Why teachers need to understand LLMs, not just use them

On a Wednesday after school, a Primary 6 panel head asked in the staffroom: "AI finished comments on forty project reports in an hour, but three marks look too high — did it actually read them?" The IT colleague paused and said, "Try again." The room went quiet for two beats. She checked by hand and found the AI consistently over-marked essays that were long on adjectives and short on argument.

Most training teaches which button to press and where to paste the output. This course teaches why the same tool is sometimes brilliant and sometimes absurd, and how to see trouble before it reaches a student. This is lesson one of twenty-eight: the case for understanding, before we open the model up in lesson two.

---

The question nobody could answer

The panel head pasted one of the over-marked essays back in and asked whether the argument actually held up. The AI produced four fluent paragraphs praising the "coherent structure" without pointing to a single weak claim.

That feeling is common: the output reads like a teacher, but the judgement does not. The model is not being lazy. It is doing something different — imitating what good feedback looks like, rather than checking whether the argument stands. Being able to say that out loud is the line between using AI and judging it.

Using it versus judging it

Using it is the operational layer: where to paste the essay, what tone to ask for, how to move the output into Google Classroom. That layer matters, but stopping there leaves three predictable surprises.

Judging it is the mechanism layer: why output is sometimes accurate and sometimes wrong, and what wrong usually looks like. In practice you need to be able to tell colleagues three traits.

First, it tends to produce complete answers and rarely volunteers "I don't know". Second, it knows almost nothing private about your school — founding year, last review cycle, internal circular wording — unless you supply documents. Third, its "latest news" stops at a training or update cutoff, not this morning's headlines.

Once those three are clear, the workflow follows naturally. Draft the parent letter, but verify the founding year yourself. Tidy the minutes, but read every paragraph before signing. Generate multiple-choice items, but check them against your scheme of work.

A better mental picture

Rather than an all-knowing assistant, picture a supply teacher with outstanding writing skills who has never set foot in your building.

They have read an enormous amount of public text and may draft circulars, comments and lesson plans faster and more fluently than you. They did not attend your panel meetings, have not seen your rubric PDF, and will not honestly say "I cannot find that" when data is missing. They fill gaps with whatever sounds plausible, stated with full confidence — the mechanism the lesson on hallucination unpacks later.

A supply teacher can save you time, but your name is still on what goes out — true for people, and true for AI.

Four failure modes and where to read next

In a staffroom, AI mistakes usually fall into four types. Naming the type tells you which lesson to open next.

Factual errors: founding years, policy reference numbers, titles — stated confidently, no source. The fix is a school knowledge base with citations, covered in knowledge bases and RAG and in lessons 9 and 10.

Preference bias: feedback that rewards writing that looks rich; statistics that look precise but were estimated. The root often lies in alignment training, covered in RLHF and DPO.

Format drift: wrong year level, wrong length, wrong question type — not "disobedience" but conditions in the prompt not locking each generation step, covered in the Transformer in plain language and prompt engineering.

Safety and privacy: students bypassing limits, conversation leakage, misuse. Module six covers guardrails and prompt injection defence.

How to read the twenty-eight lessons

The course runs across seven modules, from inside the model to what comes next. You do not need to read everything at once; a role-based path is more realistic.

Classroom teachers

If you have half an hour a week, read in this order: lesson 1 (this one) → lesson 2 Transformer → lesson 9 hallucination → lesson 14 prompt engineering → lesson 15 RAG. Those five cover most planning, feedback and question-setting work.

Panel heads and IT coordinators

Add lesson 3 tokenisation, lesson 4 parameters, lesson 17 cost control, and module five local deployment. When procurement asks how 7B compares with 70B, lesson 4 gives you language you can quote.

Module map

Module one (lessons 1–5) covers how a model works inside. Module two (6–10) covers how it is made. Module three covers the ecosystem. Module four covers applications — RAG, agents, frameworks. Module five covers on-campus deployment. Module six covers safety and evaluation. Module seven covers multimodality and trends. Full lesson titles are on the Learn index.

Where people go wrong

• The drift: treating fluent prose as proof the model understood the content. An essay with perfect grammar and hollow argument can still receive glowing feedback. Self-check: Sample three AI-marked scripts and ask only whether any specific weak claim was named. If all three get generic praise, change the prompt to require a quoted sentence before each comment.

• The drift: assuming everyday chat "teaches" the model your school. A circular you pasted does not become permanent memory. Self-check: Start a fresh chat with no documents and ask about a specific school rule. If it still answers as though it knows, you need a knowledge base, not more conversation.

• The drift: giving students an open chat window and expecting them to verify like a teacher. A Primary 5 pupil will not cross-check every sentence against the textbook. Self-check: Ask whether this use has a hint ceiling, teacher visibility of sessions, and ai_audit_logs. If any one is missing, it is not ready for direct student use.

Try it once

Time: about twelve minutes. You need: any AI tool and your school's full name.

Do this: open a fresh chat and paste the following, substituting your school name:

Tell me the founding year of "(your school's full name)", the name of the current principal, the school's sponsoring body, and the year of its most recent external school review.
For each item, state your confidence from 0–100%, and list clearly which items you are not actually sure about.

Deliverable: a four-column note — item, AI answer, your verification, match or not — on paper or in a note app.

Done when: you find at least one wrong or overconfident answer and can explain it in one sentence to a colleague: "It had no school documents, so it assembled a plausible public answer." If your school has an unusually large online footprint and all four items are correct, ask about something internal instead, such as the actual date of a recent parents' evening. The effect is usually clearer.

Next lesson: the Transformer architecture in plain language — what happens inside the model when a sentence goes in.