On the Three Worlds: an ontology for AI-mediated learning

The MEM framework distinguishes Mental, Formal, and Physical domains of educational reality, each with its own modes of error and shift. Generative AI lives natively in World 2, with consequences for how it ought to be taught with.

There is a temptation, in the present moment, to write about generative AI in education as if it were a single thing. It is not. The argument I want to make in this essay is that the most useful first move is ontological: we need a clearer picture of what kind of system a school is, and what kind of object a generative model is, before we can usefully argue about what to do.

The framework I have been working with — the Three Worlds — is borrowed and modified from Karl Popper's later epistemology and from Roger Penrose's three-worlds diagram in The Road to Reality. It distinguishes:

• World 1 — the Mental. The interior life of the learner: their perspectives, opinions, partial understandings, creative work in the act of being made.
• World 2 — the Formal. The system of theory, propositions, frameworks, and codified knowledge that schools have always asked students to traffic in.
• World 3 — the Physical. The world of action — workshops, fieldwork, embodied practice, the doing.

What the framework asks us to notice

Each of the three worlds has its own modes of error. World 1 errors are misperceptions, partial models, intuitions that fail. World 2 errors are propositional: a false claim, a broken proof, a citation that does not exist. World 3 errors are practical: a circuit that does not light, a wall that does not hold, a class that does not learn.

Generative AI sits squarely in World 2. That is what it is for. It is exquisitely good at producing surface-level, well-formed propositional artefacts — essays, summaries, code that compiles. It is structurally weak in World 1 (it has no interior) and indirect in World 3 (it can describe but not do).

Children do not get ideas; they make them.

What follows for teaching

Once the worlds are named, the practical question becomes: which world is this learning task actually in? Many tasks that schools have historically run as World 2 tasks — write me an essay on the causes of the First World War — were always also doing other work. They were probes into World 1 (do you, the student, hold a position?) and World 3 (can you sustain a long, structured argument over several days of writing?).

The arrival of a competent World 2 assistant collapses the surface. The probe stops working. But the learning we cared about was not in the World 2 surface — it was in the other two worlds, which the assistant cannot enter.

What to build

This is not, in the end, an argument about AI. It is an argument about what schools were for. The framework I propose is meant to make that visible. Subsequent posts in this series will walk through specific assessment forms — oral defence, portfolio review, process-based assessment — that recover World 1 and World 3 evidence in the presence of a World 2 assistant.