, the environmental cost of scale, and the effect on work — and reads the EU AI Act’s risk tiers, under which education systems often count as high-risk. It ends in a seven-criterion audit for any tool or lesson.
History and intellectual lineage
Where language models come from. The lecture sets them at the end of a seventy-year arc that begins in 1956, running through symbolic AI and connectionism, the AI winters, the statistical turn, and the path to attention and the transformer. The history is what lets an educator tell a genuine shift from hype.
The research module
Two further lectures apply the theory to the practice of research, treating the language model both as a measuring instrument and as the object being measured.
Quantitative research with LLMs
How to measure an LLM honestly. A single output is not an estimate: the lecture treats output as a random variable with real variance, covers the metrics a system is judged by, and the threats — contamination, overfitting, judge bias — that quietly invalidate a result.
Qualitative research with LLMs
How to interpret with an LLM without inventing findings. The lecture takes coding, grounded theory and trustworthiness seriously as method, and names the failure mode that appears when interpretation itself is handed to the model: themes that were never in the data.
Where the excellent level lives
The hardest connections are at the intersections, and strongest where the epistemology of machine knowledge meets learning science. Joining how a model changes the production of knowledge to how a person actually learns is the through-line of the whole course, carried by four pairs:
Across all six pillars the same idea recurs: epistemology treated as an engineering problem.