This text is not a lesson in teaching method: it is what a colleague would give you before a first class on the subject. The course already follows the right order, each chapter building on the one before. What is missing for anyone coming to it new is knowing what to project, what to handle out loud, how to group the nine chapters into sessions, and the traps of the subject.

The order to take the ideas in

Nine chapters, four sessions. Each line says what is projected and what is handled out loud, and why.

1. Session 1: words in pieces, projected. A sentence suggested by the class rather than a prepared example: the cutting up of a familiar word convinces more than an abstract example. Five minutes are enough.
2. Session 1: the next word, projected. The direct continuation of the previous idea, not a session of its own: change the setting in front of the class and read out the words offered at each value.
3. Session 1: why it makes things up, out loud. Nothing more to project: the mechanism seen just before is enough to explain a wrong answer stated with assurance. Ask who has already had one.
4. Session 2: where the biases come from, projected. The demonstration on a dataset shows the bias before anyone names it, which avoids the feel of a moral lecture.
5. Session 2: what you give it, out loud. The idea most useful to daily life in class is better discussed than demonstrated. A concrete question: what becomes of a piece of homework typed into a free tool?
6. Session 3: a fabricated image, projected. The demonstration of artefacts is best watched together: a whole class spots more details than one person on their own.
7. Session 3: a fabricated voice, out loud. Doubt takes hold in the ear, not on a graph: let them listen rather than projecting.
8. Session 3: watermarks, projected. The direct answer to the question the two previous ideas raise: “so how do we know?”. To be shown as a partial lead, never as a solution.
9. Session 4: when not to use it, out loud. The close of the course, as a discussion rather than a list read aloud: three questions are enough to settle a case, the cost of an error, whether it can be checked, and legitimacy.

The full contents of the course, in this order, are on the understand page.

A model session

The first session above fits into an ordinary lesson and can be used exactly as it stands.

Both demonstrations run offline and without an account: they are opened the day before on a computer, once, and then work even if the school network goes down during the lesson. In a room with no reliable network, that is often what decides whether the session happens at all.

What you may do with this content

The course is published under the CC BY 4.0 licence: projecting it in class, printing it, translating it, putting it into existing material or a textbook are all allowed in advance, with no prior request and without telling the association. The exact credit line to include is on the licence and reuse page, and is not to be copied from here.

The traps of this subject

Two traps come up often, worth avoiding before they take hold.

Promising to detect a fabricated text is one. No detector of AI-generated text is both reliable and accurate, and the error does not fall at random: several widely used detectors wrongly classify the majority of texts written by pupils whose first language is not English as machine-generated, while correctly identifying the texts of native speakers of English. Accusing a pupil on the strength of a percentage therefore means punishing, first of all, the one for whom English is a second language. The chapter when not to use it sets out this evidence in detail, and a longer read is published on the blog: AI detectors at school.

• Testing of detection tools for AI-generated text Debora Weber-Wulff, Alla Anohina-Naumeca, Sonja Bjelobaba, Tomas Foltynek, Jean Guerrero-Dib, Olumide Popoola, Petr Sigut, Lorna Waddington, 2023. Establishes that an independent evaluation of fourteen detection tools, including commercial tools sold to education, shows no exception that is both reliable and accurate.
• GPT detectors are biased against non-native English writers Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu, James Zou, 2023. Establishes that several widely used detectors wrongly classify the majority of texts by non-native speakers of English as AI-generated.

The second trap is presenting as settled what is still debated. The chapter why it makes things up marks explicitly the part on which there is consensus and the part on which there is not. Carrying that distinction into class, rather than smoothing it over, costs nothing and saves having to correct yourself later.

The frameworks around this work

Three public frameworks help prepare a lesson, each at a different moment.

• AI Competency Framework for Teachers UNESCO, 2024. Useful for placing your own competencies before preparing a lesson, across five dimensions and three levels of progression.
• AI Competency Framework for Students UNESCO, 2024. Useful for setting what pupils should be able to do on leaving the course, by level of progression.
• Empowering Learners for the Age of AI (“AILit Framework”) OECD and European Commission, 2026. Useful for breaking a sequence down into precise competencies: engaging with AI, creating with it, managing it, designing it.

The detail of each, with what it allows you to state, is on the official frameworks page.

What this site does not provide

Three things, and this path says so before anyone looks for them in vain.

• An official progression, year by year.
• A marked assessment, with a marking scheme.
• Compliance with a national curriculum.

If any of those pieces comes to exist on this site, it will be a module beside the course, never inside its core.