The vocabulary you'll meet in policy documents, staff training, and the classroom — in plain language, precise where it changes decisions.

Agent / agentic AI — AI that takes multi-step actions toward a goal — browsing, running code, operating software — rather than only answering. It changes what “a student used AI” even means.

AI detector — Software claiming to spot machine-written text, mostly via predictability. Unreliable in both directions and biased against non-native and novice writers — indefensible as sole evidence.

AI literacy — The mix of tool fluency, technical understanding, and critical judgement students and staff need — the goal, not just “can they use the tool”.

Cognitive offloading — Using tools to reduce mental effort (lists, GPS, now AI). Often rational — but what you don't practise, you lose.

Context window — The model's working memory: how much text it can attend to at once. Everything in it is transmitted to the provider; nothing outside it exists to the model.

Generative AI — AI that produces new content — text, images, audio, code — rather than sorting existing content. The post-2022 wave this all concerns.

Guardrails — Product-level constraints on behaviour: instructions, filters, refusal rules. In the field's best trial, guardrails were the variable that separated harm from benefit.

Hallucination — Confident, fluent, false output. A structural feature of generation, not a bug awaiting a fix — and worst on specifics: numbers, quotes, citations, recent events.

Human-in-the-loop — Any setup where AI drafts and a human decides. The professional-ownership rule for feedback, grades, and consequential documents.

LLM (large language model) — A neural network trained on vast text to predict the next word; at scale, this produces the fluent, as-if-understanding behaviour behind chatbots.

Process evidence — Drafts, version history, notes and reflections that show how work was made — the fair-process alternative to a detector's verdict, and a pedagogical good in itself.

Productive failure — The finding that struggling with a problem before instruction produces deeper learning — the learning-science case against AI that answers instantly.

Prompt / system prompt — What you type — plus the hidden standing instructions a product inserts before every chat. Most education tools differ mainly at the system-prompt layer.

RAG (retrieval-augmented generation) — Answering from fetched documents — your curriculum, a textbook, the web — instead of the model's memory. Cuts fabrication sharply, but is only as good as what it retrieves.

Sycophancy — The trained tendency to agree with, mirror, and flatter the user. A specific tutoring hazard — it will validate a wrong answer — and a risk for vulnerable users.

Two-lane assessment — Separating secured, supervised assessment of learning (Lane 1) from open, AI-inclusive assessment for learning (Lane 2) — the emerging answer to the validity crisis.