The phrase "generative model" has been flattened into "chatbot" in most classroom conversations. That flattening costs students the concept. This post gives teachers a way to hold the idea intact, without leaning on a chatbot analogy that quietly does the opposite.
What the term actually means
In active inference, a generative model is an agent's internal account of how sensations are produced by hidden causes in the world (Class E, Parr, Pezzulo, and Friston, 2022). It is not a text predictor. It is not a chatbot. It is a structured guess about what is out there and how it moves, which the agent updates as new sensations arrive.
Two features matter for the classroom:
- The model generates predictions about incoming sensations before those sensations arrive.
- The mismatch between prediction and sensation drives updates to the model, to the actions the agent takes, or to both (Class E).
That is the whole idea, in language a student can hold. The math behind it (variational free energy, KL divergence between an approximate posterior and the true posterior over causes) is worth naming once so students know a formalism exists, then setting aside for the introductory pass.
Why the chatbot metaphor collapses
A chatbot completes text. It does not, in any meaningful sense, hold a model of hidden causes producing its sensations. When students are told "a generative model is like ChatGPT," three concepts get merged:
- The statistical machinery that produced the chatbot.
- The active-inference notion of a generative model as an agent's causal account of its world.
- The everyday idea of "making stuff up."
After that merge, students cannot reason about perception, action, or learning as unified processes. They think the term means "a thing that generates output." The active-inference reading (a model that generates predictions of sensations, against which incoming sensations are checked) is gone.
We teach UNI, not LLM tooling. See teach UNI not LLM tooling for the full fence.
Three classroom demonstrations that make the model observable
Each of these has been used in workshop settings and can be run with no code (Class C, workshop configuration).
Demonstration 1: The covered-object prediction
Place a familiar object under a cloth. Ask students to write down, without lifting the cloth, what they expect to feel when they reach under. Then have them reach under briefly, without looking, and write down what they actually felt. Compare the two lists.
Point out: their prediction was generated by a model. The mismatch (a corner they did not expect, a temperature that surprised them) is prediction error. This is exactly the loop active inference describes, at human speed and human scale.
Demonstration 2: The out-of-order sentence
Read a sentence aloud with one word replaced by a plausible wrong word ("The dog barked at the mailman" becomes "The dog barked at the envelope"). Ask students to name the moment they noticed. Ask what they had been predicting.
This surfaces the fact that comprehension is prediction, not passive reception. Students discover their own generative model in the act of it failing.
Demonstration 3: The two-camera drawing
Give students a small object and ask them to draw it from one angle. Then ask them to predict, before rotating it, what the back will look like, and draw that too. Rotate the object. Compare the predicted back to the actual back.
This makes visible the point that a generative model is a model of hidden causes, not visible surfaces. The back of the object was hidden. The student's drawing was a prediction about a cause (the object's three-dimensional shape) they could not see directly.
What this teaches, and what it does not
These demonstrations teach the shape of the concept. They do not teach the variational math, and they should not pretend to. When a student asks "is this how UNI works?", the honest answer is: UNI is a working hypothesis on an attainable path toward General Natural Intelligence, a natural, active-inference approach whose evidence is growing, evidence-classed, and tested in the open. Do not take the claim on faith. Test the build, inspect the gates, and help us find where it fails.
That framing keeps the classroom honest and keeps the student's curiosity pointed at the science, not at a product.
Where to go next
- What active inference actually says for the vocabulary underneath these demonstrations.
- The teacher's workbench tour for how to run the demonstrations in a class period.
- Markov blanket as a boundary idea for the companion concept that makes "hidden causes" precise.
- /workshop if you want the guided version with a facilitator.