A classroom is a room full of prediction machines. When we forget that, we lose the student.
Every learner walks in with a working model of the world. They expect the door to open, the teacher to greet them, the lesson to feel like the last one. When something violates a prediction, attention spikes. That spike is where learning lives. If we hand the prediction, and the checking, and the sense-making to a chatbot, the spike goes to the chatbot, not to the child.
This post is about protecting the spike.
The learner is an active inference agent
The active inference literature frames a mind as an agent that carries a generative model of its world, generates predictions from that model, samples evidence, and updates the model to reduce surprise (Class E, per Parr, Pezzulo, and Friston, 2022). That framing is not a metaphor imported into education. It is a description of what a student's brain is already doing while the lesson is running.
A few implications follow directly, and they are worth stating plainly.
A student who is not predicting is not learning. They may be present, compliant, even accurate on a worksheet. But without a prediction to test, there is nothing for evidence to update. The worksheet becomes copying, not thinking.
A student whose predictions are never wrong is not learning either. Zero prediction error is a signal that the task is beneath the model, or that the student has quietly stopped generating predictions and is pattern-matching to the answer key. Both are failures of the lesson, not the student.
A student whose predictions are always wrong, and who has no safe way to be wrong, is not learning in a useful direction. They are learning that prediction is dangerous. That is a separate problem, and a serious one, and it belongs to the adults in the room to fix. (This is where trauma-informed, non-clinical practice matters: we lower the cost of a wrong prediction without pretending the wrong prediction did not happen.)
What a generative-model classroom looks like
If we take the framing seriously, three habits move to the center of the lesson.
First, we ask for the prediction before we show the answer. Not as a warm-up, but as the core move. "Before I open this beaker, write down what you think will happen, and one reason." "Before we read the next paragraph, tell me what the narrator is about to admit." The prediction has to be committed, in writing or out loud, before evidence lands. Otherwise the student's memory will quietly rewrite the prediction to match the answer, and no update happens.
Second, we surface the error, gently and specifically. The useful sentence is not "you were wrong." It is "your model expected X and the world gave you Y. What changes?" That sentence assumes the student has a model, respects it, and points at the exact place an update is needed. It is also, not accidentally, a trauma-informed sentence: the student is not the error, the model is.
Third, we protect the checking. If the check is a chatbot response the student reads and nods at, no evidence has been sampled by the student's own model. The check has to be something the student does: an experiment, a re-read, a peer disagreement, a worked example the student walks through with a pencil. Reading a confident paragraph from a generative model is not evidence. It is a summary of somebody else's prediction.
What goes missing when we outsource sense-making
Large generative models produce fluent, plausible, often-correct text. In a classroom, that fluency is the danger, not the feature.
When a student hands a question to a chatbot and takes back a paragraph, three things fail to happen (Class B, this is observable in the artifacts students produce, and in the followup questions they cannot answer).
The prediction step is skipped. The student never committed to what they thought. So there is nothing to update.
The evidence step is skipped. A paragraph from a language model is not an observation of the world. It is a probability-weighted next-token stream conditioned on the prompt. It can be right, wrong, or confidently in between, and the student has no way, from inside the paragraph, to tell which. (Class F, the falsifier here is simple: ask the student to defend one specific claim in the paragraph from a source they did not get from the same tool. If they cannot, the paragraph was not evidence, it was decor.)
The update step is skipped. Because there was no prediction and no evidence, there is nothing to reduce surprise about. The student's working model of, say, photosynthesis or the French Revolution is unchanged. The homework got done. The learning did not.
This is not an argument that generative tools have no place near students. It is an argument that the sequence matters. Prediction first, evidence second, model update third. Tools that enter before the prediction is committed corrode agency. Tools that enter after can, sometimes, serve as one more source of evidence, as long as the student is still the one doing the sampling and the updating.
Agency, work, and the honest conversation
A fair objection at this point: aren't we preparing students for a world where these tools are everywhere, including at work?
Yes. And the honest human-market picture is worth reading directly. Themesis published a conversation with Jay Kumar Chimata on JobFirst.ai and the actual shape of the current AI job market (link); our one-line frame in our voice is that the market rewards people who can think under, around, and past these tools, not people who can prompt them. That reading, if it holds, is exactly the case for protecting the prediction habit in school. The employable skill is the update loop, not the paragraph.
What we teach on top of this
There is a technical program behind these practices. Publicly, we describe it as a working hypothesis: 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. (Class U at the program level; individual pieces carry their own evidence classes.)
For a classroom, the useful export is smaller and more concrete. It is the prediction-then-check habit, the "your model expected X, the world gave you Y" sentence, the refusal to let a chatbot do the sampling for a student, and the willingness to slow the lesson down so an actual update can happen.
That is what a generative-model classroom protects. Not the model in the API. The one in the child.
Where to go next
- Curriculum design that keeps the prediction step honest: curriculum gates that teach thinking.
- The daily-lesson version of the core move: prediction then check in daily lessons.
- The failure mode this post names: outsourcing thinking, what goes missing.
- A student-facing artifact schools have adapted: student declarations, a template.
- For teachers and school leaders who want the practitioner-track walkthrough, the workshop is where we teach the sequence directly.