Active inference has a reputation for being hard. Most of that reputation comes from the notation, not the idea. Underneath the equations there is a picture a teacher can hold, and once you can hold the picture the equations stop feeling like a wall.
This post walks the core claim in plain language, keeps the technical terms visible so you can look them up later, and lands on what a classroom can honestly do with them.
The core claim, one sentence
An agent, whether a brain, a cell, or a well-behaved learning system, keeps a working model of its situation and updates that model to reduce the gap between what it expects and what it senses. In the Parr, Pezzulo, and Friston formulation (Class E), that gap has a name: variational free energy. Reducing it is the whole game.
That is it. The rest is machinery for making the sentence precise.
Generative model, in classroom terms
A generative model is the internal story an agent tells about how its world produces the signals it receives. A student solving a word problem has a generative model of what the problem "is doing." A reader has a generative model of where a sentence is going. Neither of those models is complete. Both are running.
The active-inference claim is that the agent is not just passively predicting. It is choosing actions that will make its next observations easier to explain under the model it already holds, or actions that will teach the model something it did not know.
Two moves, always both running:
- Update the model to match the world (perception).
- Act on the world to match the model (action, including asking a question).
Free energy, without the equation
The formal object is variational free energy, and in the Friston line of work it upper-bounds the surprise an agent would otherwise have to compute directly (Class E). You do not need the integral to teach the intuition. What matters for a classroom:
- Surprise is not the emotional feeling. It is a measurement of how badly an observation fits the current model.
- Free energy is the accountant's version of surprise: a quantity the agent can actually compute from its own model plus what it just saw.
- Minimizing free energy means the agent is either changing its mind or changing the world so the two line up better.
When a student says "wait, that does not make sense," they have just registered a spike in free energy. What they do next, reread, ask, guess, walk away, is an action policy. Active inference says the policy is chosen because it is expected to reduce free energy going forward, not just at this moment.
Why this matters for teaching
The active-inference frame gives teachers three usable handles:
- Prediction is doing work all the time, not just during "assessment." A quiet student is still running a model.
- Confusion is a signal about the model, not a verdict on the student. It is where learning is possible.
- Acting to learn (asking, testing, moving) is part of the same loop as thinking. Separating "hands-on" from "cognitive" misses the point.
None of this requires you to run the math. It does require you to trust that the vocabulary is doing real work when you use it, which is why we keep the terms visible instead of hiding them behind analogies.
Where this fits in the family
Themesis publishes a resource map that lists several pathways into active inference for 2026 (Class E). See Where to Start with Active Inference, A Resource Map for 2026. In our voice: it is the reference we point teachers to when they want the wider landscape before picking a lane.
Our lane is educator readiness. We teach the concepts so you can read the field, not so you can pretend to have run the experiments.