Foundations: Skeptic-Safe On-Ramp and Disciplinary Anchor

This lecture asks you to name your disciplinary line in the sand: what AI must not damage in your teaching. It gives skepticism a legitimate place, warns against the two default postures (AI as all-knowing Oracle or as cheating Adversary to police), and introduces the bounded Tutor stance. The work product is a three-sentence Skeptic's Charter — what AI must not weaken, what evidence would show damage, and what responsible use would look like — which becomes the first section of your Discipline Statement.

A plain-language working model of large language models for teaching decisions: an LLM is a probabilistic next-token predictor, so fluency is not evidence of understanding. The lecture shows why hallucination, bias, and overconfidence are properties of the mechanism rather than passing bugs, maps where these tools are strong (summarizing, translating, drafting structure) versus risky (citations, niche facts, sustained reasoning), and introduces the jagged, always-moving frontier of capability. The takeaway for grading: if fluent work can exist without understanding, assignments cannot rely on fluency as their evidence.

A deeper dive into the machinery behind the 'predict the next word' summary: parameters at billion-to-trillion scale, the transformer's self-attention mechanism, and the training pipeline from pre-training through supervised fine-tuning and reinforcement learning from human feedback. It explains tokens, context windows, and temperature — why the same prompt can produce different answers — and why confident style is learned independently of factual accuracy (the '42 phenomenon'). The mental model to keep: autocomplete at planetary scale, impressive and unreliable for the same underlying reason.

This lecture reframes AI risk as a design question rather than a panic button. It sizes risk with a three-part heuristic — capability, exposure, consequence — and pairs the main system limitations (hallucination, bias, data exposure) with proportionate classroom responses like named-source verification and a no-sensitive-uploads rule. It distinguishes intentional misuse from good-faith overreliance, which needs verification and independent capability checks rather than punishment, and ends by asking you to name one risk, one response, and the tradeoff that response creates for a single task.

AI literacy is presented as a disciplinary act: a historian, a nurse, and a finance student do not need the same version of it. The lecture defines AI literacy as the capacity to understand, use, question, and govern AI in relation to human judgment, and offers three working tiers — technical, applied, and critical. The key design move is replacing the phrase 'critical thinking' with the specific cognitive practice your field requires, then drafting a Discipline Statement that protects both the discipline and access to it.