Foundations for AI Educators

A course in the theory required to teach with and about artificial intelligence: how large language models operate, where their outputs acquire epistemic standing, and how people learn. Five lectures across the technical, epistemological, cognitive, regulatory, and historical dimensions of the field, with a companion pair on research method.

It is addressed to educators — those who instruct others in the use of AI rather than those who only use it — and it takes the underlying theory, rather than any particular tool or interface, as its stable object of study.

The course proceeds on the premise that competent instruction in this domain presupposes an account of the systems themselves. Work with agentic, language-model-based systems depends on understanding their engineering constituents — tokenization, the context window, retrieval, evaluation — and the epistemic questions these raise concerning provenance, justification, and the limits of machine-generated knowledge. That account is presented as a single body of theory, organised as five lectures across the technical, epistemological, cognitive, regulatory, and historical dimensions of the field, with a companion pair on quantitative and qualitative research method.

01

Theoretical foundations

Five lectures, one per pillar. The first three are the core — the technology, its epistemology, and the science of learning; the remaining two set the frame: ethics and regulation, and intellectual history.

01.01How language models work

An account of large language models below the level of metaphor. The lecture proceeds from tokens and embeddings to attention and the context window, then to how a model is pre-trained and adapted, the mechanism of hallucination, and what calibration and the scaling laws describe. The governing method is to refer each observed behaviour to a specific component of the architecture rather than to treat it as a property of the system as a whole.

tokenizationembeddingsattentioncontext windowlost-in-the-middlepre- & post-trainingprompting vs RAG vs fine-tuninghallucination & calibrationscaling lawsinference-time reasoning

01.02The epistemology of machine knowledge

An examination of the conditions under which a model’s output may be treated as knowledge rather than as a fluent guess. The lecture distinguishes fluency from truth, defines a claim as knowledge only when it carries a verifiable source and an exact location, and analyses model output through the epistemology of testimony, drawing on Frankfurt’s account of indifference to truth and Popper’s criterion of falsifiability.

testimonyprovenanceknowledge = claim + source + locationfluency ≠ truthFrankfurt on bullshitPopper & falsifiabilityWorld 3calibration of trust

01.03Cognitive and learning science

An account of how people learn and of the demands this places on any instrument interposed between a learner and a subject. The lecture treats cognitive load, the constructionist thesis that understanding is built rather than transmitted, scaffolding designed to be withdrawn, and difficulties that aid retention, arriving at the question of which cognitive operations may be delegated to a model and which must remain with the learner.

cognitive load (Sweller)constructionism (Papert)ZPD & scaffolding (Vygotsky)desirable difficulties (Bjork)expertise-reversal effecttransfermetacognitioncognitive offloading

01.04Ethics, society, and regulation

A survey of the ethical and regulatory questions raised by language models, addressed individually: representational harm, the data and human labour on which a model rests, copyright and privacy, the environmental cost of scale, and the consequences for work. The lecture reads the risk tiers of the EU AI Act, under which educational systems frequently qualify as high-risk, and consolidates the material into a seven-criterion audit.

representational harmdata provenanceghost workcopyrightGDPR & privacyenergy & water costautomation vs augmentationEU AI Act risk tiersseven-criterion audit

01.05History and intellectual lineage

A placement of language models within a seventy-year intellectual lineage beginning in 1956. The lecture traces the succession from symbolic artificial intelligence and connectionism through the AI winters and the statistical turn to attention and the transformer, on the principle that this history is what allows an educator to distinguish a genuine advance from its surrounding rhetoric.

1956 Dartmouthsymbolic AI vs connectionismAI wintersthe statistical turnattention & the transformerVaswani et al., 2017

02

Applied research methods

Two further lectures apply the theory to the conduct of research, treating the language model both as a measuring instrument and as the object under study.

02.01Quantitative research with LLMs

A treatment of the language model as a measuring instrument. The lecture begins from the observation that a single output is not an estimate, treats output as a random variable with quantifiable variance, surveys the metrics by which systems are evaluated, and addresses the threats — contamination, overfitting, judge bias — that invalidate an apparent result.

output as a random variablesampling & variancebootstrap & CIsprecision · recall · F1nDCGcalibration errorLLM-as-judge biasCohen’s κ · Krippendorff’s αp-values & effect sizemultiple-comparison correctiondata contaminationbenchmark overfitting

02.02Qualitative research with LLMs

A treatment of interpretive method under conditions of machine assistance. The lecture takes coding, grounded theory and the criteria of trustworthiness seriously as method, and identifies the characteristic failure that arises when interpretation itself is delegated to the model: the report of themes with no basis in the data.

open · axial · selective codinggrounded theorytheoretical saturation (Glaser & Strauss)thematic analysis (Braun & Clarke)trustworthinessaudit trail as provenance (Lincoln & Guba)hallucinated themes