AI in every subject, not just computing

The slot does not exist

The progression lesson ended on the practical problem: AI has no exam, so it has no slot, so it is taught in gaps, if at all. The solution most schools reach is to give it to the computing teacher, if there is one, and the computing teacher covers the mechanism in two lessons and moves on to the syllabus.

The better solution is to notice that most of what students need to know about AI is already a piece of some subject’s method, and to distribute the unit across the timetable. One meeting, one agreed list of who teaches what, and the progression happens across the year in the subjects best placed to teach each piece.

What each subject naturally owns

History: provenance and evidence. Historians ask where a source came from, who made it, why, and what it can and cannot tell you. That is the deepfake lesson’s “provenance, not appearance” and the reliability gradient, in a subject that has taught it for a century. The check question’s lesson: a generated account of a local event beside the real documents (a newspaper, a photograph, an interview) and the question “which of these is evidence, and what makes it so?” Students who have done this once read a chatbot’s confident history of their town differently for life.

Science: evaluation and the fair test. Science teaches how to test a claim. Give students twenty questions, a chatbot, and a mark scheme, and ask them to measure how often it is right; then ask whether the twenty were a fair test, and what a fair test would need. The “confident wrong answer” activity is a science lesson on the reliability of instruments. The sorting game’s white-background failure is a lesson on confounding variables.

Languages: the reliability gradient and translation. A language teacher can show, in ten minutes, that the model is excellent at French and shaky at Hausa, and ask why. Translation of a passage by the model beside the class’s own translation, with the errors found and explained, teaches both the language and the mechanism. And “same question, different answer” is visible instantly when the same sentence is translated twice.

Maths: the arithmetic failure. The calculator lesson’s demonstration (two three-digit numbers, checked by hand) belongs in a maths room. So does the bigram project’s tally, which is probability made physical. A maths teacher can also teach the one idea most adults lack: that a machine predicting from patterns produces the most common answer, not the correct one, and why those differ.

English: the authorship question. What does it mean to write something? Who wrote a paragraph that a student prompted, selected and edited? The citing lesson’s unsettled question is a real question in literature, and a class that reads a generated poem beside a human one and argues about which is better, and why, and whether it matters who wrote it, is doing English.

Art: the training data and the artists. Image generators learned from artists’ work, mostly without permission, and whether that was lawful is being fought in courts in several countries. Art students have a direct stake, and a lesson that generates an image “in the style of” a living artist and asks whether that is theft, homage or neither is a citizenship lesson wearing an art lesson’s clothes.

Citizenship, PSHE, religious studies: the companion apps, the deepfakes, and who profits. The two pastoral lessons in this block belong here, along with the question most units skip: these are products, made by companies, and the choices inside them were made by people with interests. Who chose the training data? Who decided what it refuses? Who is paid when a student spends two hours with a companion? A subject that already asks who holds power and why is the right home for that.

Computing, where it exists: the mechanism and the project. The bigram build, Teachable Machine, and the step toward real code. Computing teaches the how; the other subjects teach the so what, and both are needed.

The meeting

One hour. The progression lesson’s age map on the board. Each department claims a piece and a year group. A one-line record of what will be taught, by whom, in which term. A shared folder for the materials, and the printed transcripts and activity sheets go in it, because half of these lessons need no device.

The output is a unit that exists across the year without any timetable change, that is taught by the people best placed to teach each part, and that a student meets six or seven times in different rooms rather than twice in one. That repetition, across subjects, is how an idea like “it is confident when it is wrong” becomes something a student owns rather than something they were told.

The honest limit

Distributed teaching has a failure mode: everyone assumes someone else covered the basics. The prediction activity and the sorting game have to happen somewhere, first, and the meeting must name where. And a department that is stretched will drop its AI lesson the week the exam approaches. Put the lessons early in the year, in the weeks before the pressure, and accept that in a bad year some of them will not happen. Four subjects teaching it is still four more than one computing lesson in a slot that does not exist.